update code

minigemini/__init__.py

@@ -0,0 +1,3 @@
+import timm
+import open_clip
+from .model import MiniGeminiLlamaForCausalLM

minigemini/constants.py

@@ -0,0 +1,27 @@
+CONTROLLER_HEART_BEAT_EXPIRATION = 30
+WORKER_HEART_BEAT_INTERVAL = 15
+
+LOGDIR = "."
+
+# Model Constants
+IGNORE_INDEX = -100
+IMAGE_TOKEN_INDEX = -200
+PREDICT_TOKEN_INDEX = -300
+DEFAULT_IMAGE_TOKEN = "<image>"
+DEFAULT_IMAGE_PATCH_TOKEN = "<im_patch>"
+DEFAULT_IM_START_TOKEN = "<im_start>"
+DEFAULT_IM_END_TOKEN = "<im_end>"
+IMAGE_PLACEHOLDER = "<image-placeholder>"
+DEFAULT_PREDICT_TOKEN = "<predict>"
+
+DESCRIPT_PROMPT = [
+    "Describe this image thoroughly.",
+    "Provide a detailed description in this picture.",
+    "Detail every aspect of what's in this picture.",
+    "Explain this image with precision and detail.",
+    "Give a comprehensive description of this visual.",
+    "Elaborate on the specifics within this image.",
+    "Offer a detailed account of this picture's contents.",
+    "Describe in detail what this image portrays.",
+    "Break down this image into detailed descriptions.",
+    "Provide a thorough description of the elements in this image."]

minigemini/conversation.py

@@ -0,0 +1,460 @@
+import dataclasses
+from enum import auto, Enum
+from typing import List, Tuple
+import base64
+from io import BytesIO
+from PIL import Image
+
+
+class SeparatorStyle(Enum):
+    """Different separator style."""
+    SINGLE = auto()
+    TWO = auto()
+    MPT = auto()
+    PLAIN = auto()
+    LLAMA_2 = auto()
+    GEMMA = auto()
+
+
+@dataclasses.dataclass
+class Conversation:
+    """A class that keeps all conversation history."""
+    system: str
+    roles: List[str]
+    messages: List[List[str]]
+    offset: int
+    sep_style: SeparatorStyle = SeparatorStyle.SINGLE
+    sep: str = "###"
+    sep2: str = None
+    version: str = "Unknown"
+
+    skip_next: bool = False
+
+    def get_prompt(self):
+        messages = self.messages
+        if len(messages) > 0 and type(messages[0][1]) is tuple:
+            messages = self.messages.copy()
+            init_role, init_msg = messages[0].copy()
+            init_msg = init_msg[0].replace("<image>", "").strip()
+            if 'mmtag' in self.version:
+                messages[0] = (init_role, init_msg)
+                messages.insert(0, (self.roles[0], "<Image><image></Image>"))
+                messages.insert(1, (self.roles[1], "Received."))
+            else:
+                messages[0] = (init_role, "<image>\n" + init_msg)
+
+        if self.sep_style == SeparatorStyle.SINGLE:
+            ret = self.system + self.sep
+            for role, message in messages:
+                if message:
+                    if type(message) is tuple:
+                        message = message[0]
+                    ret += role + ": " + message + self.sep
+                else:
+                    ret += role + ":"
+        elif self.sep_style == SeparatorStyle.TWO:
+            seps = [self.sep, self.sep2]
+            ret = self.system + seps[0]
+            for i, (role, message) in enumerate(messages):
+                if message:
+                    if type(message) is tuple:
+                        message = message[0]
+                    ret += role + ": " + message + seps[i % 2]
+                else:
+                    ret += role + ":"
+        elif self.sep_style == SeparatorStyle.MPT:
+            ret = self.system + self.sep
+            for role, message in messages:
+                if message:
+                    if type(message) is tuple:
+                        message = message[0]
+                    ret += role + message + self.sep
+                else:
+                    ret += role
+        elif self.sep_style == SeparatorStyle.LLAMA_2:
+            wrap_sys = lambda msg: f"<<SYS>>\n{msg}\n<</SYS>>\n\n" if len(msg) > 0 else msg
+            wrap_inst = lambda msg: f"[INST] {msg} [/INST]"
+            ret = ""
+
+            for i, (role, message) in enumerate(messages):
+                if i == 0:
+                    assert message, "first message should not be none"
+                    assert role == self.roles[0], "first message should come from user"
+                if message:
+                    if type(message) is tuple:
+                        message = message[0]
+                    if i == 0: message = wrap_sys(self.system) + message
+                    if i % 2 == 0:
+                        message = wrap_inst(message)
+                        ret += self.sep + message
+                    else:
+                        ret += " " + message + " " + self.sep2
+                else:
+                    ret += ""
+            ret = ret.lstrip(self.sep)
+        elif self.sep_style == SeparatorStyle.GEMMA:
+            seps = [self.sep, self.sep2]
+            ret = self.system + seps[0]
+            for i, (role, message) in enumerate(messages):
+                if message:
+                    if type(message) is tuple:
+                        message = message[0]
+                    ret += "<start_of_turn>" + role + "\n" + message + "<end_of_turn>\n" + seps[i % 2]
+                else:
+                    ret += "<start_of_turn>" + role + "\n"
+        elif self.sep_style == SeparatorStyle.PLAIN:
+            seps = [self.sep, self.sep2]
+            ret = self.system
+            for i, (role, message) in enumerate(messages):
+                if message:
+                    if type(message) is tuple:
+                        message = message[0]
+                    ret += message + seps[i % 2]
+                else:
+                    ret += ""
+        else:
+            raise ValueError(f"Invalid style: {self.sep_style}")
+
+        return ret
+
+    def append_message(self, role, message):
+        self.messages.append([role, message])
+
+    def process_image(self, image, image_process_mode, return_pil=False, image_format='PNG', max_len=1344, min_len=672):
+        if image_process_mode == "Pad":
+            def expand2square(pil_img, background_color=(122, 116, 104)):
+                width, height = pil_img.size
+                if width == height:
+                    return pil_img
+                elif width > height:
+                    result = Image.new(pil_img.mode, (width, width), background_color)
+                    result.paste(pil_img, (0, (width - height) // 2))
+                    return result
+                else:
+                    result = Image.new(pil_img.mode, (height, height), background_color)
+                    result.paste(pil_img, ((height - width) // 2, 0))
+                    return result
+            image = expand2square(image)
+        elif image_process_mode in ["Default", "Crop"]:
+            pass
+        elif image_process_mode == "Resize":
+            image = image.resize((336, 336))
+        else:
+            raise ValueError(f"Invalid image_process_mode: {image_process_mode}")
+        if max(image.size) > max_len:
+            max_hw, min_hw = max(image.size), min(image.size)
+            aspect_ratio = max_hw / min_hw
+            shortest_edge = int(min(max_len / aspect_ratio, min_len, min_hw))
+            longest_edge = int(shortest_edge * aspect_ratio)
+            W, H = image.size
+            if H > W:
+                H, W = longest_edge, shortest_edge
+            else:
+                H, W = shortest_edge, longest_edge
+            image = image.resize((W, H))
+        if return_pil:
+            return image
+        else:
+            buffered = BytesIO()
+            image.save(buffered, format=image_format)
+            img_b64_str = base64.b64encode(buffered.getvalue()).decode()
+            return img_b64_str
+
+    def get_images(self, return_pil=False):
+        images = []
+        for i, (role, msg) in enumerate(self.messages[self.offset:]):
+            if i % 2 == 0:
+                if type(msg) is tuple:
+                    msg, image, image_process_mode = msg
+                    image = self.process_image(image, image_process_mode, return_pil=return_pil)
+                    images.append(image)
+        return images
+
+    def to_gradio_chatbot(self):
+        ret = []
+        for i, (role, msg) in enumerate(self.messages[self.offset:]):
+            if i % 2 == 0:
+                if type(msg) is tuple:
+                    msg, image, image_process_mode = msg
+                    img_b64_str = self.process_image(
+                        image, "Default", return_pil=False,
+                        image_format='JPEG')
+                    img_str = f'<img src="data:image/jpeg;base64,{img_b64_str}" alt="user upload image" />'
+                    msg = img_str + msg.replace('<image>', '').strip()
+                    ret.append([msg, None])
+                else:
+                    ret.append([msg, None])
+            else:
+                if type(msg) is tuple and len(msg) == 2:
+                    msg, img_b64_str = msg
+                    img_str = f'<img src="data:image/jpeg;base64,{img_b64_str}" alt="user upload image" />'
+                    msg = msg.strip() + img_str
+                ret[-1][-1] = msg
+        return ret
+
+    def copy(self):
+        return Conversation(
+            system=self.system,
+            roles=self.roles,
+            messages=[[x, y] for x, y in self.messages],
+            offset=self.offset,
+            sep_style=self.sep_style,
+            sep=self.sep,
+            sep2=self.sep2,
+            version=self.version)
+
+    def dict(self):
+        if len(self.get_images()) > 0:
+            return {
+                "system": self.system,
+                "roles": self.roles,
+                "messages": [[x, y[0] if type(y) is tuple else y] for x, y in self.messages],
+                "offset": self.offset,
+                "sep": self.sep,
+                "sep2": self.sep2,
+            }
+        return {
+            "system": self.system,
+            "roles": self.roles,
+            "messages": self.messages,
+            "offset": self.offset,
+            "sep": self.sep,
+            "sep2": self.sep2,
+        }
+
+
+conv_vicuna_v0 = Conversation(
+    system="A chat between a curious human and an artificial intelligence assistant. "
+           "The assistant gives helpful, detailed, and polite answers to the human's questions.",
+    roles=("Human", "Assistant"),
+    messages=(
+        ("Human", "What are the key differences between renewable and non-renewable energy sources?"),
+        ("Assistant",
+            "Renewable energy sources are those that can be replenished naturally in a relatively "
+            "short amount of time, such as solar, wind, hydro, geothermal, and biomass. "
+            "Non-renewable energy sources, on the other hand, are finite and will eventually be "
+            "depleted, such as coal, oil, and natural gas. Here are some key differences between "
+            "renewable and non-renewable energy sources:\n"
+            "1. Availability: Renewable energy sources are virtually inexhaustible, while non-renewable "
+            "energy sources are finite and will eventually run out.\n"
+            "2. Environmental impact: Renewable energy sources have a much lower environmental impact "
+            "than non-renewable sources, which can lead to air and water pollution, greenhouse gas emissions, "
+            "and other negative effects.\n"
+            "3. Cost: Renewable energy sources can be more expensive to initially set up, but they typically "
+            "have lower operational costs than non-renewable sources.\n"
+            "4. Reliability: Renewable energy sources are often more reliable and can be used in more remote "
+            "locations than non-renewable sources.\n"
+            "5. Flexibility: Renewable energy sources are often more flexible and can be adapted to different "
+            "situations and needs, while non-renewable sources are more rigid and inflexible.\n"
+            "6. Sustainability: Renewable energy sources are more sustainable over the long term, while "
+            "non-renewable sources are not, and their depletion can lead to economic and social instability.\n")
+    ),
+    offset=2,
+    sep_style=SeparatorStyle.SINGLE,
+    sep="###",
+)
+
+conv_vicuna_v1 = Conversation(
+    system="A chat between a curious user and an artificial intelligence assistant. "
+    "The assistant gives helpful, detailed, and polite answers to the user's questions.",
+    roles=("USER", "ASSISTANT"),
+    version="v1",
+    messages=(),
+    offset=0,
+    sep_style=SeparatorStyle.TWO,
+    sep=" ",
+    sep2="</s>",
+)
+
+conv_llama_2 = Conversation(
+    system="""You are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe.  Your answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure that your responses are socially unbiased and positive in nature.
+
+If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information.""",
+    roles=("USER", "ASSISTANT"),
+    version="llama_v2",
+    messages=(),
+    offset=0,
+    sep_style=SeparatorStyle.LLAMA_2,
+    sep="<s>",
+    sep2="</s>",
+)
+
+conv_llava_llama_2 = Conversation(
+    system="You are a helpful language and vision assistant. "
+           "You are able to understand the visual content that the user provides, "
+           "and assist the user with a variety of tasks using natural language.",
+    roles=("USER", "ASSISTANT"),
+    version="llama_v2",
+    messages=(),
+    offset=0,
+    sep_style=SeparatorStyle.LLAMA_2,
+    sep="<s>",
+    sep2="</s>",
+)
+
+conv_mpt = Conversation(
+    system="""<|im_start|>system
+A conversation between a user and an LLM-based AI assistant. The assistant gives helpful and honest answers.""",
+    roles=("<|im_start|>user\n", "<|im_start|>assistant\n"),
+    version="mpt",
+    messages=(),
+    offset=0,
+    sep_style=SeparatorStyle.MPT,
+    sep="<|im_end|>",
+)
+
+conv_llava_plain = Conversation(
+    system="",
+    roles=("", ""),
+    messages=(
+    ),
+    offset=0,
+    sep_style=SeparatorStyle.PLAIN,
+    sep="\n",
+)
+
+conv_llava_v0 = Conversation(
+    system="A chat between a curious human and an artificial intelligence assistant. "
+           "The assistant gives helpful, detailed, and polite answers to the human's questions.",
+    roles=("Human", "Assistant"),
+    messages=(
+    ),
+    offset=0,
+    sep_style=SeparatorStyle.SINGLE,
+    sep="###",
+)
+
+conv_llava_v0_mmtag = Conversation(
+    system="A chat between a curious user and an artificial intelligence assistant. "
+           "The assistant is able to understand the visual content that the user provides, and assist the user with a variety of tasks using natural language."
+           "The visual content will be provided with the following format: <Image>visual content</Image>.",
+    roles=("Human", "Assistant"),
+    messages=(
+    ),
+    offset=0,
+    sep_style=SeparatorStyle.SINGLE,
+    sep="###",
+    version="v0_mmtag",
+)
+
+conv_llava_v1 = Conversation(
+    system="A chat between a curious human and an artificial intelligence assistant. "
+           "The assistant gives helpful, detailed, and polite answers to the human's questions.",
+    roles=("USER", "ASSISTANT"),
+    version="v1",
+    messages=(),
+    offset=0,
+    sep_style=SeparatorStyle.TWO,
+    sep=" ",
+    sep2="</s>",
+)
+
+conv_vicuna_imgsp_v1 = Conversation(
+    system="A chat between a curious user and an artificial intelligence assistant. "
+    "The assistant gives helpful, detailed, and polite answers to the user's questions.",
+    roles=("USER", "ASSISTANT"),
+    version="imgsp_v1",
+    messages=(),
+    offset=0,
+    sep_style=SeparatorStyle.TWO,
+    sep=" ",
+    sep2="</s>",
+)
+
+conv_llava_plain_guided = Conversation(
+    system="",
+    roles=("", ""),
+    version="plain_guided",
+    messages=(
+    ),
+    offset=0,
+    sep_style=SeparatorStyle.PLAIN,
+    sep="\n",
+)
+
+conv_llava_v1_mmtag = Conversation(
+    system="A chat between a curious user and an artificial intelligence assistant. "
+           "The assistant is able to understand the visual content that the user provides, and assist the user with a variety of tasks using natural language."
+           "The visual content will be provided with the following format: <Image>visual content</Image>.",
+    roles=("USER", "ASSISTANT"),
+    messages=(),
+    offset=0,
+    sep_style=SeparatorStyle.TWO,
+    sep=" ",
+    sep2="</s>",
+    version="v1_mmtag",
+)
+
+conv_phi_2 = Conversation(
+    system="A chat between a curious user and an artificial intelligence assistant. "
+           "The assistant gives helpful, detailed, and polite answers to the user's questions.",
+    roles=("USER", "ASSISTANT"),
+    version="phi2",
+    messages=(),
+    offset=0,
+    sep_style=SeparatorStyle.TWO,
+    sep=" ",
+    sep2="<|endoftext|>",
+)
+
+conv_mistral_instruct = Conversation(
+    system="",
+    roles=("USER", "ASSISTANT"),
+    version="llama_v2",
+    messages=(),
+    offset=0,
+    sep_style=SeparatorStyle.LLAMA_2,
+    sep="<s>",
+    sep2="</s>",
+)
+
+conv_gemma = Conversation(
+    system="",
+    roles=("user", "model"),
+    version="gemma",
+    messages=(),
+    offset=0,
+    sep_style=SeparatorStyle.GEMMA,
+    sep="",
+    sep2="<eos>",
+)
+
+conv_chatml_direct = Conversation(
+    system="""<|im_start|>system
+Answer the questions.""",
+    roles=("<|im_start|>user\n", "<|im_start|>assistant\n"),
+    version="mpt",
+    messages=(),
+    offset=0,
+    sep_style=SeparatorStyle.MPT,
+    sep="<|im_end|>",
+)
+
+default_conversation = conv_vicuna_v1
+conv_templates = {
+    "default": conv_vicuna_v0,
+    "v0": conv_vicuna_v0,
+    "v1": conv_vicuna_v1,
+    "vicuna_v1": conv_vicuna_v1,
+    "phi_2": conv_phi_2,
+    "gemma": conv_gemma,
+    "llama_2": conv_llama_2,
+    "imgsp_v1": conv_vicuna_imgsp_v1,
+    "plain_guided": conv_llava_plain_guided,
+    "mistral_instruct": conv_mistral_instruct,
+    "chatml_direct": conv_chatml_direct,
+    "mistral_direct": conv_chatml_direct,
+    "plain": conv_llava_plain,
+    "v0_plain": conv_llava_plain,
+    "llava_v0": conv_llava_v0,
+    "v0_mmtag": conv_llava_v0_mmtag,
+    "llava_v1": conv_llava_v1,
+    "v1_mmtag": conv_llava_v1_mmtag,
+    "llava_llama_2": conv_llava_llama_2,
+
+    "mpt": conv_mpt,
+}
+
+
+if __name__ == "__main__":
+    print(default_conversation.get_prompt())

minigemini/eval/MathVista/calculate_score.py

@@ -0,0 +1,258 @@
+import os
+import re
+import argparse
+import pandas as pd
+
+# !pip install python-Levenshtein
+from Levenshtein import distance
+
+import sys
+sys.path.append('../')
+from utilities import *
+
+
+def get_most_similar(prediction, choices):
+    """
+    Use the Levenshtein distance (or edit distance) to determine which of the choices is most similar to the given prediction
+    """
+    distances = [distance(prediction, choice) for choice in choices]
+    ind = distances.index(min(distances))
+    return choices[ind]
+    # return min(choices, key=lambda choice: distance(prediction, choice))
+
+
+def normalize_extracted_answer(extraction, choices, question_type, answer_type, precision):
+    """
+    Normalize the extracted answer to match the answer type
+    """
+    if question_type == 'multi_choice':
+        # make sure the extraction is a string
+        if isinstance(extraction, str):
+            extraction = extraction.strip()
+        else:
+            try:
+                extraction = str(extraction)
+            except:
+                extraction = ""
+    
+        # extract "A" from "(A) text"
+        letter = re.findall(r'\(([a-zA-Z])\)', extraction)
+        if len(letter) > 0:
+            extraction = letter[0].upper()
+        
+        options = [chr(ord('A') + i) for i in range(len(choices))]
+            
+        if extraction in options:
+            # convert option letter to text, e.g. "A" -> "text"
+            ind = options.index(extraction)
+            extraction = choices[ind]
+        else:
+            # select the most similar option
+            extraction = get_most_similar(extraction, choices)
+        assert extraction in choices
+
+    elif answer_type == 'integer':
+        try:
+            extraction = str(int(float(extraction)))
+        except:
+            extraction = None
+
+    elif answer_type == 'float':
+        try:
+            extraction = str(round(float(extraction), precision))
+        except:
+            extraction = None
+        
+    elif answer_type == 'list':
+        try:
+            extraction = str(extraction)
+        except:
+            extraction = None
+
+    return extraction
+    
+
+def safe_equal(prediction, answer):
+    """
+    Check if the prediction is equal to the answer, even if they are of different types
+    """
+    try:
+        if prediction == answer:
+            return True
+        return False
+    except Exception as e:
+        print(e)
+        return False
+
+
+def get_acc_with_contion(res_pd, key, value):
+    if key == 'skills':
+        # if value in res_pd[key]:
+        total_pd = res_pd[res_pd[key].apply(lambda x: value in x)]
+    else:
+        total_pd = res_pd[res_pd[key] == value]
+
+    correct_pd = total_pd[total_pd['true_false'] == True]
+    acc = "{:.2f}".format(len(correct_pd) / len(total_pd) * 100)
+    return len(correct_pd), len(total_pd), acc
+        
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--output_file', type=str, default='output.json')
+    parser.add_argument('--score_file', type=str, default='scores.json')
+    parser.add_argument('--gt_file', type=str, default='../data/testmini.json', help='ground truth file')
+    parser.add_argument('--number', type=int, default=-1, help='number of problems to run')
+    parser.add_argument('--rerun', action='store_true', help='rerun the evaluation')
+    parser.add_argument('--caculate_gain', action='store_true', help='caculate the socre gains over random guess')
+    parser.add_argument('--random_file', type=str, default='score_random_guess.json')  
+    args = parser.parse_args()
+
+    # args
+    output_file = args.output_file
+
+    # # quick test
+    # output_file = '../results/llava-llama-2-13b/output_llava_llama_2_13b.json'
+
+    # read json
+    print(f"Reading {output_file}...")
+    results = read_json(output_file)
+
+    # read ground truth
+    print(f"Reading {args.gt_file}...")
+    gts = read_json(args.gt_file)
+
+    # full pids
+    full_pids = list(results.keys())
+    if args.number > 0:
+        full_pids = full_pids[:min(args.number, len(full_pids))]
+    print("Number of testing problems:", len(full_pids))
+    
+    ## [1] Evaluate if the prediction is true or false
+    print("\nEvaluating the predictions...")
+    update_json_flag = False
+    for pid in full_pids:
+        problem = results[pid]
+        # print(problem)
+
+        if args.rerun:
+            if 'prediction' in problem:
+                del problem['prediction']
+            if 'true_false' in problem:
+                del problem['true_false']
+
+        choices = problem['choices']
+        question_type = problem['question_type']
+        answer_type = problem['answer_type']
+        precision = problem['precision']
+        extraction = problem['extraction']
+
+        if 'answer' in problem:
+            answer = problem['answer']
+        else:
+            answer = gts[pid]['answer']
+            problem['answer'] = answer
+
+        # normalize the extracted answer to match the answer type
+        prediction = normalize_extracted_answer(extraction, choices, question_type, answer_type, precision)
+
+        # verify the prediction is true or false
+        true_false = safe_equal(prediction, answer)
+        
+        # update the problem
+        if "true_false" not in problem:
+            update_json_flag = True
+
+        elif true_false != problem['true_false']:
+            update_json_flag = True
+
+        if "prediction" not in problem:
+            update_json_flag = True
+
+        elif prediction !=  problem['prediction']:
+            update_json_flag = True
+            
+        problem['prediction'] = prediction
+        problem['true_false'] = true_false
+
+    # save the updated json
+    if update_json_flag:
+        print("\n!!!Some problems are updated.!!!")
+        print(f"\nSaving {output_file}...")
+        save_json(results, output_file)
+
+    ## [2] Calculate the average accuracy
+    total = len(full_pids)
+    correct = 0
+    for pid in full_pids:
+        if results[pid]['true_false']:
+            correct += 1
+    accuracy = str(round(correct / total * 100, 2))
+    print(f"\nCorrect: {correct}, Total: {total}, Accuracy: {accuracy}%")
+
+    scores = {"average": {"accuracy": accuracy, "correct": correct, "total": total}}
+    
+    ## [3] Calculate the fine-grained accuracy scores
+    
+    # merge the 'metadata' attribute into the data
+    for pid in results:
+        results[pid].update(results[pid].pop('metadata'))
+
+    # convert the data to a pandas DataFrame
+    df = pd.DataFrame(results).T
+
+    print(len(df))
+    print("Number of test problems:", len(df))
+    # assert len(df) == 1000 # Important!!!
+
+    # asign the target keys for evaluation
+    target_keys = ['question_type', 'answer_type', 'language', 'source', 'category', 'task', 'context', 'grade', 'skills']
+     
+    for key in target_keys:
+        print(f"\nType: [{key}]")
+        # get the unique values of the key
+        if key == 'skills':
+            # the value is a list
+            values = []
+            for i in range(len(df)):
+                values += df[key][i]
+            values = list(set(values))
+        else:
+            values = df[key].unique()
+        #print(values)
+
+        # calculate the accuracy for each value
+        scores[key] = {}
+        for value in values:
+            correct, total, acc = get_acc_with_contion(df, key, value)
+            if total > 0:
+                print(f"[{value}]: {acc}% ({correct}/{total})")
+                scores[key][value] = {"accuracy": acc, "correct": correct, "total": total}
+        
+        # sort the scores by accuracy
+        scores[key] = dict(sorted(scores[key].items(), key=lambda item: float(item[1]['accuracy']), reverse=True))
+
+    # save the scores
+    scores_file = args.score_file
+    print(f"\nSaving {scores_file}...")
+    save_json(scores, scores_file)
+    print("\nDone!")
+
+    # [4] Calculate the score gains over random guess
+    if args.caculate_gain:
+        random_file = args.random_file
+        random_scores = json.load(open(random_file))
+
+        print("\nCalculating the score gains...")
+        for key in scores:
+            if key == 'average':
+                gain = round(float(scores[key]['accuracy']) - float(random_scores[key]['accuracy']), 2)
+                scores[key]['acc_gain'] = gain
+            else:
+                for sub_key in scores[key]:
+                    gain = round(float(scores[key][sub_key]['accuracy']) - float(random_scores[key][sub_key]['accuracy']), 2)
+                    scores[key][sub_key]['acc_gain'] = str(gain)
+
+        # save the score gains
+        print(f"\nSaving {scores_file}...")    
+        save_json(scores, scores_file)
+        print("\nDone!")

minigemini/eval/MathVista/extract_answer.py

@@ -0,0 +1,160 @@
+import os
+import re
+import time
+import argparse
+
+from tqdm import tqdm
+
+import sys
+sys.path.append('../')
+from utilities import *
+
+# OpenAI
+import openai
+
+# load demo prompt
+from prompts.ext_ans import demo_prompt
+
+
+def verify_extraction(extraction):
+    extraction = extraction.strip()
+    if extraction == "" or extraction == None:
+        return False
+    return True
+
+
+def create_test_prompt(demo_prompt, query, response):
+    demo_prompt = demo_prompt.strip()
+    test_prompt = f"{query}\n\n{response}"
+    full_prompt = f"{demo_prompt}\n\n{test_prompt}\n\nExtracted answer: "
+    return full_prompt
+
+
+def extract_answer(response, problem, quick_extract=False):
+    question_type = problem['question_type']
+    answer_type = problem['answer_type']
+    choices = problem['choices']
+    query = problem['query']
+    pid = problem['pid']
+
+    if response == "":
+        return ""
+    
+    if question_type == 'multi_choice' and response in choices:
+        return response
+    
+    if answer_type == "integer":
+        try:
+            extraction = int(response)
+            return str(extraction)
+        except:
+            pass
+
+    if answer_type == "float":
+        try:
+            extraction = str(float(response))
+            return extraction
+        except:
+            pass
+
+    # quick extraction
+    if quick_extract:
+        print("Quickly extracting answer...")
+        # The answer is "text". -> "text"
+        try:
+            result = re.search(r'The answer is "(.*)"\.', response)
+            if result:
+                extraction = result.group(1)
+                return extraction
+        except:
+            pass
+
+    # general extraction
+    try:
+        full_prompt = create_test_prompt(demo_prompt, query, response)
+        extraction = get_chat_response(full_prompt, openai.api_key, openai.api_base, model=args.llm_engine)
+        return extraction
+    except Exception as e:
+        print(e)
+        print(f"Error in extracting answer for {pid}")
+
+    return ""
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    # input
+    parser.add_argument('--output_file', type=str, default='answer.json')
+    parser.add_argument('--response_label', type=str, default='response', help='response label for the input file')
+    # model
+    parser.add_argument('--llm_engine', type=str, default='gpt-4-0613', help='llm engine',
+                        choices = ['gpt-3.5-turbo', 'gpt-3.5', 'gpt-4', 'gpt-4-0314', 'gpt-4-0613'])
+    parser.add_argument('--number', type=int, default=-1, help='number of problems to run')
+    parser.add_argument('--quick_extract', action='store_true', help='use rules to extract answer for some problems')
+    parser.add_argument('--rerun', action='store_true', help='rerun the answer extraction')
+    # openai
+    parser.add_argument("--api_key", required=True, type=str, help="OpenAI API key")
+    parser.add_argument("--api_base", default=None, type=str, help="OpenAI API base")
+    # output
+    parser.add_argument('--save_every', type=int, default=10, help='save every n problems')
+    parser.add_argument('--output_label', type=str, default='', help='label for the output file')
+    args = parser.parse_args()
+
+    # args
+    label = args.response_label
+    result_file = args.output_file
+    if args.output_label != '':
+        output_file = result_file.replace('.json', f'_{args.output_label}.json')
+    else:
+        output_file = result_file
+
+    # read results
+    print(f"Reading {result_file}...")
+    try:
+        results = read_json(output_file)
+    except:
+        samples = [json.loads(line) for line in open(result_file)]
+        results = {}
+        for sample in samples:
+            results[sample['pid']] = sample
+
+    # full pids
+    full_pids = list(results.keys())
+    if args.number > 0:
+        full_pids = full_pids[:min(args.number, len(full_pids))]
+    print("Number of testing problems:", len(full_pids))
+
+    # test pids
+    if args.rerun:
+        test_pids = full_pids
+    else:
+        test_pids = []
+        for pid in full_pids:
+            # print(pid)
+            if 'extraction' not in results[pid] or not verify_extraction(results[pid]['extraction']):
+                test_pids.append(pid)
+    
+    test_num = len(test_pids)
+    print("Number of problems to run:", test_num)
+    # print(test_pids)
+
+    # openai api
+    openai.api_key = args.api_key # Your API key here
+    if args.api_base:
+        openai.api_base = args.api_base # Your API base here
+
+    # tqdm, enumerate results
+    for i, pid in enumerate(tqdm(test_pids)):
+        problem = results[pid]
+
+        assert label in problem
+        response = problem[label]       
+
+        
+        extraction  = extract_answer(response, problem, args.quick_extract)
+        results[pid]['extraction'] = extraction
+
+        if i % args.save_every == 0 or i == test_num - 1:
+            print(f"Saving results to {output_file}...")
+            save_json(results, output_file)
+            print(f"Results saved.")

minigemini/eval/MathVista/prompts/ext_ans.py

@@ -0,0 +1,42 @@
+
+
+# pids = 852,  104,  824,  506,  540
+
+demo_prompt = """
+Please read the following example. Then extract the answer from the model response and type it at the end of the prompt.
+
+Hint: Please answer the question requiring an integer answer and provide the final value, e.g., 1, 2, 3, at the end.
+Question: Which number is missing?
+
+Model response: The number missing in the sequence is 14.
+
+Extracted answer: 14
+
+Hint: Please answer the question requiring a floating-point number with one decimal place and provide the final value, e.g., 1.2, 1.3, 1.4, at the end.
+Question: What is the fraction of females facing the camera?
+
+Model response: The fraction of females facing the camera is 0.6, which means that six out of ten females in the group are facing the camera.
+
+Extracted answer: 0.6
+
+Hint: Please answer the question requiring a floating-point number with two decimal places and provide the final value, e.g., 1.23, 1.34, 1.45, at the end.
+Question: How much money does Luca need to buy a sour apple candy and a butterscotch candy? (Unit: $)
+
+Model response: Luca needs $1.45 to buy a sour apple candy and a butterscotch candy.
+
+Extracted answer: 1.45
+
+Hint: Please answer the question requiring a Python list as an answer and provide the final list, e.g., [1, 2, 3], [1.2, 1.3, 1.4], at the end.
+Question: Between which two years does the line  graph saw its maximum peak?
+
+Model response: The line graph saw its maximum peak between 2007 and 2008.
+
+Extracted answer: [2007, 2008]
+
+Hint: Please answer the question and provide the correct option letter, e.g., A, B, C, D, at the end.
+Question: What fraction of the shape is blue?\nChoices:\n(A) 3/11\n(B) 8/11\n(C) 6/11\n(D) 3/5
+
+Model response: The correct answer is (B) 8/11.
+
+Extracted answer: B
+"""

minigemini/eval/MathVista/utilities.py

@@ -0,0 +1,200 @@
+import os
+import cv2
+import json
+import time
+import pickle
+import openai
+import re
+from word2number import w2n
+
+
+def create_dir(output_dir):
+    if not os.path.exists(output_dir):
+        os.makedirs(output_dir)
+    
+
+def read_csv(file):
+    data = []
+    with open(file, 'r') as f:
+        for line in f:
+            data.append(line.strip())
+    return data
+
+
+def read_pandas_csv(csv_path):
+    # read a pandas csv sheet 
+    import pandas as pd
+    df = pd.read_csv(csv_path)
+    return df
+
+
+def read_json(path):
+    with open(path, 'r', encoding='utf-8') as f:
+        return json.load(f)
+
+
+def read_jsonl(file):
+    with open(file, 'r') as f:
+        data = [json.loads(line) for line in f]
+    return data
+
+
+def read_pickle(path):
+    with open(path, 'rb') as f:
+        return pickle.load(f)
+
+
+def save_json(data, path):
+    with open(path, 'w') as f:
+        json.dump(data, f, indent=4)
+
+
+def save_array_img(path, image):
+    cv2.imwrite(path, image)
+
+
+def contains_digit(text):
+    # check if text contains a digit
+    if any(char.isdigit() for char in text):
+        return True
+    return False  
+    
+def contains_number_word(text):
+    # check if text contains a number word
+    ignore_words = ["a", "an", "point"]
+    words = re.findall(r'\b\w+\b', text)  # This regex pattern matches any word in the text
+    for word in words:
+        if word in ignore_words:
+            continue
+        try:
+            w2n.word_to_num(word)
+            return True  # If the word can be converted to a number, return True
+        except ValueError:
+            continue  # If the word can't be converted to a number, continue with the next word
+    
+    # check if text contains a digit
+    if any(char.isdigit() for char in text):
+        return True
+
+    return False  # If none of the words could be converted to a number, return False
+
+
+def contains_quantity_word(text, special_keep_words=[]):
+    # check if text contains a quantity word
+    quantity_words = ["most", "least", "fewest"
+                      "more", "less", "fewer", 
+                      "largest", "smallest", "greatest", 
+                      "larger", "smaller", "greater", 
+                      "highest", "lowest", "higher", "lower",
+                      "increase", "decrease",
+                      "minimum", "maximum", "max", "min",
+                      "mean", "average", "median",
+                      "total", "sum", "add", "subtract",
+                      "difference", "quotient", "gap",
+                      "half", "double", "twice", "triple",
+                      "square", "cube", "root",
+                      "approximate", "approximation",
+                      "triangle", "rectangle", "circle", "square", "cube", "sphere", "cylinder", "cone", "pyramid",
+                      "multiply", "divide",
+                      "percentage", "percent", "ratio", "proportion", "fraction", "rate", 
+                    ]
+    
+    quantity_words += special_keep_words # dataset specific words
+    
+    words = re.findall(r'\b\w+\b', text)  # This regex pattern matches any word in the text
+    if any(word in quantity_words for word in words):
+        return True
+
+    return False  # If none of the words could be converted to a number, return False
+
+
+def is_bool_word(text):
+    if text in ["Yes", "No", "True", "False", 
+                "yes", "no", "true", "false", 
+                "YES", "NO", "TRUE", "FALSE"]:
+        return True
+    return False
+
+
+def is_digit_string(text):
+    # remove ".0000"
+    text = text.strip()
+    text = re.sub(r'\.0+$', '', text)
+    try:
+        int(text)
+        return True
+    except ValueError:
+        return False
+   
+    
+def is_float_string(text):
+    # text is a float string if it contains a "." and can be converted to a float
+    if "." in text:
+        try:
+            float(text)
+            return True
+        except ValueError:
+            return False
+    return False
+
+
+def copy_image(image_path, output_image_path):
+    from shutil import copyfile
+    copyfile(image_path, output_image_path)
+
+
+def copy_dir(src_dir, dst_dir):
+    from shutil import copytree
+    # copy the source directory to the target directory
+    copytree(src_dir, dst_dir)
+
+
+import PIL.Image as Image
+def get_image_size(img_path):
+    img = Image.open(img_path)
+    width, height = img.size
+    return width, height
+
+
+def get_chat_response(promot, api_key, api_base, model="gpt-3.5-turbo", temperature=0, max_tokens=256, n=1, patience=10000000,
+ sleep_time=0):
+    messages = [
+        {"role": "user", "content": promot},
+    ]
+    # print("I am here")
+    while patience > 0:
+        patience -= 1
+        try:
+            response = openai.ChatCompletion.create(model=model,
+                                                messages=messages,
+                                                api_key=api_key,
+                                                api_base=api_base,
+                                                temperature=temperature,
+                                                max_tokens=max_tokens,
+                                                n=n)
+            if n == 1:
+                prediction = response['choices'][0]['message']['content'].strip()
+                if prediction != "" and prediction != None:
+                    return prediction
+            else:
+                prediction = [choice['message']['content'].strip() for choice in response['choices']]
+                if prediction[0] != "" and prediction[0] != None:
+                    return prediction
+
+        except Exception as e:
+            if "Rate limit" not in str(e):
+                print(e)
+
+            if "Please reduce the length of the messages" in str(e):
+                print("!!Reduce promot size")
+                # reduce input prompt and keep the tail
+                new_size = int(len(promot) * 0.9)
+                new_start = len(promot) - new_size
+                promot = promot[new_start:]
+                messages = [
+                    {"role": "user", "content": promot},
+                ]
+                
+            if sleep_time > 0:
+                time.sleep(sleep_time)
+    return ""

minigemini/eval/eval_gpt_review.py

@@ -0,0 +1,113 @@
+import argparse
+import json
+import os
+
+import openai
+import tqdm
+import ray
+import time
+
+NUM_SECONDS_TO_SLEEP = 3
+
+@ray.remote(num_cpus=4)
+def get_eval(content: str, max_tokens: int):
+    while True:
+        try:
+            response = openai.ChatCompletion.create(
+                model='gpt-4',
+                messages=[{
+                    'role': 'system',
+                    'content': 'You are a helpful and precise assistant for checking the quality of the answer.'
+                }, {
+                    'role': 'user',
+                    'content': content,
+                }],
+                temperature=0.2,  # TODO: figure out which temperature is best for evaluation
+                max_tokens=max_tokens,
+            )
+            break
+        except openai.error.RateLimitError:
+            pass
+        except Exception as e:
+            print(e)
+        time.sleep(NUM_SECONDS_TO_SLEEP)
+
+    print('success!')
+    return response['choices'][0]['message']['content']
+
+
+def parse_score(review):
+    try:
+        score_pair = review.split('\n')[0]
+        score_pair = score_pair.replace(',', ' ')
+        sp = score_pair.split(' ')
+        if len(sp) == 2:
+            return [float(sp[0]), float(sp[1])]
+        else:
+            print('error', review)
+            return [-1, -1]
+    except Exception as e:
+        print(e)
+        print('error', review)
+        return [-1, -1]
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='ChatGPT-based QA evaluation.')
+    parser.add_argument('-q', '--question')
+    # parser.add_argument('-a', '--answer')
+    parser.add_argument('-a', '--answer-list', nargs='+', default=[])
+    parser.add_argument('-r', '--rule')
+    parser.add_argument('-o', '--output')
+    parser.add_argument('--max-tokens', type=int, default=1024, help='maximum number of tokens produced in the output')
+    args = parser.parse_args()
+
+    ray.init()
+
+    f_q = open(os.path.expanduser(args.question))
+    f_ans1 = open(os.path.expanduser(args.answer_list[0]))
+    f_ans2 = open(os.path.expanduser(args.answer_list[1]))
+    rule_dict = json.load(open(os.path.expanduser(args.rule), 'r'))
+
+    review_file = open(f'{args.output}', 'w')
+
+    js_list = []
+    handles = []
+    idx = 0
+    for ques_js, ans1_js, ans2_js in zip(f_q, f_ans1, f_ans2):
+        # if idx == 1:
+        #     break
+
+        ques = json.loads(ques_js)
+        ans1 = json.loads(ans1_js)
+        ans2 = json.loads(ans2_js)
+
+        category = json.loads(ques_js)['category']
+        if category in rule_dict:
+            rule = rule_dict[category]
+        else:
+            rule = rule_dict['default']
+        prompt = rule['prompt']
+        role = rule['role']
+        content = (f'[Question]\n{ques["text"]}\n\n'
+                   f'[{role} 1]\n{ans1["text"]}\n\n[End of {role} 1]\n\n'
+                   f'[{role} 2]\n{ans2["text"]}\n\n[End of {role} 2]\n\n'
+                   f'[System]\n{prompt}\n\n')
+        js_list.append({
+            'id': idx+1,
+            'question_id': ques['question_id'],
+            'answer1_id': ans1['answer_id'],
+            'answer2_id': ans2['answer_id'],
+            'category': category})
+        idx += 1
+        handles.append(get_eval.remote(content, args.max_tokens))
+        # To avoid the rate limit set by OpenAI
+        time.sleep(NUM_SECONDS_TO_SLEEP)
+
+    reviews = ray.get(handles)
+    for idx, review in enumerate(reviews):
+        scores = parse_score(review)
+        js_list[idx]['content'] = review
+        js_list[idx]['tuple'] = scores
+        review_file.write(json.dumps(js_list[idx]) + '\n')
+    review_file.close()

minigemini/eval/eval_gpt_review_bench.py

@@ -0,0 +1,121 @@
+import argparse
+import json
+import os
+
+import openai
+import time
+
+NUM_SECONDS_TO_SLEEP = 0.5
+
+
+def get_eval(content: str, max_tokens: int):
+    while True:
+        try:
+            response = openai.ChatCompletion.create(
+                model='gpt-4-0314',
+                messages=[{
+                    'role': 'system',
+                    'content': 'You are a helpful and precise assistant for checking the quality of the answer.'
+                }, {
+                    'role': 'user',
+                    'content': content,
+                }],
+                temperature=0.2,  # TODO: figure out which temperature is best for evaluation
+                max_tokens=max_tokens,
+            )
+            break
+        except openai.error.RateLimitError:
+            pass
+        except Exception as e:
+            print(e)
+        time.sleep(NUM_SECONDS_TO_SLEEP)
+
+    return response['choices'][0]['message']['content']
+
+
+def parse_score(review):
+    try:
+        score_pair = review.split('\n')[0]
+        score_pair = score_pair.replace(',', ' ')
+        sp = score_pair.split(' ')
+        if len(sp) == 2:
+            return [float(sp[0]), float(sp[1])]
+        else:
+            print('error', review)
+            return [-1, -1]
+    except Exception as e:
+        print(e)
+        print('error', review)
+        return [-1, -1]
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='ChatGPT-based QA evaluation.')
+    parser.add_argument('-q', '--question')
+    parser.add_argument('-c', '--context')
+    parser.add_argument('-a', '--answer-list', nargs='+', default=[])
+    parser.add_argument('-r', '--rule')
+    parser.add_argument('-o', '--output')
+    parser.add_argument('--max-tokens', type=int, default=1024, help='maximum number of tokens produced in the output')
+    args = parser.parse_args()
+
+    f_q = open(os.path.expanduser(args.question))
+    f_ans1 = open(os.path.expanduser(args.answer_list[0]))
+    f_ans2 = open(os.path.expanduser(args.answer_list[1]))
+    rule_dict = json.load(open(os.path.expanduser(args.rule), 'r'))
+
+    if os.path.isfile(os.path.expanduser(args.output)):
+        cur_reviews = [json.loads(line) for line in open(os.path.expanduser(args.output))]
+    else:
+        cur_reviews = []
+
+    review_file = open(f'{args.output}', 'a')
+
+    context_list = [json.loads(line) for line in open(os.path.expanduser(args.context))]
+    image_to_context = {context['image']: context for context in context_list}
+
+    handles = []
+    idx = 0
+    for ques_js, ans1_js, ans2_js in zip(f_q, f_ans1, f_ans2):
+        ques = json.loads(ques_js)
+        ans1 = json.loads(ans1_js)
+        ans2 = json.loads(ans2_js)
+
+        inst = image_to_context[ques['image']]
+
+        if isinstance(inst['caption'], list):
+            cap_str = '\n'.join(inst['caption'])
+        else:
+            cap_str = inst['caption']
+
+        category = 'llava_bench_' + json.loads(ques_js)['category']
+        if category in rule_dict:
+            rule = rule_dict[category]
+        else:
+            assert False, f"Visual QA category not found in rule file: {category}."
+        prompt = rule['prompt']
+        role = rule['role']
+        content = (f'[Context]\n{cap_str}\n\n'
+                   f'[Question]\n{ques["text"]}\n\n'
+                   f'[{role} 1]\n{ans1["text"]}\n\n[End of {role} 1]\n\n'
+                   f'[{role} 2]\n{ans2["text"]}\n\n[End of {role} 2]\n\n'
+                   f'[System]\n{prompt}\n\n')
+        cur_js = {
+            'id': idx+1,
+            'question_id': ques['question_id'],
+            'answer1_id': ans1.get('answer_id', ans1['question_id']),
+            'answer2_id': ans2.get('answer_id', ans2['answer_id']),
+            'category': category
+        }
+        if idx >= len(cur_reviews):
+            review = get_eval(content, args.max_tokens)
+            scores = parse_score(review)
+            cur_js['content'] = review
+            cur_js['tuple'] = scores
+            review_file.write(json.dumps(cur_js) + '\n')
+            review_file.flush()
+        else:
+            print(f'Skipping {idx} as we already have it.')
+        idx += 1
+        print(idx)
+    review_file.close()

minigemini/eval/eval_gpt_review_visual.py

@@ -0,0 +1,118 @@
+import argparse
+import json
+import os
+
+import openai
+import time
+
+NUM_SECONDS_TO_SLEEP = 0.5
+
+
+def get_eval(content: str, max_tokens: int):
+    while True:
+        try:
+            response = openai.ChatCompletion.create(
+                model='gpt-4-0314',
+                messages=[{
+                    'role': 'system',
+                    'content': 'You are a helpful and precise assistant for checking the quality of the answer.'
+                }, {
+                    'role': 'user',
+                    'content': content,
+                }],
+                temperature=0.2,  # TODO: figure out which temperature is best for evaluation
+                max_tokens=max_tokens,
+            )
+            break
+        except openai.error.RateLimitError:
+            pass
+        except Exception as e:
+            print(e)
+        time.sleep(NUM_SECONDS_TO_SLEEP)
+
+    return response['choices'][0]['message']['content']
+
+
+def parse_score(review):
+    try:
+        score_pair = review.split('\n')[0]
+        score_pair = score_pair.replace(',', ' ')
+        sp = score_pair.split(' ')
+        if len(sp) == 2:
+            return [float(sp[0]), float(sp[1])]
+        else:
+            print('error', review)
+            return [-1, -1]
+    except Exception as e:
+        print(e)
+        print('error', review)
+        return [-1, -1]
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='ChatGPT-based QA evaluation.')
+    parser.add_argument('-q', '--question')
+    parser.add_argument('-c', '--context')
+    parser.add_argument('-a', '--answer-list', nargs='+', default=[])
+    parser.add_argument('-r', '--rule')
+    parser.add_argument('-o', '--output')
+    parser.add_argument('--max-tokens', type=int, default=1024, help='maximum number of tokens produced in the output')
+    args = parser.parse_args()
+
+    f_q = open(os.path.expanduser(args.question))
+    f_ans1 = open(os.path.expanduser(args.answer_list[0]))
+    f_ans2 = open(os.path.expanduser(args.answer_list[1]))
+    rule_dict = json.load(open(os.path.expanduser(args.rule), 'r'))
+
+    if os.path.isfile(os.path.expanduser(args.output)):
+        cur_reviews = [json.loads(line) for line in open(os.path.expanduser(args.output))]
+    else:
+        cur_reviews = []
+
+    review_file = open(f'{args.output}', 'a')
+
+    context_list = [json.loads(line) for line in open(os.path.expanduser(args.context))]
+    image_to_context = {context['image']: context for context in context_list}
+
+    handles = []
+    idx = 0
+    for ques_js, ans1_js, ans2_js in zip(f_q, f_ans1, f_ans2):
+        ques = json.loads(ques_js)
+        ans1 = json.loads(ans1_js)
+        ans2 = json.loads(ans2_js)
+
+        inst = image_to_context[ques['image']]
+        cap_str = '\n'.join(inst['captions'])
+        box_str = '\n'.join([f'{instance["category"]}: {instance["bbox"]}' for instance in inst['instances']])
+
+        category = json.loads(ques_js)['category']
+        if category in rule_dict:
+            rule = rule_dict[category]
+        else:
+            assert False, f"Visual QA category not found in rule file: {category}."
+        prompt = rule['prompt']
+        role = rule['role']
+        content = (f'[Context]\n{cap_str}\n\n{box_str}\n\n'
+                   f'[Question]\n{ques["text"]}\n\n'
+                   f'[{role} 1]\n{ans1["text"]}\n\n[End of {role} 1]\n\n'
+                   f'[{role} 2]\n{ans2["text"]}\n\n[End of {role} 2]\n\n'
+                   f'[System]\n{prompt}\n\n')
+        cur_js = {
+            'id': idx+1,
+            'question_id': ques['question_id'],
+            'answer1_id': ans1.get('answer_id', ans1['question_id']),
+            'answer2_id': ans2.get('answer_id', ans2['answer_id']),
+            'category': category
+        }
+        if idx >= len(cur_reviews):
+            review = get_eval(content, args.max_tokens)
+            scores = parse_score(review)
+            cur_js['content'] = review
+            cur_js['tuple'] = scores
+            review_file.write(json.dumps(cur_js) + '\n')
+            review_file.flush()
+        else:
+            print(f'Skipping {idx} as we already have it.')
+        idx += 1
+        print(idx)
+    review_file.close()

minigemini/eval/eval_pope.py

@@ -0,0 +1,81 @@
+import os
+import json
+import argparse
+
+def eval_pope(answers, label_file):
+    label_list = [json.loads(q)['label'] for q in open(label_file, 'r')]
+
+    for answer in answers:
+        text = answer['text']
+
+        # Only keep the first sentence
+        if text.find('.') != -1:
+            text = text.split('.')[0]
+
+        text = text.replace(',', '')
+        words = text.split(' ')
+        if 'No' in words or 'not' in words or 'no' in words:
+            answer['text'] = 'no'
+        else:
+            answer['text'] = 'yes'
+
+    for i in range(len(label_list)):
+        if label_list[i] == 'no':
+            label_list[i] = 0
+        else:
+            label_list[i] = 1
+
+    pred_list = []
+    for answer in answers:
+        if answer['text'] == 'no':
+            pred_list.append(0)
+        else:
+            pred_list.append(1)
+
+    pos = 1
+    neg = 0
+    yes_ratio = pred_list.count(1) / len(pred_list)
+
+    TP, TN, FP, FN = 0, 0, 0, 0
+    for pred, label in zip(pred_list, label_list):
+        if pred == pos and label == pos:
+            TP += 1
+        elif pred == pos and label == neg:
+            FP += 1
+        elif pred == neg and label == neg:
+            TN += 1
+        elif pred == neg and label == pos:
+            FN += 1
+
+    print('TP\tFP\tTN\tFN\t')
+    print('{}\t{}\t{}\t{}'.format(TP, FP, TN, FN))
+
+    precision = float(TP) / float(TP + FP)
+    recall = float(TP) / float(TP + FN)
+    f1 = 2*precision*recall / (precision + recall)
+    acc = (TP + TN) / (TP + TN + FP + FN)
+    print('Accuracy: {}'.format(acc))
+    print('Precision: {}'.format(precision))
+    print('Recall: {}'.format(recall))
+    print('F1 score: {}'.format(f1))
+    print('Yes ratio: {}'.format(yes_ratio))
+    print('%.3f, %.3f, %.3f, %.3f, %.3f' % (f1, acc, precision, recall, yes_ratio) )
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--annotation-dir", type=str)
+    parser.add_argument("--question-file", type=str)
+    parser.add_argument("--result-file", type=str)
+    args = parser.parse_args()
+
+    questions = [json.loads(line) for line in open(args.question_file)]
+    questions = {question['question_id']: question for question in questions}
+    answers = [json.loads(q) for q in open(args.result_file)]
+    for file in os.listdir(args.annotation_dir):
+        assert file.startswith('coco_pope_')
+        assert file.endswith('.json')
+        category = file[10:-5]
+        cur_answers = [x for x in answers if questions[x['question_id']]['category'] == category]
+        print('Category: {}, # samples: {}'.format(category, len(cur_answers)))
+        eval_pope(cur_answers, os.path.join(args.annotation_dir, file))
+        print("====================================")

minigemini/eval/eval_science_qa.py

@@ -0,0 +1,114 @@
+import argparse
+import json
+import os
+import re
+import random
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--base-dir', type=str)
+    parser.add_argument('--result-file', type=str)
+    parser.add_argument('--output-file', type=str)
+    parser.add_argument('--output-result', type=str)
+    parser.add_argument('--split', type=str, default='test')
+    parser.add_argument('--options', type=list, default=["A", "B", "C", "D", "E"])
+    return parser.parse_args()
+
+
+def convert_caps(results):
+    fakecaps = []
+    for result in results:
+        image_id = result['question_id']
+        caption = result['text']
+        fakecaps.append({"image_id": int(image_id), "caption": caption})
+    return fakecaps
+
+
+def get_pred_idx(prediction, choices, options):
+    """
+    Get the index (e.g. 2) from the prediction (e.g. 'C')
+    """
+    if prediction in options[:len(choices)]:
+        return options.index(prediction)
+    else:
+        return -1
+        return random.choice(range(len(choices)))
+
+
+if __name__ == "__main__":
+    args = get_args()
+
+    base_dir = args.base_dir
+    split_indices = json.load(open(os.path.join(base_dir, "pid_splits.json")))[args.split]
+    problems = json.load(open(os.path.join(base_dir, "problems.json")))
+    predictions = [json.loads(line) for line in open(args.result_file)]
+    predictions = {pred['question_id']: pred for pred in predictions}
+    split_problems = {idx: problems[idx] for idx in split_indices}
+
+    results = {'correct': [], 'incorrect': []}
+    sqa_results = {}
+    sqa_results['acc'] = None
+    sqa_results['correct'] = None
+    sqa_results['count'] = None
+    sqa_results['results'] = {}
+    sqa_results['outputs'] = {}
+
+    for prob_id, prob in split_problems.items():
+        if prob_id not in predictions:
+            pred = {'text': 'FAILED', 'prompt': 'Unknown'}
+            pred_text = 'FAILED'
+        else:
+            pred = predictions[prob_id]
+            pred_text = pred['text']
+
+        if pred_text in args.options:
+            answer = pred_text
+        elif len(pred_text) >= 3 and pred_text[0] in args.options and pred_text[1:3] == ". ":
+            answer = pred_text[0]
+        else:
+            pattern = re.compile(r'The answer is ([A-Z]).')
+            res = pattern.findall(pred_text)
+            if len(res) == 1:
+                answer = res[0]  # 'A', 'B', ...
+            else:
+                answer = "FAILED"
+
+        pred_idx = get_pred_idx(answer, prob['choices'], args.options)
+
+        analysis = {
+            'question_id': prob_id,
+            'parsed_ans': answer,
+            'ground_truth': args.options[prob['answer']],
+            'question': pred['prompt'],
+            'pred': pred_text,
+            'is_multimodal': '<image>' in pred['prompt'],
+        }
+
+        sqa_results['results'][prob_id] = get_pred_idx(answer, prob['choices'], args.options)
+        sqa_results['outputs'][prob_id] = pred_text
+
+        if pred_idx == prob['answer']:
+            results['correct'].append(analysis)
+        else:
+            results['incorrect'].append(analysis)
+
+    correct = len(results['correct'])
+    total = len(results['correct']) + len(results['incorrect'])
+
+    ###### IMG ######
+    multimodal_correct = len([x for x in results['correct'] if x['is_multimodal']])
+    multimodal_incorrect = len([x for x in results['incorrect'] if x['is_multimodal']])
+    multimodal_total = multimodal_correct + multimodal_incorrect
+    ###### IMG ######
+
+    print(f'Total: {total}, Correct: {correct}, Accuracy: {correct / total * 100:.2f}%, IMG-Accuracy: {multimodal_correct / multimodal_total * 100:.2f}%')
+
+    sqa_results['acc'] = correct / total * 100
+    sqa_results['correct'] = correct
+    sqa_results['count'] = total
+
+    with open(args.output_file, 'w') as f:
+        json.dump(results, f, indent=2)
+    with open(args.output_result, 'w') as f:
+        json.dump(sqa_results, f, indent=2)

minigemini/eval/eval_science_qa_gpt4.py

@@ -0,0 +1,104 @@
+import argparse
+import json
+import os
+import re
+import random
+from collections import defaultdict
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--base-dir', type=str)
+    parser.add_argument('--gpt4-result', type=str)
+    parser.add_argument('--our-result', type=str)
+    parser.add_argument('--split', type=str, default='test')
+    parser.add_argument('--options', type=list, default=["A", "B", "C", "D", "E"])
+    return parser.parse_args()
+
+
+def convert_caps(results):
+    fakecaps = []
+    for result in results:
+        image_id = result['question_id']
+        caption = result['text']
+        fakecaps.append({"image_id": int(image_id), "caption": caption})
+    return fakecaps
+
+
+def get_pred_idx(prediction, choices, options):
+    """
+    Get the index (e.g. 2) from the prediction (e.g. 'C')
+    """
+    if prediction in options[:len(choices)]:
+        return options.index(prediction)
+    else:
+        return random.choice(range(len(choices)))
+
+
+if __name__ == "__main__":
+    args = get_args()
+
+    base_dir = args.base_dir
+    split_indices = json.load(open(os.path.join(base_dir, "pid_splits.json")))[args.split]
+    problems = json.load(open(os.path.join(base_dir, "problems.json")))
+    our_predictions = [json.loads(line) for line in open(args.our_result)]
+    our_predictions = {pred['question_id']: pred for pred in our_predictions}
+    split_problems = {idx: problems[idx] for idx in split_indices}
+
+    gpt4_predictions = json.load(open(args.gpt4_result))['outputs']
+
+    results = defaultdict(lambda: 0)
+
+    for prob_id, prob in split_problems.items():
+        if prob_id not in our_predictions:
+            continue
+        if prob_id not in gpt4_predictions:
+            continue
+        our_pred = our_predictions[prob_id]['text']
+        gpt4_pred = gpt4_predictions[prob_id]
+
+        pattern = re.compile(r'The answer is ([A-Z]).')
+        our_res = pattern.findall(our_pred)
+        if len(our_res) == 1:
+            our_answer = our_res[0]  # 'A', 'B', ...
+        else:
+            our_answer = "FAILED"
+        gpt4_res = pattern.findall(gpt4_pred)
+        if len(gpt4_res) == 1:
+            gpt4_answer = gpt4_res[0]  # 'A', 'B', ...
+        else:
+            gpt4_answer = "FAILED"
+
+        our_pred_idx = get_pred_idx(our_answer, prob['choices'], args.options)
+        gpt4_pred_idx = get_pred_idx(gpt4_answer, prob['choices'], args.options)
+
+        if gpt4_answer == 'FAILED':
+            results['gpt4_failed'] += 1
+            # continue
+            gpt4_pred_idx = our_pred_idx
+            # if our_pred_idx != prob['answer']:
+            #     print(our_predictions[prob_id]['prompt'])
+            #     print('-----------------')
+            #     print(f'LECTURE: {prob["lecture"]}')
+            #     print(f'SOLUTION: {prob["solution"]}')
+            #     print('=====================')
+        else:
+            # continue
+            pass
+        # gpt4_pred_idx = our_pred_idx
+
+        if gpt4_pred_idx == prob['answer']:
+            results['correct'] += 1
+        else:
+            results['incorrect'] += 1
+
+
+        if gpt4_pred_idx == prob['answer'] or our_pred_idx == prob['answer']:
+            results['correct_upperbound'] += 1
+
+    correct = results['correct']
+    total = results['correct'] + results['incorrect']
+    print(f'Total: {total}, Correct: {correct}, Accuracy: {correct / total * 100:.2f}%')
+    print(f'Total: {total}, Correct (upper): {results["correct_upperbound"]}, Accuracy: {results["correct_upperbound"] / total * 100:.2f}%')
+    print(f'Total: {total}, GPT-4 NO-ANS (RANDOM): {results["gpt4_failed"]}, Percentage: {results["gpt4_failed"] / total * 100:.2f}%')
+

minigemini/eval/eval_science_qa_gpt4_requery.py

@@ -0,0 +1,149 @@
+import argparse
+import json
+import os
+import re
+import random
+from collections import defaultdict
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--base-dir', type=str)
+    parser.add_argument('--gpt4-result', type=str)
+    parser.add_argument('--requery-result', type=str)
+    parser.add_argument('--our-result', type=str)
+    parser.add_argument('--output-result', type=str)
+    parser.add_argument('--split', type=str, default='test')
+    parser.add_argument('--options', type=list, default=["A", "B", "C", "D", "E"])
+    return parser.parse_args()
+
+
+def convert_caps(results):
+    fakecaps = []
+    for result in results:
+        image_id = result['question_id']
+        caption = result['text']
+        fakecaps.append({"image_id": int(image_id), "caption": caption})
+    return fakecaps
+
+
+def get_pred_idx(prediction, choices, options):
+    """
+    Get the index (e.g. 2) from the prediction (e.g. 'C')
+    """
+    if prediction in options[:len(choices)]:
+        return options.index(prediction)
+    else:
+        return random.choice(range(len(choices)))
+
+
+if __name__ == "__main__":
+    args = get_args()
+
+    base_dir = args.base_dir
+    split_indices = json.load(open(os.path.join(base_dir, "pid_splits.json")))[args.split]
+    problems = json.load(open(os.path.join(base_dir, "problems.json")))
+    our_predictions = [json.loads(line) for line in open(args.our_result)]
+    our_predictions = {pred['question_id']: pred for pred in our_predictions}
+    split_problems = {idx: problems[idx] for idx in split_indices}
+
+    requery_predictions = [json.loads(line) for line in open(args.requery_result)]
+    requery_predictions = {pred['question_id']: pred for pred in requery_predictions}
+
+    gpt4_predictions = json.load(open(args.gpt4_result))['outputs']
+
+    results = defaultdict(lambda: 0)
+
+    sqa_results = {}
+    sqa_results['acc'] = None
+    sqa_results['correct'] = None
+    sqa_results['count'] = None
+    sqa_results['results'] = {}
+    sqa_results['outputs'] = {}
+
+    for prob_id, prob in split_problems.items():
+        if prob_id not in our_predictions:
+            assert False
+        if prob_id not in gpt4_predictions:
+            assert False
+        our_pred = our_predictions[prob_id]['text']
+        gpt4_pred = gpt4_predictions[prob_id]
+        if prob_id not in requery_predictions:
+            results['missing_requery'] += 1
+            requery_pred = "MISSING"
+        else:
+            requery_pred = requery_predictions[prob_id]['text']
+
+        pattern = re.compile(r'The answer is ([A-Z]).')
+        our_res = pattern.findall(our_pred)
+        if len(our_res) == 1:
+            our_answer = our_res[0]  # 'A', 'B', ...
+        else:
+            our_answer = "FAILED"
+
+        requery_res = pattern.findall(requery_pred)
+        if len(requery_res) == 1:
+            requery_answer = requery_res[0]  # 'A', 'B', ...
+        else:
+            requery_answer = "FAILED"
+
+        gpt4_res = pattern.findall(gpt4_pred)
+        if len(gpt4_res) == 1:
+            gpt4_answer = gpt4_res[0]  # 'A', 'B', ...
+        else:
+            gpt4_answer = "FAILED"
+
+        our_pred_idx = get_pred_idx(our_answer, prob['choices'], args.options)
+        gpt4_pred_idx = get_pred_idx(gpt4_answer, prob['choices'], args.options)
+        requery_pred_idx = get_pred_idx(requery_answer, prob['choices'], args.options)
+
+        results['total'] += 1
+
+        if gpt4_answer == 'FAILED':
+            results['gpt4_failed'] += 1
+            if gpt4_pred_idx == prob['answer']:
+                results['gpt4_correct'] += 1
+            if our_pred_idx == prob['answer']:
+                results['gpt4_ourvisual_correct'] += 1
+        elif gpt4_pred_idx == prob['answer']:
+            results['gpt4_correct'] += 1
+            results['gpt4_ourvisual_correct'] += 1
+
+        if our_pred_idx == prob['answer']:
+            results['our_correct'] += 1
+
+        if requery_answer == 'FAILED':
+            sqa_results['results'][prob_id] = our_pred_idx
+            if our_pred_idx == prob['answer']:
+                results['requery_correct'] += 1
+        else:
+            sqa_results['results'][prob_id] = requery_pred_idx
+            if requery_pred_idx == prob['answer']:
+                results['requery_correct'] += 1
+            else:
+                print(f"""
+Question ({args.options[prob['answer']]}): {our_predictions[prob_id]['prompt']}
+Our ({our_answer}): {our_pred}
+GPT-4 ({gpt4_answer}): {gpt4_pred}
+Requery ({requery_answer}): {requery_pred}
+print("=====================================")
+""")
+
+        if gpt4_pred_idx == prob['answer'] or our_pred_idx == prob['answer']:
+            results['correct_upperbound'] += 1
+
+    total = results['total']
+    print(f'Total: {total}, Our-Correct: {results["our_correct"]}, Accuracy: {results["our_correct"] / total * 100:.2f}%')
+    print(f'Total: {total}, GPT-4-Correct: {results["gpt4_correct"]}, Accuracy: {results["gpt4_correct"] / total * 100:.2f}%')
+    print(f'Total: {total}, GPT-4 NO-ANS (RANDOM): {results["gpt4_failed"]}, Percentage: {results["gpt4_failed"] / total * 100:.2f}%')
+    print(f'Total: {total}, GPT-4-OursVisual-Correct: {results["gpt4_ourvisual_correct"]}, Accuracy: {results["gpt4_ourvisual_correct"] / total * 100:.2f}%')
+    print(f'Total: {total}, Requery-Correct: {results["requery_correct"]}, Accuracy: {results["requery_correct"] / total * 100:.2f}%')
+    print(f'Total: {total}, Correct upper: {results["correct_upperbound"]}, Accuracy: {results["correct_upperbound"] / total * 100:.2f}%')
+
+    sqa_results['acc'] = results["requery_correct"] / total * 100
+    sqa_results['correct'] = results["requery_correct"]
+    sqa_results['count'] = total
+
+    with open(args.output_result, 'w') as f:
+        json.dump(sqa_results, f, indent=2)
+

minigemini/eval/eval_textvqa.py

@@ -0,0 +1,65 @@
+import os
+import argparse
+import json
+import re
+
+from minigemini.eval.m4c_evaluator import TextVQAAccuracyEvaluator
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--annotation-file', type=str)
+    parser.add_argument('--result-file', type=str)
+    parser.add_argument('--result-dir', type=str)
+    return parser.parse_args()
+
+
+def prompt_processor(prompt):
+    if prompt.startswith('OCR tokens: '):
+        pattern = r"Question: (.*?) Short answer:"
+        match = re.search(pattern, prompt, re.DOTALL)
+        question = match.group(1)
+    elif 'Reference OCR token: ' in prompt and len(prompt.split('\n')) == 3:
+        if prompt.startswith('Reference OCR token:'):
+            question = prompt.split('\n')[1]
+        else:
+            question = prompt.split('\n')[0]
+    elif len(prompt.split('\n')) == 2:
+        question = prompt.split('\n')[0]
+    else:
+        assert False
+
+    return question.lower()
+
+
+def eval_single(annotation_file, result_file):
+    experiment_name = os.path.splitext(os.path.basename(result_file))[0]
+    print(experiment_name)
+    annotations = json.load(open(annotation_file))['data']
+    annotations = {(annotation['image_id'], annotation['question'].lower()): annotation for annotation in annotations}
+    results = [json.loads(line) for line in open(result_file)]
+
+    pred_list = []
+    for result in results:
+        annotation = annotations[(result['question_id'], prompt_processor(result['prompt']))]
+        pred_list.append({
+            "pred_answer": result['text'],
+            "gt_answers": annotation['answers'],
+        })
+
+    evaluator = TextVQAAccuracyEvaluator()
+    print('Samples: {}\nAccuracy: {:.2f}%\n'.format(len(pred_list), 100. * evaluator.eval_pred_list(pred_list)))
+
+
+if __name__ == "__main__":
+    args = get_args()
+
+    if args.result_file is not None:
+        eval_single(args.annotation_file, args.result_file)
+
+    if args.result_dir is not None:
+        for result_file in sorted(os.listdir(args.result_dir)):
+            if not result_file.endswith('.jsonl'):
+                print(f'Skipping {result_file}')
+                continue
+            eval_single(args.annotation_file, os.path.join(args.result_dir, result_file))

minigemini/eval/generate_webpage_data_from_table.py

@@ -0,0 +1,111 @@
+"""Generate json file for webpage."""
+import json
+import os
+import re
+
+# models = ['llama', 'alpaca', 'gpt35', 'bard']
+models = ['vicuna']
+
+
+def read_jsonl(path: str, key: str=None):
+    data = []
+    with open(os.path.expanduser(path)) as f:
+        for line in f:
+            if not line:
+                continue
+            data.append(json.loads(line))
+    if key is not None:
+        data.sort(key=lambda x: x[key])
+        data = {item[key]: item for item in data}
+    return data
+
+
+def trim_hanging_lines(s: str, n: int) -> str:
+    s = s.strip()
+    for _ in range(n):
+        s = s.split('\n', 1)[1].strip()
+    return s
+
+
+if __name__ == '__main__':
+    questions = read_jsonl('table/question.jsonl', key='question_id')
+
+    # alpaca_answers = read_jsonl('table/answer/answer_alpaca-13b.jsonl', key='question_id')
+    # bard_answers = read_jsonl('table/answer/answer_bard.jsonl', key='question_id')
+    # gpt35_answers = read_jsonl('table/answer/answer_gpt35.jsonl', key='question_id')
+    # llama_answers = read_jsonl('table/answer/answer_llama-13b.jsonl', key='question_id')
+    vicuna_answers = read_jsonl('table/answer/answer_vicuna-13b.jsonl', key='question_id')
+    ours_answers = read_jsonl('table/results/llama-13b-hf-alpaca.jsonl', key='question_id')
+
+    review_vicuna = read_jsonl('table/review/review_vicuna-13b_llama-13b-hf-alpaca.jsonl', key='question_id')
+    # review_alpaca = read_jsonl('table/review/review_alpaca-13b_vicuna-13b.jsonl', key='question_id')
+    # review_bard = read_jsonl('table/review/review_bard_vicuna-13b.jsonl', key='question_id')
+    # review_gpt35 = read_jsonl('table/review/review_gpt35_vicuna-13b.jsonl', key='question_id')
+    # review_llama = read_jsonl('table/review/review_llama-13b_vicuna-13b.jsonl', key='question_id')
+
+    records = []
+    for qid in questions.keys():
+        r = {
+            'id': qid,
+            'category': questions[qid]['category'],
+            'question': questions[qid]['text'],
+            'answers': {
+                # 'alpaca': alpaca_answers[qid]['text'],
+                # 'llama': llama_answers[qid]['text'],
+                # 'bard': bard_answers[qid]['text'],
+                # 'gpt35': gpt35_answers[qid]['text'],
+                'vicuna': vicuna_answers[qid]['text'],
+                'ours': ours_answers[qid]['text'],
+            },
+            'evaluations': {
+                # 'alpaca': review_alpaca[qid]['text'],
+                # 'llama': review_llama[qid]['text'],
+                # 'bard': review_bard[qid]['text'],
+                'vicuna': review_vicuna[qid]['content'],
+                # 'gpt35': review_gpt35[qid]['text'],
+            },
+            'scores': {
+                'vicuna': review_vicuna[qid]['tuple'],
+                # 'alpaca': review_alpaca[qid]['score'],
+                # 'llama': review_llama[qid]['score'],
+                # 'bard': review_bard[qid]['score'],
+                # 'gpt35': review_gpt35[qid]['score'],
+            },
+        }
+
+        # cleanup data
+        cleaned_evals = {}
+        for k, v in r['evaluations'].items():
+            v = v.strip()
+            lines = v.split('\n')
+            # trim the first line if it's a pair of numbers
+            if re.match(r'\d+[, ]+\d+', lines[0]):
+                lines = lines[1:]
+            v = '\n'.join(lines)
+            cleaned_evals[k] = v.replace('Assistant 1', "**Assistant 1**").replace('Assistant 2', '**Assistant 2**')
+
+        r['evaluations'] = cleaned_evals
+        records.append(r)
+
+    # Reorder the records, this is optional
+    for r in records:
+        if r['id'] <= 20:
+            r['id'] += 60
+        else:
+            r['id'] -= 20
+    for r in records:
+        if r['id'] <= 50:
+            r['id'] += 10
+        elif 50 < r['id'] <= 60:
+            r['id'] -= 50
+    for r in records:
+        if r['id'] == 7:
+            r['id'] = 1
+        elif r['id'] < 7:
+            r['id'] += 1 
+
+    records.sort(key=lambda x: x['id'])
+
+    # Write to file
+    with open('webpage/data.json', 'w') as f:
+        json.dump({'questions': records, 'models': models}, f, indent=2)

minigemini/eval/m4c_evaluator.py

@@ -0,0 +1,334 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+import re
+
+from tqdm import tqdm
+
+
+class EvalAIAnswerProcessor:
+    """
+    Processes an answer similar to Eval AI
+        copied from
+        https://github.com/facebookresearch/mmf/blob/c46b3b3391275b4181567db80943473a89ab98ab/pythia/tasks/processors.py#L897
+    """
+
+    CONTRACTIONS = {
+        "aint": "ain't",
+        "arent": "aren't",
+        "cant": "can't",
+        "couldve": "could've",
+        "couldnt": "couldn't",
+        "couldn'tve": "couldn't've",
+        "couldnt've": "couldn't've",
+        "didnt": "didn't",
+        "doesnt": "doesn't",
+        "dont": "don't",
+        "hadnt": "hadn't",
+        "hadnt've": "hadn't've",
+        "hadn'tve": "hadn't've",
+        "hasnt": "hasn't",
+        "havent": "haven't",
+        "hed": "he'd",
+        "hed've": "he'd've",
+        "he'dve": "he'd've",
+        "hes": "he's",
+        "howd": "how'd",
+        "howll": "how'll",
+        "hows": "how's",
+        "Id've": "I'd've",
+        "I'dve": "I'd've",
+        "Im": "I'm",
+        "Ive": "I've",
+        "isnt": "isn't",
+        "itd": "it'd",
+        "itd've": "it'd've",
+        "it'dve": "it'd've",
+        "itll": "it'll",
+        "let's": "let's",
+        "maam": "ma'am",
+        "mightnt": "mightn't",
+        "mightnt've": "mightn't've",
+        "mightn'tve": "mightn't've",
+        "mightve": "might've",
+        "mustnt": "mustn't",
+        "mustve": "must've",
+        "neednt": "needn't",
+        "notve": "not've",
+        "oclock": "o'clock",
+        "oughtnt": "oughtn't",
+        "ow's'at": "'ow's'at",
+        "'ows'at": "'ow's'at",
+        "'ow'sat": "'ow's'at",
+        "shant": "shan't",
+        "shed've": "she'd've",
+        "she'dve": "she'd've",
+        "she's": "she's",
+        "shouldve": "should've",
+        "shouldnt": "shouldn't",
+        "shouldnt've": "shouldn't've",
+        "shouldn'tve": "shouldn't've",
+        "somebody'd": "somebodyd",
+        "somebodyd've": "somebody'd've",
+        "somebody'dve": "somebody'd've",
+        "somebodyll": "somebody'll",
+        "somebodys": "somebody's",
+        "someoned": "someone'd",
+        "someoned've": "someone'd've",
+        "someone'dve": "someone'd've",
+        "someonell": "someone'll",
+        "someones": "someone's",
+        "somethingd": "something'd",
+        "somethingd've": "something'd've",
+        "something'dve": "something'd've",
+        "somethingll": "something'll",
+        "thats": "that's",
+        "thered": "there'd",
+        "thered've": "there'd've",
+        "there'dve": "there'd've",
+        "therere": "there're",
+        "theres": "there's",
+        "theyd": "they'd",
+        "theyd've": "they'd've",
+        "they'dve": "they'd've",
+        "theyll": "they'll",
+        "theyre": "they're",
+        "theyve": "they've",
+        "twas": "'twas",
+        "wasnt": "wasn't",
+        "wed've": "we'd've",
+        "we'dve": "we'd've",
+        "weve": "we've",
+        "werent": "weren't",
+        "whatll": "what'll",
+        "whatre": "what're",
+        "whats": "what's",
+        "whatve": "what've",
+        "whens": "when's",
+        "whered": "where'd",
+        "wheres": "where's",
+        "whereve": "where've",
+        "whod": "who'd",
+        "whod've": "who'd've",
+        "who'dve": "who'd've",
+        "wholl": "who'll",
+        "whos": "who's",
+        "whove": "who've",
+        "whyll": "why'll",
+        "whyre": "why're",
+        "whys": "why's",
+        "wont": "won't",
+        "wouldve": "would've",
+        "wouldnt": "wouldn't",
+        "wouldnt've": "wouldn't've",
+        "wouldn'tve": "wouldn't've",
+        "yall": "y'all",
+        "yall'll": "y'all'll",
+        "y'allll": "y'all'll",
+        "yall'd've": "y'all'd've",
+        "y'alld've": "y'all'd've",
+        "y'all'dve": "y'all'd've",
+        "youd": "you'd",
+        "youd've": "you'd've",
+        "you'dve": "you'd've",
+        "youll": "you'll",
+        "youre": "you're",
+        "youve": "you've",
+    }
+
+    NUMBER_MAP = {
+        "none": "0",
+        "zero": "0",
+        "one": "1",
+        "two": "2",
+        "three": "3",
+        "four": "4",
+        "five": "5",
+        "six": "6",
+        "seven": "7",
+        "eight": "8",
+        "nine": "9",
+        "ten": "10",
+    }
+    ARTICLES = ["a", "an", "the"]
+    PERIOD_STRIP = re.compile(r"(?!<=\d)(\.)(?!\d)")
+    COMMA_STRIP = re.compile(r"(?<=\d)(\,)+(?=\d)")
+    PUNCTUATIONS = [
+        ";",
+        r"/",
+        "[",
+        "]",
+        '"',
+        "{",
+        "}",
+        "(",
+        ")",
+        "=",
+        "+",
+        "\\",
+        "_",
+        "-",
+        ">",
+        "<",
+        "@",
+        "`",
+        ",",
+        "?",
+        "!",
+    ]
+
+    def __init__(self, *args, **kwargs):
+        pass
+
+    def word_tokenize(self, word):
+        word = word.lower()
+        word = word.replace(",", "").replace("?", "").replace("'s", " 's")
+        return word.strip()
+
+    def process_punctuation(self, in_text):
+        out_text = in_text
+        for p in self.PUNCTUATIONS:
+            if (p + " " in in_text or " " + p in in_text) or (
+                re.search(self.COMMA_STRIP, in_text) is not None
+            ):
+                out_text = out_text.replace(p, "")
+            else:
+                out_text = out_text.replace(p, " ")
+        out_text = self.PERIOD_STRIP.sub("", out_text, re.UNICODE)
+        return out_text
+
+    def process_digit_article(self, in_text):
+        out_text = []
+        temp_text = in_text.lower().split()
+        for word in temp_text:
+            word = self.NUMBER_MAP.setdefault(word, word)
+            if word not in self.ARTICLES:
+                out_text.append(word)
+            else:
+                pass
+        for word_id, word in enumerate(out_text):
+            if word in self.CONTRACTIONS:
+                out_text[word_id] = self.CONTRACTIONS[word]
+        out_text = " ".join(out_text)
+        return out_text
+
+    def __call__(self, item):
+        item = self.word_tokenize(item)
+        item = item.replace("\n", " ").replace("\t", " ").strip()
+        item = self.process_punctuation(item)
+        item = self.process_digit_article(item)
+        return item
+
+
+class TextVQAAccuracyEvaluator:
+    def __init__(self):
+        self.answer_processor = EvalAIAnswerProcessor()
+
+    def _compute_answer_scores(self, raw_answers):
+        """
+        compute the accuracy (soft score) of human answers
+        """
+        answers = [self.answer_processor(a) for a in raw_answers]
+        assert len(answers) == 10
+        gt_answers = list(enumerate(answers))
+        unique_answers = set(answers)
+        unique_answer_scores = {}
+
+        for unique_answer in unique_answers:
+            accs = []
+            for gt_answer in gt_answers:
+                other_answers = [item for item in gt_answers if item != gt_answer]
+                matching_answers = [
+                    item for item in other_answers if item[1] == unique_answer
+                ]
+                acc = min(1, float(len(matching_answers)) / 3)
+                accs.append(acc)
+            unique_answer_scores[unique_answer] = sum(accs) / len(accs)
+
+        return unique_answer_scores
+
+    def eval_pred_list(self, pred_list):
+        pred_scores = []
+        for entry in tqdm(pred_list):
+            pred_answer = self.answer_processor(entry["pred_answer"])
+            unique_answer_scores = self._compute_answer_scores(entry["gt_answers"])
+            score = unique_answer_scores.get(pred_answer, 0.0)
+            pred_scores.append(score)
+
+        accuracy = sum(pred_scores) / len(pred_scores)
+        return accuracy
+
+
+class STVQAAccuracyEvaluator:
+    def __init__(self):
+        self.answer_processor = EvalAIAnswerProcessor()
+
+    def eval_pred_list(self, pred_list):
+        pred_scores = []
+        for entry in pred_list:
+            pred_answer = self.answer_processor(entry["pred_answer"])
+            gts = [self.answer_processor(a) for a in entry["gt_answers"]]
+            score = 1.0 if pred_answer in gts else 0.0
+            pred_scores.append(score)
+
+        accuracy = sum(pred_scores) / len(pred_scores)
+        return accuracy
+
+
+class STVQAANLSEvaluator:
+    def __init__(self):
+        import editdistance  # install with `pip install editdistance`
+
+        self.get_edit_distance = editdistance.eval
+
+    def get_anls(self, s1, s2):
+        s1 = s1.lower().strip()
+        s2 = s2.lower().strip()
+        iou = 1 - self.get_edit_distance(s1, s2) / max(len(s1), len(s2))
+        anls = iou if iou >= 0.5 else 0.0
+        return anls
+
+    def eval_pred_list(self, pred_list):
+        pred_scores = []
+        for entry in pred_list:
+            anls = max(
+                self.get_anls(entry["pred_answer"], gt) for gt in entry["gt_answers"]
+            )
+            pred_scores.append(anls)
+
+        accuracy = sum(pred_scores) / len(pred_scores)
+        return accuracy
+
+
+class TextCapsBleu4Evaluator:
+    def __init__(self):
+        # The following script requires Java 1.8.0 and pycocotools installed.
+        # The pycocoevalcap can be installed with pip as
+        # pip install git+https://github.com/ronghanghu/coco-caption.git@python23
+        # Original pycocoevalcap code is at https://github.com/tylin/coco-caption
+        # but has no python3 support yet.
+        try:
+            from pycocoevalcap.bleu.bleu import Bleu
+            from pycocoevalcap.tokenizer.ptbtokenizer import PTBTokenizer
+        except ModuleNotFoundError:
+            print(
+                "Please install pycocoevalcap module using "
+                "pip install git+https://github.com/ronghanghu/coco-caption.git@python23"  # noqa
+            )
+            raise
+
+        self.tokenizer = PTBTokenizer()
+        self.scorer = Bleu(4)
+
+    def eval_pred_list(self, pred_list):
+        # Create reference and hypotheses captions.
+        gts = {}
+        res = {}
+        for idx, entry in enumerate(pred_list):
+            gts[idx] = [{"caption": a} for a in entry["gt_answers"]]
+            res[idx] = [{"caption": entry["pred_answer"]}]
+
+        gts = self.tokenizer.tokenize(gts)
+        res = self.tokenizer.tokenize(res)
+        score, _ = self.scorer.compute_score(gts, res)
+
+        bleu4 = score[3]  # score is (Bleu-1, Bleu-2, Bleu-3, Bleu-4)
+        return bleu4

minigemini/eval/model_math_vista.py

@@ -0,0 +1,237 @@
+import argparse
+import torch
+import os
+import json
+from tqdm import tqdm
+
+from minigemini.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+from minigemini.conversation import conv_templates, SeparatorStyle
+from minigemini.model.builder import load_pretrained_model
+from minigemini.utils import disable_torch_init
+from minigemini.mm_utils import tokenizer_image_token, process_images, get_model_name_from_path
+
+from PIL import Image
+import math
+
+def split_list(lst, n):
+    """Split a list into n (roughly) equal-sized chunks"""
+    chunk_size = math.ceil(len(lst) / n)  # integer division
+    return [lst[i:i+chunk_size] for i in range(0, len(lst), chunk_size)]
+
+
+def get_chunk(lst, n, k):
+    chunks = split_list(lst, n)
+    return chunks[k]
+
+
+def create_one_query(problem, shot_num, shot_type, use_caption):
+
+
+    ### [1] Demo prompt
+    demo_prompt = ""
+
+    ### [2] Test query
+    # problem info
+    question = problem['question']
+    unit = problem['unit']
+    choices = problem['choices']
+    # caption = problem['caption']
+    precision = problem['precision']
+    question_type = problem['question_type']
+    answer_type = problem['answer_type']
+
+    # hint
+    if shot_type == 'solution':
+        if question_type == "multi_choice":
+            assert answer_type == "text"
+            hint_text = f"Hint: Please answer the question and provide the correct option letter, e.g., A, B, C, D, at the end."
+        else:
+            assert answer_type in ["integer", "float", "list"]
+            if answer_type == "integer":
+                hint_text = f"Hint: Please answer the question requiring an integer answer and provide the final value, e.g., 1, 2, 3, at the end."
+            
+            elif answer_type == "float" and precision == 1:
+                hint_text = f"Hint: Please answer the question requiring a floating-point number with one decimal place and provide the final value, e.g., 1.2, 1.3, 1.4, at the end."
+            
+            elif answer_type == "float" and precision == 2:
+                hint_text = f"Hint: Please answer the question requiring a floating-point number with two decimal places and provide the final value, e.g., 1.23, 1.34, 1.45, at the end."
+            
+            elif answer_type == "list":
+                hint_text = f"Hint: Please answer the question requiring a Python list as an answer and provide the final list, e.g., [1, 2, 3], [1.2, 1.3, 1.4], at the end."
+    else:
+        assert shot_type == 'code'
+        hint_text = "Hint: Please generate a python code to solve the problem"
+
+    # question
+    question_text = f"Question: {question}"
+    if unit:
+        question_text += f" (Unit: {unit})"
+
+    # choices
+    if choices:
+        # choices: (A) 1.2 (B) 1.3 (C) 1.4 (D) 1.5
+        texts = ["Choices:"]
+        for i, choice in enumerate(choices):
+            texts.append(f"({chr(ord('A')+i)}) {choice}")
+        choices_text = "\n".join(texts)
+    else:
+        choices_text = ""
+
+    # prompt
+    if shot_type == 'solution':
+        prompt = "Solution: "
+    else:
+        assert shot_type == 'code'
+        prompt = "Python code: "
+    
+    elements = [hint_text, question_text, choices_text]
+    test_query = "\n".join([e for e in elements if e != ""])
+
+    ### [3] Final query
+    query = demo_prompt + "\n\n" + test_query
+    query = query.strip()
+    return query
+
+
+def eval_model(args):
+    # Model
+    disable_torch_init()
+    model_path = os.path.expanduser(args.model_path)
+    model_name = get_model_name_from_path(model_path)
+    tokenizer, model, image_processor, context_len = load_pretrained_model(model_path, args.model_base, model_name,
+                                                                           load_8bit=args.load_8bit)
+
+    questions = json.load(open(os.path.expanduser(args.question_file), "r"))
+    questions = [dict(pid=pid, info=qs) for pid, qs in questions.items()]
+    questions = get_chunk(questions, args.num_chunks, args.chunk_idx)
+
+    answers_file = os.path.expanduser(args.answers_file)
+    os.makedirs(os.path.dirname(answers_file), exist_ok=True)
+
+    if os.path.exists(answers_file):
+        file = open(answers_file, "r")
+        pred_contents = [json.loads(line) for line in file]
+        done_pid = [sample['pid'] for sample in pred_contents]
+    else:
+        done_pid = []
+    ans_file = open(answers_file, "a")
+
+    for i, line in enumerate(tqdm(questions)):
+        idx = line['pid']
+        info = line['info']
+        if idx in done_pid:
+            continue
+
+        qs = create_one_query(
+            problem = info, 
+            shot_num = 0,
+            shot_type = 'solution',
+            use_caption = False,
+        )
+        query = qs
+
+        if 'image' in info:
+            image_file = info["image"]
+            image = Image.open(os.path.join(args.image_folder, image_file))
+            
+            if hasattr(model.config, 'image_size_aux'):
+                if not hasattr(image_processor, 'image_size_raw'):
+                    image_processor.image_size_raw = image_processor.crop_size.copy()
+                image_processor.crop_size['height'] = model.config.image_size_aux
+                image_processor.crop_size['width'] = model.config.image_size_aux
+                image_processor.size['shortest_edge'] = model.config.image_size_aux
+            
+            image_tensor = process_images([image], image_processor, model.config)[0]
+            
+            image_grid = getattr(model.config, 'image_grid', 1)
+            if hasattr(model.config, 'image_size_aux'):
+                raw_shape = [image_processor.image_size_raw['height'] * image_grid, 
+                            image_processor.image_size_raw['width'] * image_grid]
+                image_tensor_aux = image_tensor
+                image_tensor = torch.nn.functional.interpolate(image_tensor[None], 
+                                                            size=raw_shape, 
+                                                            mode='bilinear', 
+                                                            align_corners=False)[0]
+            else:
+                image_tensor_aux = []
+
+            if image_grid >= 2:            
+                raw_image = image_tensor.reshape(3, 
+                                                image_grid,
+                                                image_processor.image_size_raw['height'],
+                                                image_grid,
+                                                image_processor.image_size_raw['width'])
+                raw_image = raw_image.permute(1, 3, 0, 2, 4)
+                raw_image = raw_image.reshape(-1, 3,
+                                            image_processor.image_size_raw['height'],
+                                            image_processor.image_size_raw['width'])
+                
+                if getattr(model.config, 'image_global', False):
+                    global_image = image_tensor
+                    if len(global_image.shape) == 3:
+                        global_image = global_image[None]
+                    global_image = torch.nn.functional.interpolate(global_image, 
+                                                            size=[image_processor.image_size_raw['height'],
+                                                                image_processor.image_size_raw['width']], 
+                                                            mode='bilinear', 
+                                                            align_corners=False)
+                    # [image_crops, image_global]
+                    raw_image = torch.cat([raw_image, global_image], dim=0)
+                image_tensor = raw_image.contiguous()
+            
+            images = image_tensor[None].to(dtype=model.dtype, device='cuda', non_blocking=True)
+            images_aux = image_tensor_aux[None].to(dtype=model.dtype, device='cuda', non_blocking=True) if len(image_tensor_aux)>0 else None
+            if getattr(model.config, 'mm_use_im_start_end', False):
+                qs = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\n' + qs
+            else:
+                qs = DEFAULT_IMAGE_TOKEN + '\n' + qs
+        else:
+            images = None
+            images_aux = None
+
+        conv = conv_templates[args.conv_mode].copy()
+        conv.append_message(conv.roles[0], qs)
+        conv.append_message(conv.roles[1], None)
+        prompt = conv.get_prompt()
+
+        input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).cuda()
+        
+        with torch.inference_mode():
+            output_ids = model.generate(
+                input_ids,
+                images=images,
+                images_aux=images_aux,
+                do_sample=True if args.temperature > 0 else False,
+                temperature=args.temperature,
+                max_new_tokens=1024,
+                bos_token_id=tokenizer.bos_token_id,  # Begin of sequence token
+                eos_token_id=tokenizer.eos_token_id,  # End of sequence token
+                pad_token_id=tokenizer.pad_token_id,  # Pad token
+                use_cache=True,
+            )
+
+        outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0].strip()
+
+        info['query'] = query
+        info['response'] = outputs
+        ans_file.write(json.dumps(info) + "\n")
+        ans_file.flush()
+    ans_file.close()
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model-path", type=str, default="facebook/opt-350m")
+    parser.add_argument("--model-base", type=str, default=None)
+    parser.add_argument("--image-folder", type=str, default="")
+    parser.add_argument("--question-file", type=str, default="tables/question.json")
+    parser.add_argument("--answers-file", type=str, default="answer.jsonl")
+    parser.add_argument("--conv-mode", type=str, default="llava_v0")
+    parser.add_argument("--num-chunks", type=int, default=1)
+    parser.add_argument("--chunk-idx", type=int, default=0)
+    parser.add_argument("--temperature", type=float, default=0.2)
+    parser.add_argument("--answer-prompter", action="store_true")
+    parser.add_argument('--load_8bit', type=bool, default=False)
+    parser.add_argument("--single-pred-prompt", action="store_true")
+    args = parser.parse_args()
+
+    eval_model(args)

minigemini/eval/model_qa.py

@@ -0,0 +1,64 @@
+import argparse
+from transformers import AutoTokenizer, AutoModelForCausalLM, StoppingCriteria
+import torch
+import os
+import json
+from tqdm import tqdm
+import shortuuid
+
+from minigemini.conversation import default_conversation
+from minigemini.utils import disable_torch_init
+
+
+@torch.inference_mode()
+def eval_model(model_name, questions_file, answers_file):
+    # Model
+    disable_torch_init()
+    model_name = os.path.expanduser(model_name)
+    tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=False)
+    model = AutoModelForCausalLM.from_pretrained(model_name,
+        torch_dtype=torch.float16).cuda()
+
+
+    ques_file = open(os.path.expanduser(questions_file), "r")
+    ans_file = open(os.path.expanduser(answers_file), "w")
+    for i, line in enumerate(tqdm(ques_file)):
+        idx = json.loads(line)["question_id"]
+        qs = json.loads(line)["text"]
+        cat = json.loads(line)["category"]
+        conv = default_conversation.copy()
+        conv.append_message(conv.roles[0], qs)
+        prompt = conv.get_prompt()
+        inputs = tokenizer([prompt])
+        input_ids = torch.as_tensor(inputs.input_ids).cuda()
+        output_ids = model.generate(
+            input_ids,
+            do_sample=True,
+            use_cache=True,
+            temperature=0.7,
+            max_new_tokens=1024,)
+        outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0]
+        try:
+            index = outputs.index(conv.sep, len(prompt))
+        except ValueError:
+            outputs += conv.sep
+            index = outputs.index(conv.sep, len(prompt))
+
+        outputs = outputs[len(prompt) + len(conv.roles[1]) + 2:index].strip()
+        ans_id = shortuuid.uuid()
+        ans_file.write(json.dumps({"question_id": idx,
+                                   "text": outputs,
+                                   "answer_id": ans_id,
+                                   "model_id": model_name,
+                                   "metadata": {}}) + "\n")
+        ans_file.flush()
+    ans_file.close()
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model-name", type=str, default="facebook/opt-350m")
+    parser.add_argument("--question-file", type=str, default="tables/question.jsonl")
+    parser.add_argument("--answers-file", type=str, default="answer.jsonl")
+    args = parser.parse_args()
+
+    eval_model(args.model_name, args.question_file, args.answers_file)

minigemini/eval/model_vqa.py

@@ -0,0 +1,154 @@
+import argparse
+import torch
+import os
+import json
+from tqdm import tqdm
+import shortuuid
+
+from minigemini.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+from minigemini.conversation import conv_templates, SeparatorStyle
+from minigemini.model.builder import load_pretrained_model
+from minigemini.utils import disable_torch_init
+from minigemini.mm_utils import tokenizer_image_token, process_images, get_model_name_from_path
+
+from PIL import Image
+import math
+
+
+def split_list(lst, n):
+    """Split a list into n (roughly) equal-sized chunks"""
+    chunk_size = math.ceil(len(lst) / n)  # integer division
+    return [lst[i:i+chunk_size] for i in range(0, len(lst), chunk_size)]
+
+
+def get_chunk(lst, n, k):
+    chunks = split_list(lst, n)
+    return chunks[k]
+
+
+def eval_model(args):
+    # Model
+    disable_torch_init()
+    model_path = os.path.expanduser(args.model_path)
+    model_name = get_model_name_from_path(model_path)
+    tokenizer, model, image_processor, context_len = load_pretrained_model(model_path, args.model_base, model_name)
+
+    questions = [json.loads(q) for q in open(os.path.expanduser(args.question_file), "r")]
+    questions = get_chunk(questions, args.num_chunks, args.chunk_idx)
+    answers_file = os.path.expanduser(args.answers_file)
+    os.makedirs(os.path.dirname(answers_file), exist_ok=True)
+    ans_file = open(answers_file, "w")
+    for line in tqdm(questions):
+        idx = line["question_id"]
+        image_file = line["image"]
+        qs = line["text"]
+        cur_prompt = qs
+        
+        if hasattr(model, "update_prompt"):
+            model.update_prompt([[cur_prompt]])
+
+        if model.config.mm_use_im_start_end:
+            qs = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\n' + qs
+        else:
+            qs = DEFAULT_IMAGE_TOKEN + '\n' + qs
+
+        conv = conv_templates[args.conv_mode].copy()
+        conv.append_message(conv.roles[0], qs)
+        conv.append_message(conv.roles[1], None)
+        prompt = conv.get_prompt()
+
+        input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).cuda()
+
+        image = Image.open(os.path.join(args.image_folder, image_file)).convert('RGB')
+        
+        if hasattr(model.config, 'image_size_aux'):
+            if not hasattr(image_processor, 'image_size_raw'):
+                image_processor.image_size_raw = image_processor.crop_size.copy()
+            image_processor.crop_size['height'] = model.config.image_size_aux
+            image_processor.crop_size['width'] = model.config.image_size_aux
+            image_processor.size['shortest_edge'] = model.config.image_size_aux
+        
+        image_tensor = process_images([image], image_processor, model.config)[0]
+
+        image_grid = getattr(model.config, 'image_grid', 1)
+        if hasattr(model.config, 'image_size_aux'):
+            raw_shape = [image_processor.image_size_raw['height'] * image_grid, 
+                        image_processor.image_size_raw['width'] * image_grid]
+            image_tensor_aux = image_tensor
+            image_tensor = torch.nn.functional.interpolate(image_tensor[None], 
+                                                        size=raw_shape, 
+                                                        mode='bilinear', 
+                                                        align_corners=False)[0]
+        else:
+            image_tensor_aux = []
+
+        if image_grid >= 2:            
+            raw_image = image_tensor.reshape(3, 
+                                            image_grid,
+                                            image_processor.image_size_raw['height'],
+                                            image_grid,
+                                            image_processor.image_size_raw['width'])
+            raw_image = raw_image.permute(1, 3, 0, 2, 4)
+            raw_image = raw_image.reshape(-1, 3,
+                                        image_processor.image_size_raw['height'],
+                                        image_processor.image_size_raw['width'])
+            
+            if getattr(model.config, 'image_global', False):
+                global_image = image_tensor
+                if len(global_image.shape) == 3:
+                    global_image = global_image[None]
+                global_image = torch.nn.functional.interpolate(global_image, 
+                                                        size=[image_processor.image_size_raw['height'],
+                                                            image_processor.image_size_raw['width']], 
+                                                        mode='bilinear', 
+                                                        align_corners=False)
+                # [image_crops, image_global]
+                raw_image = torch.cat([raw_image, global_image], dim=0)
+            image_tensor = raw_image.contiguous()
+
+        images = image_tensor[None].to(dtype=model.dtype, device='cuda', non_blocking=True)
+        images_aux = image_tensor_aux[None].to(dtype=model.dtype, device='cuda', non_blocking=True) if len(image_tensor_aux)>0 else None
+
+        with torch.inference_mode():
+            output_ids = model.generate(
+                input_ids,
+                images=images,
+                images_aux=images_aux,
+                do_sample=True if args.temperature > 0 else False,
+                temperature=args.temperature,
+                top_p=args.top_p,
+                num_beams=args.num_beams,
+                max_new_tokens=1024,
+                bos_token_id=tokenizer.bos_token_id,  # Begin of sequence token
+                eos_token_id=tokenizer.eos_token_id,  # End of sequence token
+                pad_token_id=tokenizer.pad_token_id,  # Pad token
+                use_cache=True)
+
+        outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0].strip()
+
+        ans_id = shortuuid.uuid()
+        ans_file.write(json.dumps({"question_id": idx,
+                                   "prompt": cur_prompt,
+                                   "text": outputs,
+                                   "answer_id": ans_id,
+                                   "model_id": model_name,
+                                   "metadata": {}}) + "\n")
+        ans_file.flush()
+    ans_file.close()
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model-path", type=str, default="facebook/opt-350m")
+    parser.add_argument("--model-base", type=str, default=None)
+    parser.add_argument("--image-folder", type=str, default="")
+    parser.add_argument("--question-file", type=str, default="tables/question.jsonl")
+    parser.add_argument("--answers-file", type=str, default="answer.jsonl")
+    parser.add_argument("--conv-mode", type=str, default="llava_v1")
+    parser.add_argument("--num-chunks", type=int, default=1)
+    parser.add_argument("--chunk-idx", type=int, default=0)
+    parser.add_argument("--temperature", type=float, default=0.2)
+    parser.add_argument("--top_p", type=float, default=None)
+    parser.add_argument("--num_beams", type=int, default=1)
+    args = parser.parse_args()
+
+    eval_model(args)

minigemini/eval/model_vqa_loader.py

@@ -0,0 +1,187 @@
+import argparse
+import torch
+import os
+import json
+from tqdm import tqdm
+import shortuuid
+
+from minigemini.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+from minigemini.conversation import conv_templates, SeparatorStyle
+from minigemini.model.builder import load_pretrained_model
+from minigemini.utils import disable_torch_init
+from minigemini.mm_utils import tokenizer_image_token, process_images, get_model_name_from_path
+from torch.utils.data import Dataset, DataLoader
+
+from PIL import Image
+import math
+
+def split_list(lst, n):
+    """Split a list into n (roughly) equal-sized chunks"""
+    chunk_size = math.ceil(len(lst) / n)  # integer division
+    return [lst[i:i+chunk_size] for i in range(0, len(lst), chunk_size)]
+
+
+def get_chunk(lst, n, k):
+    chunks = split_list(lst, n)
+    return chunks[k]
+
+
+# Custom dataset class
+class CustomDataset(Dataset):
+    def __init__(self, questions, image_folder, tokenizer, image_processor, model_config):
+        self.questions = questions
+        self.image_folder = image_folder
+        self.tokenizer = tokenizer
+        self.image_processor = image_processor
+        self.model_config = model_config
+
+    def __getitem__(self, index):
+        line = self.questions[index]
+        image_file = line["image"]
+        qs = line["text"]
+        
+        if self.model_config.mm_use_im_start_end:
+            qs = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\n' + qs
+        else:
+            qs = DEFAULT_IMAGE_TOKEN + '\n' + qs
+
+        conv = conv_templates[args.conv_mode].copy()
+        conv.append_message(conv.roles[0], qs)
+        conv.append_message(conv.roles[1], None)
+        prompt = conv.get_prompt()
+
+        image = Image.open(os.path.join(self.image_folder, image_file)).convert('RGB')
+                
+        if hasattr(self.model_config, 'image_size_aux'):
+            if not hasattr(self.image_processor, 'image_size_raw'):
+                self.image_processor.image_size_raw = self.image_processor.crop_size.copy()
+            self.image_processor.crop_size['height'] = self.model_config.image_size_aux
+            self.image_processor.crop_size['width'] = self.model_config.image_size_aux
+            self.image_processor.size['shortest_edge'] = self.model_config.image_size_aux
+        
+        image_tensor = process_images([image], self.image_processor, self.model_config)[0]
+
+        input_ids = tokenizer_image_token(prompt, self.tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt')
+        
+        image_grid = getattr(self.model_config, 'image_grid', 1)
+        if hasattr(self.model_config, 'image_size_aux'):
+            raw_shape = [self.image_processor.image_size_raw['height'] * image_grid, 
+                         self.image_processor.image_size_raw['width'] * image_grid]
+            image_tensor_aux = image_tensor
+            image_tensor = torch.nn.functional.interpolate(image_tensor[None], 
+                                                           size=raw_shape, 
+                                                           mode='bilinear', 
+                                                           align_corners=False)[0]
+        else:
+            image_tensor_aux = []
+
+        if image_grid >= 2:            
+            raw_image = image_tensor.reshape(3, 
+                                             image_grid,
+                                             self.image_processor.image_size_raw['height'],
+                                             image_grid,
+                                             self.image_processor.image_size_raw['width'])
+            raw_image = raw_image.permute(1, 3, 0, 2, 4)
+            raw_image = raw_image.reshape(-1, 3,
+                                          self.image_processor.image_size_raw['height'],
+                                          self.image_processor.image_size_raw['width'])
+            
+            if getattr(self.model_config, 'image_global', False):
+                global_image = image_tensor
+                if len(global_image.shape) == 3:
+                    global_image = global_image[None]
+                global_image = torch.nn.functional.interpolate(global_image, 
+                                                        size=[self.image_processor.image_size_raw['height'],
+                                                              self.image_processor.image_size_raw['width']], 
+                                                        mode='bilinear', 
+                                                        align_corners=False)
+                # [image_crops, image_global]
+                raw_image = torch.cat([raw_image, global_image], dim=0)
+            image_tensor = raw_image.contiguous()
+
+        return input_ids, image_tensor, image_tensor_aux
+    
+    def __len__(self):
+        return len(self.questions)
+
+
+# DataLoader
+def create_data_loader(questions, image_folder, tokenizer, image_processor, model_config, batch_size=1, num_workers=4):
+    assert batch_size == 1, "batch_size must be 1"
+    dataset = CustomDataset(questions, image_folder, tokenizer, image_processor, model_config)
+    data_loader = DataLoader(dataset, batch_size=batch_size, num_workers=num_workers, shuffle=False)
+    return data_loader
+
+
+def eval_model(args):
+    # Model
+    disable_torch_init()
+    model_path = os.path.expanduser(args.model_path)
+    model_name = get_model_name_from_path(model_path)
+    tokenizer, model, image_processor, context_len = load_pretrained_model(model_path, args.model_base, model_name, load_8bit=args.load_8bit)
+
+    questions = [json.loads(q) for q in open(os.path.expanduser(args.question_file), "r")]
+    questions = get_chunk(questions, args.num_chunks, args.chunk_idx)
+    answers_file = os.path.expanduser(args.answers_file)
+    os.makedirs(os.path.dirname(answers_file), exist_ok=True)
+    ans_file = open(answers_file, "w")
+
+    if 'plain' in args.conv_mode and 'finetune' not in model_name.lower() and 'mmtag' not in args.conv_mode:
+        args.conv_mode = args.conv_mode + '_mmtag'
+        print(f'It seems that this is a plain model, but it is not using a mmtag prompt, auto switching to {args.conv_mode}.')
+
+    data_loader = create_data_loader(questions, args.image_folder, tokenizer, image_processor, model.config)
+
+    for (input_ids, image_tensor, image_tensor_aux), line in tqdm(zip(data_loader, questions), total=len(questions)):
+        idx = line["question_id"]
+        cur_prompt = line["text"]
+        
+        input_ids = input_ids.to(device=model.device, non_blocking=True)
+        if hasattr(model, "update_prompt"):
+            model.update_prompt([[cur_prompt]])
+
+        with torch.inference_mode():
+            output_ids = model.generate(
+                input_ids,
+                images=image_tensor.to(dtype=model.dtype, device=model.device, non_blocking=True),
+                images_aux=image_tensor_aux.to(dtype=model.dtype, device=model.device, non_blocking=True) if len(image_tensor_aux)>0 else None,
+                do_sample=True if args.temperature > 0 else False,
+                temperature=args.temperature,
+                top_p=args.top_p,
+                num_beams=args.num_beams,
+                max_new_tokens=args.max_new_tokens,
+                bos_token_id=tokenizer.bos_token_id,  # Begin of sequence token
+                eos_token_id=tokenizer.eos_token_id,  # End of sequence token
+                pad_token_id=tokenizer.pad_token_id,  # Pad token
+                use_cache=True)
+
+        outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0].strip()
+
+        ans_id = shortuuid.uuid()
+        ans_file.write(json.dumps({"question_id": idx,
+                                   "prompt": cur_prompt,
+                                   "text": outputs,
+                                   "answer_id": ans_id,
+                                   "model_id": model_name,
+                                   "metadata": {}}) + "\n")
+        # ans_file.flush()
+    ans_file.close()
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model-path", type=str, default="facebook/opt-350m")
+    parser.add_argument("--model-base", type=str, default=None)
+    parser.add_argument("--image-folder", type=str, default="")
+    parser.add_argument("--question-file", type=str, default="tables/question.jsonl")
+    parser.add_argument("--answers-file", type=str, default="answer.jsonl")
+    parser.add_argument("--conv-mode", type=str, default="llava_v1")
+    parser.add_argument("--num-chunks", type=int, default=1)
+    parser.add_argument("--chunk-idx", type=int, default=0)
+    parser.add_argument("--temperature", type=float, default=0.2)
+    parser.add_argument("--top_p", type=float, default=None)
+    parser.add_argument("--num_beams", type=int, default=1)
+    parser.add_argument('--load_8bit', type=bool, default=False)
+    parser.add_argument("--max_new_tokens", type=int, default=128)
+    args = parser.parse_args()
+
+    eval_model(args)

minigemini/eval/model_vqa_mmbench.py

@@ -0,0 +1,212 @@
+import argparse
+import torch
+import os
+import json
+import pandas as pd
+from tqdm import tqdm
+import shortuuid
+
+from minigemini.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+from minigemini.conversation import conv_templates, SeparatorStyle
+from minigemini.model.builder import load_pretrained_model
+from minigemini.utils import disable_torch_init
+from minigemini.mm_utils import tokenizer_image_token, process_images, load_image_from_base64, get_model_name_from_path
+
+from PIL import Image
+import math
+
+
+all_options = ['A', 'B', 'C', 'D']
+
+
+def split_list(lst, n):
+    """Split a list into n (roughly) equal-sized chunks"""
+    chunk_size = math.ceil(len(lst) / n)  # integer division
+    return [lst[i:i+chunk_size] for i in range(0, len(lst), chunk_size)]
+
+
+def get_chunk(lst, n, k):
+    chunks = split_list(lst, n)
+    return chunks[k]
+
+
+def is_none(value):
+    if value is None:
+        return True
+    if type(value) is float and math.isnan(value):
+        return True
+    if type(value) is str and value.lower() == 'nan':
+        return True
+    if type(value) is str and value.lower() == 'none':
+        return True
+    return False
+
+def get_options(row, options):
+    parsed_options = []
+    for option in options:
+        option_value = row[option]
+        if is_none(option_value):
+            break
+        parsed_options.append(option_value)
+    return parsed_options
+
+
+def eval_model(args):
+    # Model
+    disable_torch_init()
+    model_path = os.path.expanduser(args.model_path)
+    model_name = get_model_name_from_path(model_path)
+    tokenizer, model, image_processor, context_len = load_pretrained_model(model_path, args.model_base, model_name)
+
+    questions = pd.read_table(os.path.expanduser(args.question_file))
+    questions = get_chunk(questions, args.num_chunks, args.chunk_idx)
+    answers_file = os.path.expanduser(args.answers_file)
+    os.makedirs(os.path.dirname(answers_file), exist_ok=True)
+    ans_file = open(answers_file, "w")
+
+    if 'plain' in model_name and 'finetune' not in model_name.lower() and 'mmtag' not in args.conv_mode:
+        args.conv_mode = args.conv_mode + '_mmtag'
+        print(f'It seems that this is a plain model, but it is not using a mmtag prompt, auto switching to {args.conv_mode}.')
+
+    for index, row in tqdm(questions.iterrows(), total=len(questions)):
+        options = get_options(row, all_options)
+        cur_option_char = all_options[:len(options)]
+
+        if args.all_rounds:
+            num_rounds = len(options)
+        else:
+            num_rounds = 1
+
+        for round_idx in range(num_rounds):
+            idx = row['index']
+            question = row['question']
+            hint = row['hint']
+            image = load_image_from_base64(row['image'])
+            if not is_none(hint):
+                question = hint + '\n' + question
+            for option_char, option in zip(all_options[:len(options)], options):
+                question = question + '\n' + option_char + '. ' + option
+            qs = cur_prompt = question
+            
+            if hasattr(model, "update_prompt"):
+                model.update_prompt([[cur_prompt]])
+            
+            if model.config.mm_use_im_start_end:
+                qs = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\n' + qs
+            else:
+                qs = DEFAULT_IMAGE_TOKEN + '\n' + qs
+
+            if args.single_pred_prompt:
+                if args.lang == 'cn':
+                    qs = qs + '\n' + "请直接回答选项字母。"
+                else:
+                    qs = qs + '\n' + "Answer with the option's letter from the given choices directly."
+
+            conv = conv_templates[args.conv_mode].copy()
+            conv.append_message(conv.roles[0], qs)
+            conv.append_message(conv.roles[1], None)
+            prompt = conv.get_prompt()
+
+            input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).cuda()
+
+            if hasattr(model.config, 'image_size_aux'):
+                if not hasattr(image_processor, 'image_size_raw'):
+                    image_processor.image_size_raw = image_processor.crop_size.copy()
+                image_processor.crop_size['height'] = model.config.image_size_aux
+                image_processor.crop_size['width'] = model.config.image_size_aux
+                image_processor.size['shortest_edge'] = model.config.image_size_aux
+
+            image_tensor = process_images([image], image_processor, model.config)[0]
+            image_grid = getattr(model.config, 'image_grid', 1)
+            if hasattr(model.config, 'image_size_aux'):
+                raw_shape = [image_processor.image_size_raw['height'] * image_grid, 
+                            image_processor.image_size_raw['width'] * image_grid]
+                image_tensor_aux = image_tensor
+                image_tensor = torch.nn.functional.interpolate(image_tensor[None], 
+                                                            size=raw_shape, 
+                                                            mode='bilinear', 
+                                                            align_corners=False)[0]
+            else:
+                image_tensor_aux = []
+
+            if image_grid >= 2:            
+                raw_image = image_tensor.reshape(3, 
+                                                image_grid,
+                                                image_processor.image_size_raw['height'],
+                                                image_grid,
+                                                image_processor.image_size_raw['width'])
+                raw_image = raw_image.permute(1, 3, 0, 2, 4)
+                raw_image = raw_image.reshape(-1, 3,
+                                            image_processor.image_size_raw['height'],
+                                            image_processor.image_size_raw['width'])
+                
+                if getattr(model.config, 'image_global', False):
+                    global_image = image_tensor
+                    if len(global_image.shape) == 3:
+                        global_image = global_image[None]
+                    global_image = torch.nn.functional.interpolate(global_image, 
+                                                            size=[image_processor.image_size_raw['height'],
+                                                                image_processor.image_size_raw['width']], 
+                                                            mode='bilinear', 
+                                                            align_corners=False)
+                    # [image_crops, image_global]
+                    raw_image = torch.cat([raw_image, global_image], dim=0)
+                image_tensor = raw_image.contiguous()
+            
+            images = image_tensor[None].to(dtype=model.dtype, device='cuda', non_blocking=True)
+            images_aux = image_tensor_aux[None].to(dtype=model.dtype, device='cuda', non_blocking=True) if len(image_tensor_aux)>0 else None
+
+            with torch.inference_mode():
+                output_ids = model.generate(
+                    input_ids,
+                    images=images,
+                    images_aux=images_aux,
+                    do_sample=True if args.temperature > 0 else False,
+                    temperature=args.temperature,
+                    top_p=args.top_p,
+                    num_beams=args.num_beams,
+                    # no_repeat_ngram_size=3,
+                    max_new_tokens=1024,
+                    bos_token_id=tokenizer.bos_token_id,  # Begin of sequence token
+                    eos_token_id=tokenizer.eos_token_id,  # End of sequence token
+                    pad_token_id=tokenizer.pad_token_id,  # Pad token
+                    use_cache=True)
+
+            outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0].strip()
+
+            ans_id = shortuuid.uuid()
+            ans_file.write(json.dumps({"question_id": idx,
+                                    "round_id": round_idx,
+                                    "prompt": cur_prompt,
+                                    "text": outputs,
+                                    "options": options,
+                                    "option_char": cur_option_char,
+                                    "answer_id": ans_id,
+                                    "model_id": model_name,
+                                    "metadata": {}}) + "\n")
+            ans_file.flush()
+
+            # rotate options
+            options = options[1:] + options[:1]
+            cur_option_char = cur_option_char[1:] + cur_option_char[:1]
+    ans_file.close()
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model-path", type=str, default="facebook/opt-350m")
+    parser.add_argument("--model-base", type=str, default=None)
+    parser.add_argument("--image-folder", type=str, default="")
+    parser.add_argument("--question-file", type=str, default="tables/question.jsonl")
+    parser.add_argument("--answers-file", type=str, default="answer.jsonl")
+    parser.add_argument("--conv-mode", type=str, default="llava_v1")
+    parser.add_argument("--num-chunks", type=int, default=1)
+    parser.add_argument("--chunk-idx", type=int, default=0)
+    parser.add_argument("--temperature", type=float, default=0.2)
+    parser.add_argument("--top_p", type=float, default=None)
+    parser.add_argument("--num_beams", type=int, default=1)
+    parser.add_argument("--all-rounds", action="store_true")
+    parser.add_argument("--single-pred-prompt", action="store_true")
+    parser.add_argument("--lang", type=str, default="en")
+    args = parser.parse_args()
+
+    eval_model(args)

minigemini/eval/model_vqa_qbench.py

@@ -0,0 +1,122 @@
+import argparse
+import torch
+from tqdm import tqdm
+import json
+
+from minigemini.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+from minigemini.conversation import conv_templates, SeparatorStyle
+from minigemini.model.builder import load_pretrained_model
+from minigemini.utils import disable_torch_init
+from minigemini.mm_utils import tokenizer_image_token, get_model_name_from_path, KeywordsStoppingCriteria
+
+from PIL import Image
+
+import requests
+from PIL import Image
+from io import BytesIO
+
+
+def load_image(image_file):
+    if image_file.startswith('http') or image_file.startswith('https'):
+        response = requests.get(image_file)
+        image = Image.open(BytesIO(response.content)).convert('RGB')
+    else:
+        image = Image.open(image_file).convert('RGB')
+    return image
+
+
+def eval_model(args):
+    # Model
+    disable_torch_init()
+
+    model_name = get_model_name_from_path(args.model_path)
+    tokenizer, model, image_processor, context_len = load_pretrained_model(args.model_path, args.model_base, model_name, True)
+
+
+
+    
+    with open(args.questions_file) as f:
+        llvqa_data = json.load(f)  
+        
+    for i, llddata in enumerate(tqdm(llvqa_data)):
+        filename = llddata["img_path"]
+        if args.lang == "en":
+            message = llddata["question"] + "\nChoose between one of the options as follows:\n"
+        elif args.lang == "zh":
+            message = llddata["question"] + "\在下列选项中选择一个:\n"
+        else:
+            raise NotImplementedError("Q-Bench does not support languages other than English (en) and Chinese (zh) yet. Contact us (https://github.com/VQAssessment/Q-Bench/) to convert  Q-Bench into more languages.")
+        for choice, ans in zip(["A.", "B.", "C.", "D."], llddata["candidates"]):
+            message += f"{choice} {ans}\n"
+        qs = message
+        
+        if model.config.mm_use_im_start_end:
+            qs = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\n' + qs
+        else:
+            qs = DEFAULT_IMAGE_TOKEN + '\n' + qs
+
+        if 'llama-2' in model_name.lower():
+            conv_mode = "llava_llama_2"
+        elif "v1" in model_name.lower():
+            conv_mode = "llava_v1"
+        elif "mpt" in model_name.lower():
+            conv_mode = "mpt"
+        else:
+            conv_mode = "llava_v0"
+
+        if args.conv_mode is not None and conv_mode != args.conv_mode:
+            print('[WARNING] the auto inferred conversation mode is {}, while `--conv-mode` is {}, using {}'.format(conv_mode, args.conv_mode, args.conv_mode))
+        else:
+            args.conv_mode = conv_mode
+
+        conv = conv_templates[args.conv_mode].copy()
+        conv.append_message(conv.roles[0], qs)
+        conv.append_message(conv.roles[1], None)
+        prompt = conv.get_prompt()
+
+        image = load_image(args.image_folder + filename)
+        image_tensor = image_processor.preprocess(image, return_tensors='pt')['pixel_values'].half().cuda()
+
+        input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).cuda()
+
+        stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2
+        keywords = [stop_str]
+        stopping_criteria = KeywordsStoppingCriteria(keywords, tokenizer, input_ids)
+        
+
+        with torch.inference_mode():
+            output_ids = model.generate(
+                input_ids,
+                images=image_tensor,
+                num_beams=1,
+                do_sample=False,
+                temperature=0,
+                max_new_tokens=1024,
+                use_cache=True,
+                stopping_criteria=[stopping_criteria])
+        
+        input_token_len = input_ids.shape[1]
+        n_diff_input_output = (input_ids != output_ids[:, :input_token_len]).sum().item()
+        if n_diff_input_output > 0:
+            print(f'[Warning] {n_diff_input_output} output_ids are not the same as the input_ids')
+        outputs = tokenizer.batch_decode(output_ids[:, input_token_len:], skip_special_tokens=True)[0]
+        outputs = outputs.strip()
+        if outputs.endswith(stop_str):
+            outputs = outputs[:-len(stop_str)]
+        outputs = outputs.strip()
+        llddata["response"] = outputs
+        with open(args.answers_file, "a") as wf:
+            json.dump(llddata, wf)
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model-path", type=str, default="llava-v1.5")
+    parser.add_argument("--model-base", type=str, default=None)
+    parser.add_argument("--image-folder", type=str, default="./playground/data/qbench/images_llvisionqa")
+    parser.add_argument("--questions-file", type=str, default="./playground/data/qbench/llvisionqa_dev.json")
+    parser.add_argument("--answers-file", type=str, default="answer.jsonl")
+    parser.add_argument("--conv-mode", type=str, default="llava_v1")
+    parser.add_argument("--lang", type=str, default="en")
+    args = parser.parse_args()
+
+    eval_model(args)

minigemini/eval/model_vqa_science.py

@@ -0,0 +1,162 @@
+import argparse
+import torch
+import os
+import json
+from tqdm import tqdm
+import shortuuid
+
+from minigemini.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+from minigemini.conversation import conv_templates, SeparatorStyle
+from minigemini.model.builder import load_pretrained_model
+from minigemini.utils import disable_torch_init
+from minigemini.mm_utils import tokenizer_image_token, process_images, get_model_name_from_path
+
+from PIL import Image
+import math
+
+def split_list(lst, n):
+    """Split a list into n (roughly) equal-sized chunks"""
+    chunk_size = math.ceil(len(lst) / n)  # integer division
+    return [lst[i:i+chunk_size] for i in range(0, len(lst), chunk_size)]
+
+
+def get_chunk(lst, n, k):
+    chunks = split_list(lst, n)
+    return chunks[k]
+
+
+def eval_model(args):
+    # Model
+    disable_torch_init()
+    model_path = os.path.expanduser(args.model_path)
+    model_name = get_model_name_from_path(model_path)
+    tokenizer, model, image_processor, context_len = load_pretrained_model(model_path, args.model_base, model_name)
+
+    questions = json.load(open(os.path.expanduser(args.question_file), "r"))
+    questions = get_chunk(questions, args.num_chunks, args.chunk_idx)
+    answers_file = os.path.expanduser(args.answers_file)
+    os.makedirs(os.path.dirname(answers_file), exist_ok=True)
+    ans_file = open(answers_file, "w")
+
+    for i, line in enumerate(tqdm(questions)):
+        idx = line["id"]
+        question = line['conversations'][0]
+        qs = question['value'].replace('<image>', '').strip()
+        cur_prompt = qs
+
+        if 'image' in line:
+            image_file = line["image"]
+            image = Image.open(os.path.join(args.image_folder, image_file))
+            
+            if hasattr(model.config, 'image_size_aux'):
+                if not hasattr(image_processor, 'image_size_raw'):
+                    image_processor.image_size_raw = image_processor.crop_size.copy()
+                image_processor.crop_size['height'] = model.config.image_size_aux
+                image_processor.crop_size['width'] = model.config.image_size_aux
+                image_processor.size['shortest_edge'] = model.config.image_size_aux
+            
+            image_tensor = process_images([image], image_processor, model.config)[0]
+            
+            image_grid = getattr(model.config, 'image_grid', 1)
+            if hasattr(model.config, 'image_size_aux'):
+                raw_shape = [image_processor.image_size_raw['height'] * image_grid, 
+                            image_processor.image_size_raw['width'] * image_grid]
+                image_tensor_aux = image_tensor
+                image_tensor = torch.nn.functional.interpolate(image_tensor[None], 
+                                                            size=raw_shape, 
+                                                            mode='bilinear', 
+                                                            align_corners=False)[0]
+            else:
+                image_tensor_aux = []
+
+            if image_grid >= 2:            
+                raw_image = image_tensor.reshape(3, 
+                                                image_grid,
+                                                image_processor.image_size_raw['height'],
+                                                image_grid,
+                                                image_processor.image_size_raw['width'])
+                raw_image = raw_image.permute(1, 3, 0, 2, 4)
+                raw_image = raw_image.reshape(-1, 3,
+                                            image_processor.image_size_raw['height'],
+                                            image_processor.image_size_raw['width'])
+                
+                if getattr(model.config, 'image_global', False):
+                    global_image = image_tensor
+                    if len(global_image.shape) == 3:
+                        global_image = global_image[None]
+                    global_image = torch.nn.functional.interpolate(global_image, 
+                                                            size=[image_processor.image_size_raw['height'],
+                                                                image_processor.image_size_raw['width']], 
+                                                            mode='bilinear', 
+                                                            align_corners=False)
+                    # [image_crops, image_global]
+                    raw_image = torch.cat([raw_image, global_image], dim=0)
+                image_tensor = raw_image.contiguous()
+            
+            images = image_tensor[None].to(dtype=model.dtype, device='cuda', non_blocking=True)
+            images_aux = image_tensor_aux[None].to(dtype=model.dtype, device='cuda', non_blocking=True) if len(image_tensor_aux)>0 else None
+            if getattr(model.config, 'mm_use_im_start_end', False):
+                qs = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\n' + qs
+            else:
+                qs = DEFAULT_IMAGE_TOKEN + '\n' + qs
+            cur_prompt = '<image>' + '\n' + cur_prompt
+        else:
+            images = None
+            images_aux = None
+
+        if args.single_pred_prompt:
+            qs = qs + '\n' + "Answer with the option's letter from the given choices directly."
+            cur_prompt = cur_prompt + '\n' + "Answer with the option's letter from the given choices directly."
+
+        conv = conv_templates[args.conv_mode].copy()
+        conv.append_message(conv.roles[0], qs)
+        conv.append_message(conv.roles[1], None)
+        prompt = conv.get_prompt()
+
+        input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).cuda()
+        
+        if hasattr(model, "update_prompt"):
+            model.update_prompt([[cur_prompt]])
+        
+        with torch.inference_mode():
+            output_ids = model.generate(
+                input_ids,
+                images=images,
+                images_aux=images_aux,
+                do_sample=True if args.temperature > 0 else False,
+                temperature=args.temperature,
+                max_new_tokens=1024,
+                bos_token_id=tokenizer.bos_token_id,  # Begin of sequence token
+                eos_token_id=tokenizer.eos_token_id,  # End of sequence token
+                pad_token_id=tokenizer.pad_token_id,  # Pad token
+                use_cache=True,
+            )
+
+        outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0].strip()
+
+        ans_id = shortuuid.uuid()
+        ans_file.write(json.dumps({"question_id": idx,
+                                   "prompt": cur_prompt,
+                                   "text": outputs,
+                                   "answer_id": ans_id,
+                                   "model_id": model_name,
+                                   "metadata": {}}) + "\n")
+        ans_file.flush()
+    ans_file.close()
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model-path", type=str, default="facebook/opt-350m")
+    parser.add_argument("--model-base", type=str, default=None)
+    parser.add_argument("--image-folder", type=str, default="")
+    parser.add_argument("--question-file", type=str, default="tables/question.json")
+    parser.add_argument("--answers-file", type=str, default="answer.jsonl")
+    parser.add_argument("--conv-mode", type=str, default="llava_v0")
+    parser.add_argument("--num-chunks", type=int, default=1)
+    parser.add_argument("--chunk-idx", type=int, default=0)
+    parser.add_argument("--temperature", type=float, default=0.2)
+    parser.add_argument("--answer-prompter", action="store_true")
+    parser.add_argument("--single-pred-prompt", action="store_true")
+    args = parser.parse_args()
+
+    eval_model(args)

minigemini/eval/qa_baseline_gpt35.py

@@ -0,0 +1,74 @@
+"""Generate answers with GPT-3.5"""
+# Note: you need to be using OpenAI Python v0.27.0 for the code below to work
+import argparse
+import json
+import os
+import time
+import concurrent.futures
+
+import openai
+import tqdm
+import shortuuid
+
+MODEL = 'gpt-3.5-turbo'
+MODEL_ID = 'gpt-3.5-turbo:20230327'
+
+def get_answer(question_id: int, question: str, max_tokens: int):
+    ans = {
+        'answer_id': shortuuid.uuid(),
+        'question_id': question_id,
+        'model_id': MODEL_ID,
+    }
+    for _ in range(3):
+        try:
+            response = openai.ChatCompletion.create(
+                model=MODEL,
+                messages=[{
+                    'role': 'system',
+                    'content': 'You are a helpful assistant.'
+                }, {
+                    'role': 'user',
+                    'content': question,
+                }],
+                max_tokens=max_tokens,
+            )
+            ans['text'] = response['choices'][0]['message']['content']
+            return ans
+        except Exception as e:
+            print('[ERROR]', e)
+            ans['text'] = '#ERROR#'
+            time.sleep(1)
+    return ans
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='ChatGPT answer generation.')
+    parser.add_argument('-q', '--question')
+    parser.add_argument('-o', '--output')
+    parser.add_argument('--max-tokens', type=int, default=1024, help='maximum number of tokens produced in the output')
+    args = parser.parse_args()
+
+    questions_dict = {}
+    with open(os.path.expanduser(args.question)) as f:
+        for line in f:
+            if not line:
+                continue
+            q = json.loads(line)
+            questions_dict[q['question_id']] = q['text']
+
+    answers = []
+
+    with concurrent.futures.ThreadPoolExecutor(max_workers=32) as executor:
+        futures = []
+        for qid, question in questions_dict.items():
+            future = executor.submit(get_answer, qid, question, args.max_tokens)
+            futures.append(future)
+
+        for future in tqdm.tqdm(concurrent.futures.as_completed(futures), total=len(futures)):
+            answers.append(future.result())
+
+    answers.sort(key=lambda x: x['question_id'])
+
+    with open(os.path.expanduser(args.output), 'w') as f:
+        table = [json.dumps(ans) for ans in answers]
+        f.write('\n'.join(table))

minigemini/eval/run_llava.py

@@ -0,0 +1,143 @@
+import argparse
+import torch
+
+from minigemini.constants import (
+    IMAGE_TOKEN_INDEX,
+    DEFAULT_IMAGE_TOKEN,
+    DEFAULT_IM_START_TOKEN,
+    DEFAULT_IM_END_TOKEN,
+    IMAGE_PLACEHOLDER,
+)
+from minigemini.conversation import conv_templates, SeparatorStyle
+from minigemini.model.builder import load_pretrained_model
+from minigemini.utils import disable_torch_init
+from minigemini.mm_utils import (
+    process_images,
+    tokenizer_image_token,
+    get_model_name_from_path,
+)
+
+from PIL import Image
+
+import requests
+from PIL import Image
+from io import BytesIO
+import re
+
+
+def image_parser(args):
+    out = args.image_file.split(args.sep)
+    return out
+
+
+def load_image(image_file):
+    if image_file.startswith("http") or image_file.startswith("https"):
+        response = requests.get(image_file)
+        image = Image.open(BytesIO(response.content)).convert("RGB")
+    else:
+        image = Image.open(image_file).convert("RGB")
+    return image
+
+
+def load_images(image_files):
+    out = []
+    for image_file in image_files:
+        image = load_image(image_file)
+        out.append(image)
+    return out
+
+
+def eval_model(args):
+    # Model
+    disable_torch_init()
+
+    model_name = get_model_name_from_path(args.model_path)
+    tokenizer, model, image_processor, context_len = load_pretrained_model(
+        args.model_path, args.model_base, model_name
+    )
+
+    qs = args.query
+    image_token_se = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN
+    if IMAGE_PLACEHOLDER in qs:
+        if model.config.mm_use_im_start_end:
+            qs = re.sub(IMAGE_PLACEHOLDER, image_token_se, qs)
+        else:
+            qs = re.sub(IMAGE_PLACEHOLDER, DEFAULT_IMAGE_TOKEN, qs)
+    else:
+        if model.config.mm_use_im_start_end:
+            qs = image_token_se + "\n" + qs
+        else:
+            qs = DEFAULT_IMAGE_TOKEN + "\n" + qs
+
+    if "llama-2" in model_name.lower():
+        conv_mode = "llava_llama_2"
+    elif "mistral" in model_name.lower():
+        conv_mode = "mistral_instruct"
+    elif "v1.6-34b" in model_name.lower():
+        conv_mode = "chatml_direct"
+    elif "v1" in model_name.lower():
+        conv_mode = "llava_v1"
+    elif "mpt" in model_name.lower():
+        conv_mode = "mpt"
+    else:
+        conv_mode = "llava_v0"
+
+    if args.conv_mode is not None and conv_mode != args.conv_mode:
+        print(
+            "[WARNING] the auto inferred conversation mode is {}, while `--conv-mode` is {}, using {}".format(
+                conv_mode, args.conv_mode, args.conv_mode
+            )
+        )
+    else:
+        args.conv_mode = conv_mode
+
+    conv = conv_templates[args.conv_mode].copy()
+    conv.append_message(conv.roles[0], qs)
+    conv.append_message(conv.roles[1], None)
+    prompt = conv.get_prompt()
+
+    image_files = image_parser(args)
+    images = load_images(image_files)
+    images_tensor = process_images(
+        images,
+        image_processor,
+        model.config
+    ).to(model.device, dtype=torch.float16)
+
+    input_ids = (
+        tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt")
+        .unsqueeze(0)
+        .cuda()
+    )
+
+    with torch.inference_mode():
+        output_ids = model.generate(
+            input_ids,
+            images=images_tensor,
+            do_sample=True if args.temperature > 0 else False,
+            temperature=args.temperature,
+            top_p=args.top_p,
+            num_beams=args.num_beams,
+            max_new_tokens=args.max_new_tokens,
+            use_cache=True,
+        )
+
+    outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0].strip()
+    print(outputs)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model-path", type=str, default="facebook/opt-350m")
+    parser.add_argument("--model-base", type=str, default=None)
+    parser.add_argument("--image-file", type=str, required=True)
+    parser.add_argument("--query", type=str, required=True)
+    parser.add_argument("--conv-mode", type=str, default=None)
+    parser.add_argument("--sep", type=str, default=",")
+    parser.add_argument("--temperature", type=float, default=0.2)
+    parser.add_argument("--top_p", type=float, default=None)
+    parser.add_argument("--num_beams", type=int, default=1)
+    parser.add_argument("--max_new_tokens", type=int, default=512)
+    args = parser.parse_args()
+
+    eval_model(args)

minigemini/eval/summarize_gpt_review.py

@@ -0,0 +1,60 @@
+import json
+import os
+from collections import defaultdict
+
+import numpy as np
+
+import argparse
+
+def parse_args():
+    parser = argparse.ArgumentParser(description='ChatGPT-based QA evaluation.')
+    parser.add_argument('-d', '--dir', default=None)
+    parser.add_argument('-v', '--version', default=None)
+    parser.add_argument('-s', '--select', nargs='*', default=None)
+    parser.add_argument('-f', '--files', nargs='*', default=[])
+    parser.add_argument('-i', '--ignore', nargs='*', default=[])
+    return parser.parse_args()
+
+
+if __name__ == '__main__':
+    args = parse_args()
+
+    if args.ignore is not None:
+        args.ignore = [int(x) for x in args.ignore]
+
+    if len(args.files) > 0:
+        review_files = args.files
+    else:
+        review_files = [x for x in os.listdir(args.dir) if x.endswith('.jsonl') and (x.startswith('gpt4_text') or x.startswith('reviews_') or x.startswith('review_') or 'review' in args.dir)]
+
+    for review_file in sorted(review_files):
+        config = os.path.basename(review_file).replace('gpt4_text_', '').replace('.jsonl', '')
+        if args.select is not None and any(x not in config for x in args.select):
+            continue
+        if '0613' in config:
+            version = '0613'
+        else:
+            version = '0314'
+        if args.version is not None and args.version != version:
+            continue
+        scores = defaultdict(list)
+        print(config)
+        with open(os.path.join(args.dir, review_file) if args.dir is not None else review_file) as f:
+            for review_str in f:
+                review = json.loads(review_str)
+                if review['question_id'] in args.ignore:
+                    continue
+                if 'category' in review:
+                    scores[review['category']].append(review['tuple'])
+                    scores['all'].append(review['tuple'])
+                else:
+                    if 'tuple' in review:
+                        scores['all'].append(review['tuple'])
+                    else:
+                        scores['all'].append(review['score'])
+        for k, v in sorted(scores.items()):
+            stats = np.asarray(v).mean(0).tolist()
+            stats = [round(x, 3) for x in stats]
+            # print(k, stats, round(stats[1]/stats[0]*100, 1))
+            print(k, round(stats[1]/stats[0]*100, 1), round(stats[0] * 10, 1), round(stats[1] * 10, 1))
+        print('=================================')

minigemini/mm_utils.py

@@ -0,0 +1,105 @@
+from PIL import Image
+from io import BytesIO
+import base64
+
+import torch
+from transformers import StoppingCriteria
+from minigemini.constants import IMAGE_TOKEN_INDEX
+
+
+def load_image_from_base64(image):
+    return Image.open(BytesIO(base64.b64decode(image)))
+
+
+def expand2square(pil_img, background_color):
+    width, height = pil_img.size
+    if width == height:
+        return pil_img
+    elif width > height:
+        result = Image.new(pil_img.mode, (width, width), background_color)
+        result.paste(pil_img, (0, (width - height) // 2))
+        return result
+    else:
+        result = Image.new(pil_img.mode, (height, height), background_color)
+        result.paste(pil_img, ((height - width) // 2, 0))
+        return result
+
+
+def process_images(images, image_processor, model_cfg):
+    image_aspect_ratio = getattr(model_cfg, "image_aspect_ratio", None)
+    new_images = []
+    if image_aspect_ratio == 'pad':
+        for image in images:
+            image = expand2square(image.convert('RGB'), tuple(int(x*255) for x in image_processor.image_mean))
+            image = image_processor.preprocess(image, return_tensors='pt')['pixel_values'][0]
+            new_images.append(image)
+    else:
+        return image_processor(images, return_tensors='pt')['pixel_values']
+    if all(x.shape == new_images[0].shape for x in new_images):
+        new_images = torch.stack(new_images, dim=0)
+    return new_images
+
+
+def tokenizer_image_token(prompt, tokenizer, image_token_index=IMAGE_TOKEN_INDEX, return_tensors=None):
+    prompt_chunks = [tokenizer(chunk).input_ids for chunk in prompt.split('<image>')]
+
+    def insert_separator(X, sep):
+        return [ele for sublist in zip(X, [sep]*len(X)) for ele in sublist][:-1]
+
+    input_ids = []
+    offset = 0
+    if len(prompt_chunks) > 0 and len(prompt_chunks[0]) > 0 and prompt_chunks[0][0] == tokenizer.bos_token_id:
+        offset = 1
+        input_ids.append(prompt_chunks[0][0])
+
+    for x in insert_separator(prompt_chunks, [image_token_index] * (offset + 1)):
+        input_ids.extend(x[offset:])
+
+    if return_tensors is not None:
+        if return_tensors == 'pt':
+            return torch.tensor(input_ids, dtype=torch.long)
+        raise ValueError(f'Unsupported tensor type: {return_tensors}')
+    return input_ids
+
+
+def get_model_name_from_path(model_path):
+    model_path = model_path.strip("/")
+    model_paths = model_path.split("/")
+    if model_paths[-1].startswith('checkpoint-'):
+        return model_paths[-2] + "_" + model_paths[-1]
+    else:
+        return model_paths[-1]
+
+class KeywordsStoppingCriteria(StoppingCriteria):
+    def __init__(self, keywords, tokenizer, input_ids):
+        self.keywords = keywords
+        self.keyword_ids = []
+        self.max_keyword_len = 0
+        for keyword in keywords:
+            cur_keyword_ids = tokenizer(keyword).input_ids
+            if len(cur_keyword_ids) > 1 and cur_keyword_ids[0] == tokenizer.bos_token_id:
+                cur_keyword_ids = cur_keyword_ids[1:]
+            if len(cur_keyword_ids) > self.max_keyword_len:
+                self.max_keyword_len = len(cur_keyword_ids)
+            self.keyword_ids.append(torch.tensor(cur_keyword_ids))
+        self.tokenizer = tokenizer
+        self.start_len = input_ids.shape[1]
+    
+    def call_for_batch(self, output_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
+        offset = min(output_ids.shape[1] - self.start_len, self.max_keyword_len)
+        self.keyword_ids = [keyword_id.to(output_ids.device) for keyword_id in self.keyword_ids]
+        for keyword_id in self.keyword_ids:
+            truncated_output_ids = output_ids[0, -keyword_id.shape[0]:]
+            if torch.equal(truncated_output_ids, keyword_id):
+                return True
+        outputs = self.tokenizer.batch_decode(output_ids[:, -offset:], skip_special_tokens=True)[0]
+        for keyword in self.keywords:
+            if keyword in outputs:
+                return True
+        return False
+    
+    def __call__(self, output_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
+        outputs = []
+        for i in range(output_ids.shape[0]):
+            outputs.append(self.call_for_batch(output_ids[i].unsqueeze(0), scores))
+        return all(outputs)

minigemini/model/__init__.py

@@ -0,0 +1,7 @@
+from .language_model.mini_gemini_llama import MiniGeminiLlamaForCausalLM
+try:
+    from .language_model.mini_gemini_mistral import MiniGeminiMistralForCausalLM
+    from .language_model.mini_gemini_mixtral import MiniGeminiMixtralForCausalLM
+    from .language_model.mini_gemini_gemma import MiniGeminiGemmaForCausalLM
+except:
+    ImportWarning("New model not imported. Try to update Transformers.")

minigemini/model/builder.py

@@ -0,0 +1,140 @@
+#    Copyright 2023 Haotian Liu
+#
+#    Licensed under the Apache License, Version 2.0 (the "License");
+#    you may not use this file except in compliance with the License.
+#    You may obtain a copy of the License at
+#
+#        http://www.apache.org/licenses/LICENSE-2.0
+#
+#    Unless required by applicable law or agreed to in writing, software
+#    distributed under the License is distributed on an "AS IS" BASIS,
+#    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#    See the License for the specific language governing permissions and
+#    limitations under the License.
+# ------------------------------------------------------------------------
+# Modified from LLaVA (https://github.com/haotian-liu/LLaVA)
+# Copyright 2024 Yanwei Li
+# ------------------------------------------------------------------------
+
+import os
+import warnings
+import logging
+
+from transformers import AutoTokenizer, AutoModelForCausalLM, AutoConfig, BitsAndBytesConfig
+import torch
+from minigemini.model import *
+from minigemini.constants import DEFAULT_IMAGE_PATCH_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+
+def load_pretrained_model(model_path, model_base, model_name, load_8bit=False, load_4bit=False, device_map="auto", device="cuda", use_flash_attn=False, **kwargs):
+    kwargs = {"device_map": device_map, **kwargs}
+
+    if device != "cuda":
+        kwargs['device_map'] = {"": device}
+
+    if load_8bit:
+        kwargs['load_in_8bit'] = True
+    elif load_4bit:
+        kwargs['load_in_4bit'] = True
+        kwargs['quantization_config'] = BitsAndBytesConfig(
+            load_in_4bit=True,
+            bnb_4bit_compute_dtype=torch.float16,
+            bnb_4bit_use_double_quant=True,
+            bnb_4bit_quant_type='nf4'
+        )
+    else:
+        kwargs['torch_dtype'] = torch.float16
+
+    if use_flash_attn:
+        kwargs['attn_implementation'] = 'flash_attention_2'
+    
+    logging.getLogger("transformers").setLevel(logging.ERROR)
+    
+    if 'mini-gemini' in model_name.lower():        
+        # Load MiniGemini model
+        if model_base is not None:
+            # this may be mm projector only
+            print('Loading MiniGemini from base model...')
+            
+            if "8x7b" in model_name.lower():
+                tokenizer = AutoTokenizer.from_pretrained(model_base)
+                model = MiniGeminiMixtralForCausalLM.from_pretrained(model_base, low_cpu_mem_usage=True, **kwargs)
+            elif "2b" in model_name.lower():
+                tokenizer = AutoTokenizer.from_pretrained(model_base)
+                model = MiniGeminiGemmaForCausalLM.from_pretrained(model_base, low_cpu_mem_usage=True, **kwargs)
+            else:
+                tokenizer = AutoTokenizer.from_pretrained(model_base, use_fast=False)
+                model = MiniGeminiLlamaForCausalLM.from_pretrained(model_base, low_cpu_mem_usage=True, **kwargs)
+            mm_projector_weights = torch.load(os.path.join(model_path, 'mm_projector.bin'), map_location='cpu')
+            mm_projector_weights = {k: v.to(torch.float16) for k, v in mm_projector_weights.items()}
+            model.load_state_dict(mm_projector_weights, strict=False)
+        else:
+            if "8x7b" in model_name.lower():
+                tokenizer = AutoTokenizer.from_pretrained(model_path)
+                model = MiniGeminiMixtralForCausalLM.from_pretrained(model_path, **kwargs)
+            elif "2b" in model_name.lower():
+                tokenizer = AutoTokenizer.from_pretrained(model_path)
+                model = MiniGeminiGemmaForCausalLM.from_pretrained(model_path, low_cpu_mem_usage=True, **kwargs)
+            else:
+                tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False)
+                model = MiniGeminiLlamaForCausalLM.from_pretrained(model_path, low_cpu_mem_usage=True, **kwargs)
+
+    else:
+        # Load language model
+        if model_base is not None:
+            # PEFT model
+            from peft import PeftModel
+            tokenizer = AutoTokenizer.from_pretrained(model_base, use_fast=False)
+            model = AutoModelForCausalLM.from_pretrained(model_base, low_cpu_mem_usage=True, **kwargs)
+            print(f"Loading LoRA weights from {model_path}")
+            model = PeftModel.from_pretrained(model, model_path)
+            print(f"Merging weights")
+            model = model.merge_and_unload()
+            print('Convert to FP16...')
+            model.to(torch.float16)
+        else:
+            if 'mpt' in model_name.lower():
+                tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=True)
+                model = AutoModelForCausalLM.from_pretrained(model_path, low_cpu_mem_usage=True, trust_remote_code=True, **kwargs)
+            else:
+                tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False)
+                model = AutoModelForCausalLM.from_pretrained(model_path, low_cpu_mem_usage=True, **kwargs)
+
+    image_processor = None
+
+    mm_use_im_start_end = getattr(model.config, "mm_use_im_start_end", False)
+    mm_use_im_patch_token = getattr(model.config, "mm_use_im_patch_token", True)
+    if mm_use_im_patch_token:
+        tokenizer.add_tokens([DEFAULT_IMAGE_PATCH_TOKEN], special_tokens=True)
+    if mm_use_im_start_end:
+        tokenizer.add_tokens([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN], special_tokens=True)
+    
+    model.resize_token_embeddings(len(tokenizer))
+
+    vision_tower = model.get_vision_tower()
+    if not vision_tower.is_loaded:
+        vision_tower.load_model()
+    vision_tower.to(device=device, dtype=torch.float16)
+    image_processor = vision_tower.image_processor
+    
+    if 'mini-gemini' in model_name.lower():
+        vision_tower_aux = model.get_vision_tower_aux()
+        if not vision_tower_aux.is_loaded:
+            vision_tower_aux.load_model()
+        vision_tower_aux.to(device=device, dtype=torch.float16)
+        
+        # initialize attention modules
+        model.config.model_path = model_path
+        model.get_model().initialize_uni_modules(model.config, for_eval=True)
+
+        model.get_model().vlm_uni_query_projector.to(device=device)
+        model.get_model().vlm_uni_aux_projector.to(device=device)
+        model.get_model().vlm_uni_val_projector.to(device=device)
+    
+    if hasattr(model.config, "max_sequence_length"):
+        context_len = model.config.max_sequence_length
+    else:
+        context_len = 2048
+    
+    logging.getLogger("transformers").setLevel(logging.WARNING)
+    
+    return tokenizer, model, image_processor, context_len

minigemini/model/consolidate.py

@@ -0,0 +1,29 @@
+"""
+Usage:
+python3 -m minigemini.model.consolidate --src ~/model_weights/llava-7b --dst ~/model_weights/llava-7b_consolidate
+"""
+import argparse
+
+import torch
+from transformers import AutoTokenizer, AutoModelForCausalLM
+from minigemini.model import *
+from minigemini.model.utils import auto_upgrade
+
+
+def consolidate_ckpt(src_path, dst_path):
+    print("Loading model")
+    auto_upgrade(src_path)
+    src_model = AutoModelForCausalLM.from_pretrained(src_path, torch_dtype=torch.float16, low_cpu_mem_usage=True)
+    src_tokenizer = AutoTokenizer.from_pretrained(src_path, use_fast=False)
+    src_model.save_pretrained(dst_path)
+    src_tokenizer.save_pretrained(dst_path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--src", type=str, required=True)
+    parser.add_argument("--dst", type=str, required=True)
+
+    args = parser.parse_args()
+
+    consolidate_ckpt(args.src, args.dst)

minigemini/model/language_model/mini_gemini_gemma.py

@@ -0,0 +1,164 @@
+#    Copyright 2023 Haotian Liu
+#
+#    Licensed under the Apache License, Version 2.0 (the "License");
+#    you may not use this file except in compliance with the License.
+#    You may obtain a copy of the License at
+#
+#        http://www.apache.org/licenses/LICENSE-2.0
+#
+#    Unless required by applicable law or agreed to in writing, software
+#    distributed under the License is distributed on an "AS IS" BASIS,
+#    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#    See the License for the specific language governing permissions and
+#    limitations under the License.
+# ------------------------------------------------------------------------
+# Modified from LLaVA (https://github.com/haotian-liu/LLaVA)
+# Copyright 2024 Yanwei Li
+# ------------------------------------------------------------------------
+
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+
+try:
+    from transformers import AutoConfig, AutoModelForCausalLM, \
+                            GemmaConfig, GemmaModel, GemmaForCausalLM
+except:
+    print("New model not imported. Try to update Transformers to 4.38.0 or later.")
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from transformers.generation.utils import GenerateOutput
+from transformers.generation.utils import logging
+
+from ..mini_gemini_arch import MiniGeminiMetaModel, MiniGeminiMetaForCausalLM
+
+logger = logging.get_logger(__name__)
+
+class MiniGeminiConfig(GemmaConfig):
+    model_type = "mini_gemini_gemma"
+
+
+class MiniGeminiGemmaModel(MiniGeminiMetaModel, GemmaModel):
+    config_class = MiniGeminiConfig
+    
+    def __init__(self, config: GemmaConfig):
+        super(MiniGeminiGemmaModel, self).__init__(config)
+
+class MiniGeminiGemmaForCausalLM(GemmaForCausalLM, MiniGeminiMetaForCausalLM):
+    config_class = MiniGeminiConfig
+
+    def __init__(self, config):
+        super(GemmaForCausalLM, self).__init__(config)
+        self.model = MiniGeminiGemmaModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_model(self):
+        return self.model
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None, 
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        images: Optional[torch.FloatTensor] = None,
+        images_aux: Optional[torch.FloatTensor] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+
+        if inputs_embeds is None:
+            (
+                input_ids,
+                position_ids,
+                attention_mask,
+                past_key_values,
+                inputs_embeds,
+                labels,
+            ) = self.prepare_inputs_labels_for_multimodal(
+                input_ids,
+                position_ids,
+                attention_mask,
+                past_key_values,
+                labels,
+                images,
+                images_aux
+            )
+
+        return super().forward(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            labels=labels,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict
+        )
+
+    @torch.no_grad()
+    def generate(
+        self,
+        inputs: Optional[torch.Tensor] = None,
+        images: Optional[torch.Tensor] = None,
+        images_aux: Optional[torch.FloatTensor] = None,
+        **kwargs,
+    ) -> Union[GenerateOutput, torch.LongTensor]:
+        position_ids = kwargs.pop("position_ids", None)
+        attention_mask = kwargs.pop("attention_mask", None)
+        if "inputs_embeds" in kwargs:
+            raise NotImplementedError("`inputs_embeds` is not supported")
+
+        if images is not None:
+            (
+                inputs,
+                position_ids,
+                attention_mask,
+                _,
+                inputs_embeds,
+                _
+            ) = self.prepare_inputs_labels_for_multimodal(
+                inputs,
+                position_ids,
+                attention_mask,
+                None,
+                None,
+                images,
+                images_aux
+            )
+        else:
+            inputs_embeds = self.get_model().embed_tokens(inputs)
+
+        return super().generate(
+            position_ids=position_ids,
+            attention_mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+            **kwargs
+        )
+
+    def prepare_inputs_for_generation(self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs):
+        images = kwargs.pop("images", None)
+        images_aux = kwargs.pop("images_aux", None)
+        _inputs = super().prepare_inputs_for_generation(
+            input_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, **kwargs
+        )
+        if images is not None:
+            _inputs['images'] = images
+        if images_aux is not None:
+            _inputs['images_aux'] = images_aux
+        return _inputs
+
+AutoConfig.register("mini_gemini_gemma", MiniGeminiConfig)
+AutoModelForCausalLM.register(MiniGeminiConfig, MiniGeminiGemmaForCausalLM)

minigemini/model/language_model/mini_gemini_llama.py

@@ -0,0 +1,203 @@
+#    Copyright 2023 Haotian Liu
+#
+#    Licensed under the Apache License, Version 2.0 (the "License");
+#    you may not use this file except in compliance with the License.
+#    You may obtain a copy of the License at
+#
+#        http://www.apache.org/licenses/LICENSE-2.0
+#
+#    Unless required by applicable law or agreed to in writing, software
+#    distributed under the License is distributed on an "AS IS" BASIS,
+#    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#    See the License for the specific language governing permissions and
+#    limitations under the License.
+# ------------------------------------------------------------------------
+# Modified from LLaVA (https://github.com/haotian-liu/LLaVA)
+# Copyright 2024 Yanwei Li
+# ------------------------------------------------------------------------
+
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers import AutoConfig, AutoModelForCausalLM, \
+                         LlamaConfig, LlamaModel, LlamaForCausalLM
+
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from transformers.utils import logging
+from transformers.generation.utils import GenerateOutput
+
+from minigemini.model.mini_gemini_arch import MiniGeminiMetaModel, MiniGeminiMetaForCausalLM
+from torch.nn import CrossEntropyLoss
+
+
+logger = logging.get_logger(__name__)
+
+class MiniGeminiConfig(LlamaConfig):
+    model_type = "mini_gemini"
+
+class MiniGeminiLlamaModel(MiniGeminiMetaModel, LlamaModel):
+    config_class = MiniGeminiConfig
+    
+    def __init__(self, config: LlamaConfig):
+        super(MiniGeminiLlamaModel, self).__init__(config)
+
+
+class MiniGeminiLlamaForCausalLM(LlamaForCausalLM, MiniGeminiMetaForCausalLM):
+    config_class = MiniGeminiConfig
+
+    def __init__(self, config):
+        super(LlamaForCausalLM, self).__init__(config)
+        self.model = MiniGeminiLlamaModel(config)
+        self.pretraining_tp = config.pretraining_tp
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_model(self):
+        return self.model
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        images: Optional[torch.FloatTensor] = None,
+        images_aux: Optional[torch.FloatTensor] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if inputs_embeds is None:
+            (
+                input_ids,
+                position_ids,
+                attention_mask,
+                past_key_values,
+                inputs_embeds,
+                labels
+            ) = self.prepare_inputs_labels_for_multimodal(
+                input_ids,
+                position_ids,
+                attention_mask,
+                past_key_values,
+                labels,
+                images,
+                images_aux
+            )
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs[0]
+        if self.pretraining_tp > 1:
+            lm_head_slices = self.lm_head.weight.split(self.vocab_size // self.pretraining_tp, dim=0)
+            logits = [F.linear(hidden_states, lm_head_slices[i]) for i in range(self.pretraining_tp)]
+            logits = torch.cat(logits, dim=-1)
+        else:
+            logits = self.lm_head(hidden_states)
+        logits = logits.float()
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    @torch.no_grad()
+    def generate(
+        self,
+        inputs: Optional[torch.Tensor] = None,
+        images: Optional[torch.Tensor] = None,
+        images_aux: Optional[torch.FloatTensor] = None,
+        **kwargs,
+    ) -> Union[GenerateOutput, torch.LongTensor]:
+        position_ids = kwargs.pop("position_ids", None)
+        attention_mask = kwargs.pop("attention_mask", None)
+        if "inputs_embeds" in kwargs:
+            raise NotImplementedError("`inputs_embeds` is not supported")
+
+        if images is not None:
+            (
+                inputs,
+                position_ids,
+                attention_mask,
+                _,
+                inputs_embeds,
+                _
+            ) = self.prepare_inputs_labels_for_multimodal(
+                inputs,
+                position_ids,
+                attention_mask,
+                None,
+                None,
+                images,
+                images_aux
+            )
+        else:
+            inputs_embeds = self.get_model().embed_tokens(inputs)
+
+        return super().generate(
+            position_ids=position_ids,
+            attention_mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+            **kwargs
+        )
+
+    def prepare_inputs_for_generation(self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs):
+        images = kwargs.pop("images", None)
+        images_aux = kwargs.pop("images_aux", None)
+        _inputs = super().prepare_inputs_for_generation(
+            input_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, **kwargs
+        )
+        if images is not None:
+            _inputs['images'] = images
+        if images_aux is not None:
+            _inputs['images_aux'] = images_aux
+        return _inputs
+
+AutoConfig.register("mini_gemini", MiniGeminiConfig)
+AutoModelForCausalLM.register(MiniGeminiConfig, MiniGeminiLlamaForCausalLM)

minigemini/model/language_model/mini_gemini_mistral.py

@@ -0,0 +1,162 @@
+#    Copyright 2023 Haotian Liu
+#
+#    Licensed under the Apache License, Version 2.0 (the "License");
+#    you may not use this file except in compliance with the License.
+#    You may obtain a copy of the License at
+#
+#        http://www.apache.org/licenses/LICENSE-2.0
+#
+#    Unless required by applicable law or agreed to in writing, software
+#    distributed under the License is distributed on an "AS IS" BASIS,
+#    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#    See the License for the specific language governing permissions and
+#    limitations under the License.
+# ------------------------------------------------------------------------
+# Modified from LLaVA (https://github.com/haotian-liu/LLaVA)
+# Copyright 2024 Yanwei Li
+# ------------------------------------------------------------------------
+
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+
+from transformers import AutoConfig, AutoModelForCausalLM, \
+                         MistralConfig, MistralModel, MistralForCausalLM
+
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from transformers.generation.utils import GenerateOutput
+from transformers.generation.utils import logging
+
+from ..mini_gemini_arch import MiniGeminiMetaModel, MiniGeminiMetaForCausalLM
+
+logger = logging.get_logger(__name__)
+
+class MiniGeminiConfig(MistralConfig):
+    model_type = "mini_gemini_mistral"
+
+
+class MiniGeminiMistralModel(MiniGeminiMetaModel, MistralModel):
+    config_class = MiniGeminiConfig
+    
+    def __init__(self, config: MistralConfig):
+        super(MiniGeminiMistralModel, self).__init__(config)
+        # self.max_pos_idx = 0
+
+class MiniGeminiMistralForCausalLM(MistralForCausalLM, MiniGeminiMetaForCausalLM):
+    config_class = MiniGeminiConfig
+
+    def __init__(self, config):
+        super(MistralForCausalLM, self).__init__(config)
+        self.model = MiniGeminiMistralModel(config)
+        # self.pretraining_tp = config.pretraining_tp
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_model(self):
+        return self.model
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        images: Optional[torch.FloatTensor] = None,
+        images_aux: Optional[torch.FloatTensor] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+
+        if inputs_embeds is None:
+            (
+                input_ids,
+                position_ids,
+                attention_mask,
+                past_key_values,
+                inputs_embeds,
+                labels
+            ) = self.prepare_inputs_labels_for_multimodal(
+                input_ids,
+                position_ids,
+                attention_mask,
+                past_key_values,
+                labels,
+                images,
+                images_aux
+            )
+
+        return super().forward(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            labels=labels,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict
+        )
+
+    @torch.no_grad()
+    def generate(
+        self,
+        inputs: Optional[torch.Tensor] = None,
+        images: Optional[torch.Tensor] = None,
+        images_aux: Optional[torch.FloatTensor] = None,
+        **kwargs,
+    ) -> Union[GenerateOutput, torch.LongTensor]:
+        position_ids = kwargs.pop("position_ids", None)
+        attention_mask = kwargs.pop("attention_mask", None)
+        if "inputs_embeds" in kwargs:
+            raise NotImplementedError("`inputs_embeds` is not supported")
+
+        if images is not None:
+            (
+                inputs,
+                position_ids,
+                attention_mask,
+                _,
+                inputs_embeds,
+                _
+            ) = self.prepare_inputs_labels_for_multimodal(
+                inputs,
+                position_ids,
+                attention_mask,
+                None,
+                None,
+                images,
+                images_aux
+            )
+        else:
+            inputs_embeds = self.get_model().embed_tokens(inputs)
+
+        return super().generate(
+            position_ids=position_ids,
+            attention_mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+            **kwargs
+        )
+
+    def prepare_inputs_for_generation(self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs):
+        images = kwargs.pop("images", None)
+        images_aux = kwargs.pop("images_aux", None)
+        _inputs = super().prepare_inputs_for_generation(
+            input_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, **kwargs
+        )
+        if images is not None:
+            _inputs['images'] = images
+        if images_aux is not None:
+            _inputs['images_aux'] = images_aux
+        return _inputs
+
+AutoConfig.register("mini_gemini_mistral", MiniGeminiConfig)
+AutoModelForCausalLM.register(MiniGeminiConfig, MiniGeminiMistralForCausalLM)

minigemini/model/language_model/mini_gemini_mixtral.py

@@ -0,0 +1,162 @@
+#    Copyright 2023 Haotian Liu
+#
+#    Licensed under the Apache License, Version 2.0 (the "License");
+#    you may not use this file except in compliance with the License.
+#    You may obtain a copy of the License at
+#
+#        http://www.apache.org/licenses/LICENSE-2.0
+#
+#    Unless required by applicable law or agreed to in writing, software
+#    distributed under the License is distributed on an "AS IS" BASIS,
+#    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#    See the License for the specific language governing permissions and
+#    limitations under the License.
+# ------------------------------------------------------------------------
+# Modified from LLaVA (https://github.com/haotian-liu/LLaVA)
+# Copyright 2024 Yanwei Li
+# ------------------------------------------------------------------------
+
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+
+from transformers import AutoConfig, AutoModelForCausalLM, \
+                         MixtralConfig, MixtralModel, MixtralForCausalLM
+
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from transformers.generation.utils import GenerateOutput
+from transformers.generation.utils import logging
+
+from ..mini_gemini_arch import MiniGeminiMetaModel, MiniGeminiMetaForCausalLM
+
+logger = logging.get_logger(__name__)
+
+class MiniGeminiConfig(MixtralConfig):
+    model_type = "mini_gemini_mixtral"
+
+
+class MiniGeminiMixtralModel(MiniGeminiMetaModel, MixtralModel):
+    config_class = MiniGeminiConfig
+    
+    def __init__(self, config: MixtralConfig):
+        super(MiniGeminiMixtralModel, self).__init__(config)
+        # self.max_pos_idx = 0
+
+class MiniGeminiMixtralForCausalLM(MixtralForCausalLM, MiniGeminiMetaForCausalLM):
+    config_class = MiniGeminiConfig
+
+    def __init__(self, config):
+        super(MixtralForCausalLM, self).__init__(config)
+        self.model = MiniGeminiMixtralModel(config)
+        # self.pretraining_tp = config.pretraining_tp
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_model(self):
+        return self.model
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        images: Optional[torch.FloatTensor] = None,
+        images_aux: Optional[torch.FloatTensor] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+
+        if inputs_embeds is None:
+            (
+                input_ids,
+                position_ids,
+                attention_mask,
+                past_key_values,
+                inputs_embeds,
+                labels
+            ) = self.prepare_inputs_labels_for_multimodal(
+                input_ids,
+                position_ids,
+                attention_mask,
+                past_key_values,
+                labels,
+                images,
+                images_aux
+            )        
+
+        return super().forward(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            labels=labels,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict
+        )
+
+    @torch.no_grad()
+    def generate(
+        self,
+        inputs: Optional[torch.Tensor] = None,
+        images: Optional[torch.Tensor] = None,
+        images_aux: Optional[torch.FloatTensor] = None,
+        **kwargs,
+    ) -> Union[GenerateOutput, torch.LongTensor]:
+        position_ids = kwargs.pop("position_ids", None)
+        attention_mask = kwargs.pop("attention_mask", None)
+        if "inputs_embeds" in kwargs:
+            raise NotImplementedError("`inputs_embeds` is not supported")
+
+        if images is not None:
+            (
+                inputs,
+                position_ids,
+                attention_mask,
+                _,
+                inputs_embeds,
+                _
+            ) = self.prepare_inputs_labels_for_multimodal(
+                inputs,
+                position_ids,
+                attention_mask,
+                None,
+                None,
+                images,
+                images_aux
+            )
+        else:
+            inputs_embeds = self.get_model().embed_tokens(inputs)
+
+        return super().generate(
+            position_ids=position_ids,
+            attention_mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+            **kwargs
+        )
+
+    def prepare_inputs_for_generation(self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs):
+        images = kwargs.pop("images", None)
+        images_aux = kwargs.pop("images_aux", None)
+        _inputs = super().prepare_inputs_for_generation(
+            input_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, **kwargs
+        )
+        if images is not None:
+            _inputs['images'] = images
+        if images_aux is not None:
+            _inputs['images_aux'] = images_aux
+        return _inputs
+
+AutoConfig.register("mini_gemini_mixtral", MiniGeminiConfig)
+AutoModelForCausalLM.register(MiniGeminiConfig, MiniGeminiMixtralForCausalLM)

minigemini/model/llava_arch.py

@@ -0,0 +1,299 @@
+#    Copyright 2023 Haotian Liu
+#
+#    Licensed under the Apache License, Version 2.0 (the "License");
+#    you may not use this file except in compliance with the License.
+#    You may obtain a copy of the License at
+#
+#        http://www.apache.org/licenses/LICENSE-2.0
+#
+#    Unless required by applicable law or agreed to in writing, software
+#    distributed under the License is distributed on an "AS IS" BASIS,
+#    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#    See the License for the specific language governing permissions and
+#    limitations under the License.
+
+
+from abc import ABC, abstractmethod
+
+import torch
+import torch.nn as nn
+
+from .multimodal_encoder.builder import build_vision_tower
+from .multimodal_projector.builder import build_vision_projector
+
+from minigemini.constants import IGNORE_INDEX, IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_PATCH_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+
+
+class LlavaMetaModel:
+
+    def __init__(self, config):
+        super(LlavaMetaModel, self).__init__(config)
+
+        if hasattr(config, "mm_vision_tower"):
+            self.vision_tower = build_vision_tower(config, delay_load=True)
+            self.mm_projector = build_vision_projector(config)
+
+    def get_vision_tower(self):
+        vision_tower = getattr(self, 'vision_tower', None)
+        if type(vision_tower) is list:
+            vision_tower = vision_tower[0]
+        return vision_tower
+
+    def initialize_vision_modules(self, model_args, fsdp=None):
+        vision_tower = model_args.vision_tower
+        mm_vision_select_layer = model_args.mm_vision_select_layer
+        mm_vision_select_feature = model_args.mm_vision_select_feature
+        pretrain_mm_mlp_adapter = model_args.pretrain_mm_mlp_adapter
+
+        self.config.mm_vision_tower = vision_tower
+
+        if self.get_vision_tower() is None:
+            vision_tower = build_vision_tower(model_args)
+
+            if fsdp is not None and len(fsdp) > 0:
+                self.vision_tower = [vision_tower]
+            else:
+                self.vision_tower = vision_tower
+        else:
+            if fsdp is not None and len(fsdp) > 0:
+                vision_tower = self.vision_tower[0]
+            else:
+                vision_tower = self.vision_tower
+            vision_tower.load_model()
+
+        self.config.use_mm_proj = True
+        self.config.mm_projector_type = getattr(model_args, 'mm_projector_type', 'linear')
+        self.config.mm_hidden_size = vision_tower.hidden_size
+        self.config.mm_vision_select_layer = mm_vision_select_layer
+        self.config.mm_vision_select_feature = mm_vision_select_feature
+
+        if getattr(self, 'mm_projector', None) is None:
+            self.mm_projector = build_vision_projector(self.config)
+        else:
+            # In case it is frozen by LoRA
+            for p in self.mm_projector.parameters():
+                p.requires_grad = True
+
+        if pretrain_mm_mlp_adapter is not None:
+            mm_projector_weights = torch.load(pretrain_mm_mlp_adapter, map_location='cpu')
+            def get_w(weights, keyword):
+                return {k.split(keyword + '.')[1]: v for k, v in weights.items() if keyword in k}
+
+            if 'model' in mm_projector_weights.keys():
+                mm_projector_weights = mm_projector_weights['model']
+                status = self.mm_projector.load_state_dict(mm_projector_weights, strict=False)
+                print('missing_keys:', status.missing_keys)
+            else:
+                status = self.mm_projector.load_state_dict(get_w(mm_projector_weights, 'mm_projector'), strict=False)
+                print('missing_keys:', status.missing_keys)
+
+            # class_embedding_weights = get_w(mm_projector_weights, 'model.vision_tower.vision_tower.vision_model.embeddings')
+            # if len(class_embedding_weights) > 0:
+            #     self.vision_tower.vision_tower.vision_model.embeddings.load_state_dict(class_embedding_weights, strict=False)
+
+
+class LlavaMetaForCausalLM(ABC):
+
+    @abstractmethod
+    def get_model(self):
+        pass
+
+    def get_vision_tower(self):
+        return self.get_model().get_vision_tower()
+
+    def encode_images(self, images=None, points=None):
+        if images is not None:
+            image_features = self.get_model().get_vision_tower()(images)
+            image_features = self.get_model().mm_projector(image_features)
+        if points is not None:
+            # use pre-computed features here
+            point_features = [self.get_model().mm_projector(_point).squeeze() for _point in points]
+        return image_features
+
+    def prepare_inputs_labels_for_multimodal(
+        self, input_ids, position_ids, attention_mask, past_key_values, labels, images=None, points=None
+    ):
+        vision_tower = self.get_vision_tower()
+        if vision_tower is None or images is None or input_ids.shape[1] == 1:
+            if past_key_values is not None and vision_tower is not None and images is not None and input_ids.shape[1] == 1:
+                target_shape = past_key_values[-1][-1].shape[-2] + 1
+                attention_mask = torch.cat((attention_mask, torch.ones(
+                    (attention_mask.shape[0], target_shape - attention_mask.shape[1]),
+                    dtype=attention_mask.dtype,
+                    device=attention_mask.device
+                )), dim=1)
+                position_ids = torch.sum(attention_mask, dim=1).unsqueeze(-1) - 1
+            return input_ids, position_ids, attention_mask, past_key_values, None, labels
+
+        if type(images) is list or images.ndim == 5:
+            concat_images = torch.cat([image for image in images], dim=0)
+            image_features = self.encode_images(concat_images)
+            split_sizes = [image.shape[0] for image in images]
+            image_features = torch.split(image_features, split_sizes, dim=0)
+            image_features = [x.flatten(0, 1).to(self.device) for x in image_features]
+        else:
+            image_features = self.encode_images(images).to(self.device)
+
+        # TODO: image start / end is not implemented here to support pretraining.
+        if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_im_start_end', False):
+            raise NotImplementedError
+
+        # Let's just add dummy tensors if they do not exist,
+        # it is a headache to deal with None all the time.
+        # But it is not ideal, and if you have a better idea,
+        # please open an issue / submit a PR, thanks.
+        _labels = labels
+        _position_ids = position_ids
+        _attention_mask = attention_mask
+        if attention_mask is None:
+            attention_mask = torch.ones_like(input_ids, dtype=torch.bool)
+        else:
+            attention_mask = attention_mask.bool()
+        if position_ids is None:
+            position_ids = torch.arange(0, input_ids.shape[1], dtype=torch.long, device=input_ids.device)
+        if labels is None:
+            labels = torch.full_like(input_ids, IGNORE_INDEX)
+
+        # remove the padding using attention_mask -- TODO: double check
+        input_ids = [cur_input_ids[cur_attention_mask] for cur_input_ids, cur_attention_mask in zip(input_ids, attention_mask)]
+        labels = [cur_labels[cur_attention_mask] for cur_labels, cur_attention_mask in zip(labels, attention_mask)]
+
+        new_input_embeds = []
+        new_labels = []
+        cur_image_idx = 0
+        for batch_idx, cur_input_ids in enumerate(input_ids):
+            num_images = (cur_input_ids == IMAGE_TOKEN_INDEX).sum()
+            if num_images == 0:
+                cur_image_features = image_features[cur_image_idx]
+                cur_input_embeds_1 = self.get_model().embed_tokens(cur_input_ids)
+                cur_input_embeds = torch.cat([cur_input_embeds_1, cur_image_features[0:0]], dim=0)
+                new_input_embeds.append(cur_input_embeds)
+                new_labels.append(labels[batch_idx])
+                cur_image_idx += 1
+                continue
+
+            image_token_indices = [-1] + torch.where(cur_input_ids == IMAGE_TOKEN_INDEX)[0].tolist() + [cur_input_ids.shape[0]]
+            cur_input_ids_noim = []
+            cur_labels = labels[batch_idx]
+            cur_labels_noim = []
+            for i in range(len(image_token_indices) - 1):
+                cur_input_ids_noim.append(cur_input_ids[image_token_indices[i]+1:image_token_indices[i+1]])
+                cur_labels_noim.append(cur_labels[image_token_indices[i]+1:image_token_indices[i+1]])
+            split_sizes = [x.shape[0] for x in cur_labels_noim]
+            cur_input_embeds = self.get_model().embed_tokens(torch.cat(cur_input_ids_noim))
+            cur_input_embeds_no_im = torch.split(cur_input_embeds, split_sizes, dim=0)
+            cur_new_input_embeds = []
+            cur_new_labels = []
+
+            for i in range(num_images + 1):
+                cur_new_input_embeds.append(cur_input_embeds_no_im[i])
+                cur_new_labels.append(cur_labels_noim[i])
+                if i < num_images:
+                    cur_image_features = image_features[cur_image_idx]
+                    cur_image_idx += 1
+                    cur_new_input_embeds.append(cur_image_features)
+                    cur_new_labels.append(torch.full((cur_image_features.shape[0],), IGNORE_INDEX, device=cur_labels.device, dtype=cur_labels.dtype))
+
+            cur_new_input_embeds = [x.to(self.device) for x in cur_new_input_embeds]
+            cur_new_input_embeds = torch.cat(cur_new_input_embeds)
+            cur_new_labels = torch.cat(cur_new_labels)
+
+            new_input_embeds.append(cur_new_input_embeds)
+            new_labels.append(cur_new_labels)
+
+        # Truncate sequences to max length as image embeddings can make the sequence longer
+        tokenizer_model_max_length = getattr(self.config, 'tokenizer_model_max_length', None)
+        if tokenizer_model_max_length is not None:
+            new_input_embeds = [x[:tokenizer_model_max_length] for x in new_input_embeds]
+            new_labels = [x[:tokenizer_model_max_length] for x in new_labels]
+
+        # Combine them
+        max_len = max(x.shape[0] for x in new_input_embeds)
+        batch_size = len(new_input_embeds)
+
+        new_input_embeds_padded = []
+        new_labels_padded = torch.full((batch_size, max_len), IGNORE_INDEX, dtype=new_labels[0].dtype, device=new_labels[0].device)
+        attention_mask = torch.zeros((batch_size, max_len), dtype=attention_mask.dtype, device=attention_mask.device)
+        position_ids = torch.zeros((batch_size, max_len), dtype=position_ids.dtype, device=position_ids.device)
+
+        for i, (cur_new_embed, cur_new_labels) in enumerate(zip(new_input_embeds, new_labels)):
+            cur_len = cur_new_embed.shape[0]
+            if getattr(self.config, 'tokenizer_padding_side', 'right') == "left":
+                new_input_embeds_padded.append(torch.cat((
+                    torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device),
+                    cur_new_embed
+                ), dim=0))
+                if cur_len > 0:
+                    new_labels_padded[i, -cur_len:] = cur_new_labels
+                    attention_mask[i, -cur_len:] = True
+                    position_ids[i, -cur_len:] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)
+            else:
+                new_input_embeds_padded.append(torch.cat((
+                    cur_new_embed,
+                    torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device)
+                ), dim=0))
+                if cur_len > 0:
+                    new_labels_padded[i, :cur_len] = cur_new_labels
+                    attention_mask[i, :cur_len] = True
+                    position_ids[i, :cur_len] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)
+
+        new_input_embeds = torch.stack(new_input_embeds_padded, dim=0)
+
+        if _labels is None:
+            new_labels = None
+        else:
+            new_labels = new_labels_padded
+
+        if _attention_mask is None:
+            attention_mask = None
+        else:
+            attention_mask = attention_mask.to(dtype=_attention_mask.dtype)
+
+        if _position_ids is None:
+            position_ids = None
+
+        return None, position_ids, attention_mask, past_key_values, new_input_embeds, new_labels
+
+    def initialize_vision_tokenizer(self, model_args, tokenizer):
+        if model_args.mm_use_im_patch_token:
+            tokenizer.add_tokens([DEFAULT_IMAGE_PATCH_TOKEN], special_tokens=True)
+            self.resize_token_embeddings(len(tokenizer))
+
+        if model_args.mm_use_im_start_end:
+            num_new_tokens = tokenizer.add_tokens([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN], special_tokens=True)
+            self.resize_token_embeddings(len(tokenizer))
+
+            if num_new_tokens > 0:
+                input_embeddings = self.get_input_embeddings().weight.data
+                output_embeddings = self.get_output_embeddings().weight.data
+
+                input_embeddings_avg = input_embeddings[:-num_new_tokens].mean(
+                    dim=0, keepdim=True)
+                output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(
+                    dim=0, keepdim=True)
+
+                input_embeddings[-num_new_tokens:] = input_embeddings_avg
+                output_embeddings[-num_new_tokens:] = output_embeddings_avg
+
+            if model_args.tune_mm_mlp_adapter:
+                for p in self.get_input_embeddings().parameters():
+                    p.requires_grad = True
+                for p in self.get_output_embeddings().parameters():
+                    p.requires_grad = False
+
+            if model_args.pretrain_mm_mlp_adapter:
+                mm_projector_weights = torch.load(model_args.pretrain_mm_mlp_adapter, map_location='cpu')
+                embed_tokens_weight = mm_projector_weights['model.embed_tokens.weight']
+                assert num_new_tokens == 2
+                if input_embeddings.shape == embed_tokens_weight.shape:
+                    input_embeddings[-num_new_tokens:] = embed_tokens_weight[-num_new_tokens:]
+                elif embed_tokens_weight.shape[0] == num_new_tokens:
+                    input_embeddings[-num_new_tokens:] = embed_tokens_weight
+                else:
+                    raise ValueError(f"Unexpected embed_tokens_weight shape. Pretrained: {embed_tokens_weight.shape}. Current: {input_embeddings.shape}. Numer of new tokens: {num_new_tokens}.")
+        elif model_args.mm_use_im_patch_token:
+            if model_args.tune_mm_mlp_adapter:
+                for p in self.get_input_embeddings().parameters():
+                    p.requires_grad = False
+                for p in self.get_output_embeddings().parameters():
+                    p.requires_grad = False

minigemini/model/mini_gemini_arch.py

@@ -0,0 +1,497 @@
+#    Copyright 2023 Haotian Liu
+#
+#    Licensed under the Apache License, Version 2.0 (the "License");
+#    you may not use this file except in compliance with the License.
+#    You may obtain a copy of the License at
+#
+#        http://www.apache.org/licenses/LICENSE-2.0
+#
+#    Unless required by applicable law or agreed to in writing, software
+#    distributed under the License is distributed on an "AS IS" BASIS,
+#    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#    See the License for the specific language governing permissions and
+#    limitations under the License.
+# ------------------------------------------------------------------------
+# Modified from LLaVA (https://github.com/haotian-liu/LLaVA)
+# Copyright 2024 Yanwei Li
+# ------------------------------------------------------------------------
+
+from abc import ABC, abstractmethod
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import json
+import os
+import transformers
+import safetensors
+from transformers.deepspeed import is_deepspeed_zero3_enabled
+import deepspeed
+
+from .multimodal_encoder.builder import build_vision_tower, build_vision_tower_aux
+from .multimodal_projector.builder import build_vision_projector
+
+from minigemini.constants import (IGNORE_INDEX, IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_PATCH_TOKEN, 
+                             DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN)
+
+IS_NEW_TRANSFORMERS = transformers.__version__ >= "4.34.0"
+
+class MiniGeminiMetaModel:
+
+    def __init__(self, config):
+        super(MiniGeminiMetaModel, self).__init__(config)
+
+        if hasattr(config, "mm_vision_tower"):
+            self.vision_tower = build_vision_tower(config, delay_load=True)
+            self.mm_projector = build_vision_projector(config)
+
+        if hasattr(config, "mm_vision_tower_aux"):
+            self.vision_tower_aux = build_vision_tower_aux(config, delay_load=True)
+
+    def get_vision_tower(self):
+        vision_tower = getattr(self, 'vision_tower', None)
+        if type(vision_tower) is list:
+            vision_tower = vision_tower[0]
+        return vision_tower
+
+    def get_vision_tower_aux(self):
+        vision_tower_aux = getattr(self, 'vision_tower_aux', None)
+        if type(vision_tower_aux) is list:
+            vision_tower_aux = vision_tower_aux[0]
+        return vision_tower_aux
+
+    def initialize_vision_modules(self, model_args, fsdp=None):
+        vision_tower = model_args.vision_tower
+        vision_tower_aux = model_args.vision_tower_aux
+        mm_vision_select_layer = model_args.mm_vision_select_layer
+        mm_vision_select_feature = model_args.mm_vision_select_feature
+        pretrain_mm_mlp_adapter = model_args.pretrain_mm_mlp_adapter
+
+        self.config.mm_vision_tower = vision_tower
+        self.config.mm_vision_tower_aux = vision_tower_aux
+
+        if self.get_vision_tower() is None:
+            vision_tower = build_vision_tower(model_args)
+
+            if fsdp is not None and len(fsdp) > 0:
+                self.vision_tower = [vision_tower]
+            else:
+                self.vision_tower = vision_tower
+        else:
+            if fsdp is not None and len(fsdp) > 0:
+                vision_tower = self.vision_tower[0]
+            else:
+                vision_tower = self.vision_tower
+            vision_tower.load_model()
+
+        if vision_tower_aux is not None:
+            if self.get_vision_tower_aux() is None:
+                vision_tower_aux = build_vision_tower_aux(model_args)
+
+                if fsdp is not None and len(fsdp) > 0:
+                    self.vision_tower_aux = [vision_tower_aux]
+                else:
+                    self.vision_tower_aux = vision_tower_aux
+            else:
+                if fsdp is not None and len(fsdp) > 0:
+                    vision_tower_aux = self.vision_tower_aux[0]
+                else:
+                    vision_tower_aux = self.vision_tower_aux
+                vision_tower_aux.load_model()
+            self.config.mm_hidden_size_aux = vision_tower_aux.hidden_size
+
+        self.config.use_mm_proj = True
+        self.config.mm_projector_type = getattr(model_args, 'mm_projector_type', 'linear')
+        self.config.mm_hidden_size = vision_tower.hidden_size
+        self.config.mm_vision_select_layer = mm_vision_select_layer
+        self.config.mm_vision_select_feature = mm_vision_select_feature
+
+        if getattr(self, 'mm_projector', None) is None:
+            self.mm_projector = build_vision_projector(self.config)
+        else:
+            # In case it is frozen by LoRA
+            for p in self.mm_projector.parameters():
+                p.requires_grad = True
+
+        if pretrain_mm_mlp_adapter is not None:
+            mm_projector_weights = torch.load(pretrain_mm_mlp_adapter, map_location='cpu')
+            def get_w(weights, keyword):
+                return {k.split(keyword + '.')[1]: v for k, v in weights.items() if keyword + '.' in k}
+
+            if 'model' in mm_projector_weights.keys():
+                mm_projector_weights = mm_projector_weights['model']
+                if is_deepspeed_zero3_enabled():
+                    if len(mm_projector_weights) > 0:
+                        with deepspeed.zero.GatheredParameters(mm_projector_weights, modifier_rank=0):
+                            if torch.distributed.get_rank() == 0:
+                                self.mm_projector.load_state_dict(mm_projector_weights)
+                else:
+                    status = self.mm_projector.load_state_dict(mm_projector_weights, strict=False)
+                    print('missing_keys:', status.missing_keys)
+            else:
+                if is_deepspeed_zero3_enabled():
+                    named_parameters = get_w(mm_projector_weights, 'mm_projector')
+                    if len(named_parameters) > 0:
+                        with deepspeed.zero.GatheredParameters(named_parameters, modifier_rank=0):
+                            if torch.distributed.get_rank() == 0:
+                                self.mm_projector.load_state_dict(named_parameters)
+                else:
+                    status = self.mm_projector.load_state_dict(get_w(mm_projector_weights, 'mm_projector'), strict=False)
+                    print('missing_keys:', status.missing_keys)
+            self.mm_projector = self.mm_projector.to(device=self.device)
+
+    def initialize_uni_modules(self, model_args, for_eval=False):  
+        pretrain_mm_mlp_adapter = getattr(model_args, "pretrain_mm_mlp_adapter", None)
+        self.config.image_size_aux = getattr(model_args, 'image_size_aux', 320)
+        self.config.optimize_vision_tower = getattr(model_args, 'optimize_vision_tower', False)
+        self.config.optimize_vision_tower_aux = getattr(model_args, 'optimize_vision_tower_aux', False)
+
+        self.vlm_uni_query_projector  = nn.Sequential(nn.LayerNorm(self.config.mm_hidden_size), 
+                                                      nn.Linear(self.config.mm_hidden_size, self.config.mm_hidden_size))
+        self.vlm_uni_aux_projector  = nn.Sequential(nn.LayerNorm(self.config.mm_hidden_size_aux),
+                                                    nn.Linear(self.config.mm_hidden_size_aux, self.config.mm_hidden_size))
+        self.vlm_uni_val_projector  = nn.Sequential(nn.LayerNorm(self.config.mm_hidden_size_aux),
+                                                    nn.Linear(self.config.mm_hidden_size_aux, self.config.mm_hidden_size))
+        
+        if pretrain_mm_mlp_adapter is not None:
+            projector_weights = torch.load(pretrain_mm_mlp_adapter, map_location='cpu')
+        else:
+            trainable_module = ['vlm_uni', 'vision_fpn', 'vision_stages']
+            if hasattr(model_args, 'model_name_or_path'):
+                model_save_path = model_args.model_name_or_path
+            else:
+                model_save_path = model_args.model_path
+            model_idx_path = getattr(model_args, 'model_path', model_save_path)
+            if IS_NEW_TRANSFORMERS:
+                try:
+                    weight_file = json.load(open(os.path.join(model_idx_path, 'model.safetensors.index.json'), 'r'))['weight_map']
+                except:
+                    weight_file = json.load(open(os.path.join(model_idx_path, 'pytorch_model.bin.index.json'), 'r'))['weight_map']
+            else:
+                weight_file = json.load(open(os.path.join(model_idx_path, 'pytorch_model.bin.index.json'), 'r'))['weight_map']
+            model_path = set([weight_file[_key] for _key in weight_file if any([_module in _key for _module in trainable_module])])
+            projector_weights = {}
+            for _model in model_path:
+                if not IS_NEW_TRANSFORMERS:
+                    projector_weights.update(torch.load(os.path.join(model_idx_path, _model), map_location='cpu'))
+                else:
+                    with safetensors.safe_open(os.path.join(model_idx_path, _model), framework="pt", device='cpu') as f:
+                        for _key in f.keys():
+                            projector_weights.update({_key: f.get_tensor(_key)})
+            if len(projector_weights) == 0:
+                return
+
+        def get_w(weights, keyword, main_module, sub_module):
+            if getattr(main_module, sub_module, None) is None:
+                return
+            
+            pretrain_weight = {k.split(keyword + '.')[1]: v for k, v in weights.items() if keyword + '.' in k}
+            if len(pretrain_weight) == 0:
+                return
+            if is_deepspeed_zero3_enabled():
+                named_parameters = [v for k, v in getattr(main_module, sub_module).named_parameters()]
+                if len(named_parameters) > 0:
+                    # because zero3 puts placeholders in model params, this context
+                    # manager gathers (unpartitions) the params of the current layer, then loads from
+                    # the state dict and then re-partitions them again
+                    with deepspeed.zero.GatheredParameters(named_parameters, modifier_rank=0):
+                        if torch.distributed.get_rank() == 0:
+                            getattr(main_module, sub_module).load_state_dict(pretrain_weight)
+                    with deepspeed.zero.GatheredParameters(self.mm_projector[0].weight, modifier_rank=None):
+                        weight_type = self.mm_projector[0].weight.dtype
+                        device_type = self.mm_projector[0].weight.device
+            else:
+                weight_type = self.mm_projector[0].weight.dtype
+                device_type = self.mm_projector[0].weight.device
+                getattr(main_module, sub_module).load_state_dict(pretrain_weight)
+            if weight_type == torch.uint8 or weight_type == torch.int8 or weight_type == torch.int16:
+                weight_type = torch.float16
+
+            getattr(main_module, sub_module).to(dtype=weight_type)
+            print(f"Loading {sub_module} weights...")
+        
+        # load pretrained weights
+        get_w(projector_weights, 'vision_tower.vision_tower', self.vision_tower, 'vision_tower')
+
+        # load pretrained weights
+        if self.config.optimize_vision_tower_aux:
+            # not optimize vision stem, just used to check
+            get_w(projector_weights, 'vision_tower_aux.vision_stem', self.vision_tower_aux, 'vision_stem')
+            get_w(projector_weights, 'vision_tower_aux.vision_stages', self.vision_tower_aux, 'vision_stages')
+        get_w(projector_weights, 'vlm_uni_query_projector', self, 'vlm_uni_query_projector')
+        get_w(projector_weights, 'vlm_uni_aux_projector', self, 'vlm_uni_aux_projector')
+        get_w(projector_weights, 'vlm_uni_val_projector', self, 'vlm_uni_val_projector')
+    
+class MiniGeminiMetaForCausalLM(ABC):
+
+    @abstractmethod
+    def get_model(self):
+        pass
+
+    def get_vision_tower(self):
+        return self.get_model().get_vision_tower()
+
+    def get_vision_tower_aux(self):
+        return self.get_model().get_vision_tower_aux()
+
+    def encode_images(self, images, images_aux=None, is_video=False):
+        image_grid = getattr(self.config, 'image_grid', 1)
+        image_global = getattr(self.config, 'image_global', False)
+        if image_grid > 1:
+            batch_size = images.shape[0]
+            if image_global:
+                global_images = images[:, -1:].flatten(0,1).contiguous()
+                grid_images = images[:, :-1].flatten(0,1).contiguous()
+                images = torch.cat([grid_images, global_images], dim=0)
+            else:
+                images = images.flatten(0,1).contiguous()
+        
+        image_features = self.get_model().get_vision_tower()(images)
+        
+        if image_global:
+            image_feat_global = image_features[-len(global_images):]
+            image_features = image_features[:len(grid_images)]
+        
+        if images_aux is not None:
+            image_aux_features_raw = self.get_model().get_vision_tower_aux()(images_aux).to(
+                dtype=image_features.dtype, device=image_features.device)
+            
+            if image_global:
+                image_aux_features_global = F.interpolate(image_aux_features_raw.float(), 
+                                                            scale_factor=1/image_grid, 
+                                                            mode='bilinear', 
+                                                            align_corners=False).to(dtype=image_aux_features_raw.dtype)
+                image_feat_global, image_aux_feat_global = self.unified_resampler(image_feat_global, image_aux_features_global)
+
+            if image_grid > 1:
+                image_aux_features_raw = image_aux_features_raw.reshape(*image_aux_features_raw.shape[:2],
+                                                                        image_grid,
+                                                                        image_aux_features_raw.shape[-2]//image_grid,
+                                                                        image_grid,
+                                                                        image_aux_features_raw.shape[-1]//image_grid)
+                image_aux_features_raw = image_aux_features_raw.permute(0, 2, 4, 1, 3, 5).flatten(1,2).flatten(0,1).contiguous()
+            image_features, image_aux_features = self.unified_resampler(image_features, image_aux_features_raw)
+            
+            if image_grid > 1:
+                image_features = image_features.reshape(batch_size, image_grid**2, *image_features.shape[1:])
+                image_features = image_features.flatten(1,2).contiguous()
+                image_aux_features = image_aux_features.reshape(batch_size, image_grid**2, *image_aux_features.shape[1:])
+                image_aux_features = image_aux_features.flatten(1,2).contiguous()
+            
+            # add global features, [global, local]
+            if image_global:
+                image_features = torch.cat([image_feat_global, image_features], dim=1)
+                image_aux_features = torch.cat([image_aux_feat_global, image_aux_features], dim=1)
+            
+            # token generation
+            image_features = image_features + image_aux_features
+        
+        # process image features after token generation
+        image_features = self.get_model().mm_projector(image_features)
+        
+        return image_features
+
+    def unified_resampler(self, images, images_aux):
+        # patchwise with square images
+        patch_num = int(images.shape[1]**0.5)
+        patch_size = images_aux.shape[-1]//patch_num
+        # within patch attention
+        images_aux = images_aux.permute(0,2,3,1)
+        images_aux = images_aux.reshape(len(images_aux), patch_num, patch_size, patch_num, patch_size, images_aux.shape[-1])
+        images_aux = images_aux.permute(0,1,3,2,4,5)
+        images_aux = images_aux.reshape(len(images_aux), patch_num**2, patch_size**2, images_aux.shape[-1]).contiguous()
+
+        # token 
+        print(self.get_model().vlm_uni_query_projector[0].weight.device)
+        embed_query = self.get_model().vlm_uni_query_projector(images)
+        embed_aux = self.get_model().vlm_uni_aux_projector(images_aux)
+        embed_value = self.get_model().vlm_uni_val_projector(images_aux) 
+        embed_att = embed_query[:,:,None] @ (embed_aux.transpose(-1,-2) / (embed_aux.shape[-1]**0.5))
+        embed_att = embed_att.nan_to_num()
+        embed_feat = (embed_att.softmax(-1) @ embed_value).mean(2)
+        
+        return images, embed_feat
+
+    def prepare_inputs_labels_for_multimodal(
+        self, input_ids, position_ids, attention_mask, past_key_values, labels, images=None, images_aux=None,
+    ):        
+        vision_tower = self.get_vision_tower()
+        if vision_tower is None or images is None or input_ids.shape[1] == 1:
+            if past_key_values is not None and vision_tower is not None and images is not None and input_ids.shape[1] == 1:                
+                target_shape = past_key_values[-1][-1].shape[-2] + 1
+                attention_mask = torch.cat((attention_mask, torch.ones(
+                    (attention_mask.shape[0], target_shape - attention_mask.shape[1]),
+                    dtype=attention_mask.dtype,
+                    device=attention_mask.device
+                )), dim=1)
+                position_ids = torch.sum(attention_mask, dim=1).unsqueeze(-1) - 1
+            return input_ids, position_ids, attention_mask, past_key_values, None, labels
+
+        image_features = self.encode_images(images, images_aux)
+
+        # TODO: image start / end is not implemented here to support pretraining.
+        if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_im_start_end', False):
+            raise NotImplementedError
+
+        # Let's just add dummy tensors if they do not exist,
+        # it is a headache to deal with None all the time.
+        # But it is not ideal, and if you have a better idea,
+        # please open an issue / submit a PR, thanks.
+        _labels = labels
+        _position_ids = position_ids
+        _attention_mask = attention_mask
+        if attention_mask is None:
+            attention_mask = torch.ones_like(input_ids, dtype=torch.bool)
+        else:
+            attention_mask = attention_mask.bool()
+        if position_ids is None:
+            position_ids = torch.arange(0, input_ids.shape[1], dtype=torch.long, device=input_ids.device)
+        if labels is None:
+            labels = torch.full_like(input_ids, IGNORE_INDEX)
+
+        # remove the padding using attention_mask -- TODO: double check
+        input_ids = [cur_input_ids[cur_attention_mask] for cur_input_ids, cur_attention_mask in zip(input_ids, attention_mask)]
+        labels = [cur_labels[cur_attention_mask] for cur_labels, cur_attention_mask in zip(labels, attention_mask)]
+
+        new_input_embeds = []
+        new_labels = []
+        cur_image_idx = 0
+        for batch_idx, cur_input_ids in enumerate(input_ids):
+            num_images = (cur_input_ids == IMAGE_TOKEN_INDEX).sum()
+            if num_images == 0:
+                cur_image_features = image_features[cur_image_idx]
+                cur_input_embeds_1 = self.get_model().embed_tokens(cur_input_ids)
+                cur_input_embeds = torch.cat([cur_input_embeds_1, cur_image_features[0:0]], dim=0)
+                new_input_embeds.append(cur_input_embeds)
+                new_labels.append(labels[batch_idx])
+                cur_image_idx += 1
+                continue
+
+            image_token_indices = [-1] + torch.where(cur_input_ids == IMAGE_TOKEN_INDEX)[0].tolist() + [cur_input_ids.shape[0]]
+            cur_input_ids_noim = []
+            cur_labels = labels[batch_idx]
+            cur_labels_noim = []
+            for i in range(len(image_token_indices) - 1):
+                cur_input_ids_noim.append(cur_input_ids[image_token_indices[i]+1:image_token_indices[i+1]])
+                cur_labels_noim.append(cur_labels[image_token_indices[i]+1:image_token_indices[i+1]])
+            split_sizes = [x.shape[0] for x in cur_labels_noim]
+            cur_input_embeds = self.get_model().embed_tokens(torch.cat(cur_input_ids_noim))
+            cur_input_embeds_no_im = torch.split(cur_input_embeds, split_sizes, dim=0)
+            cur_new_input_embeds = []
+            cur_new_labels = []
+            
+            max_pos_id = 0
+            for i in range(num_images + 1):
+                cur_new_input_embeds.append(cur_input_embeds_no_im[i])
+                cur_new_labels.append(cur_labels_noim[i])
+                max_pos_id += cur_input_embeds_no_im[i].shape[0]
+                if i < num_images:
+                    cur_image_features = image_features[cur_image_idx]
+                    cur_image_idx += 1
+                    cur_new_input_embeds.append(cur_image_features)
+                    cur_new_labels.append(torch.full((cur_image_features.shape[0],), IGNORE_INDEX, device=cur_labels.device, dtype=cur_labels.dtype))
+                    max_pos_id += cur_image_features.shape[0]
+            
+            cur_new_input_embeds = [x.to(device=cur_input_embeds.device) for x in cur_new_input_embeds]
+            cur_new_input_embeds = torch.cat(cur_new_input_embeds)
+            cur_new_labels = torch.cat(cur_new_labels)
+
+            new_input_embeds.append(cur_new_input_embeds)
+            new_labels.append(cur_new_labels)
+
+        # Truncate sequences to max length as image embeddings can make the sequence longer
+        tokenizer_model_max_length = getattr(self.config, 'tokenizer_model_max_length', None)
+        if tokenizer_model_max_length is not None:
+            new_input_embeds = [x[:tokenizer_model_max_length] for x in new_input_embeds]
+            new_labels = [x[:tokenizer_model_max_length] for x in new_labels]
+
+        # Combine them
+        max_len = max(x.shape[0] for x in new_input_embeds)
+        batch_size = len(new_input_embeds)
+
+        new_input_embeds_padded = []
+        new_labels_padded = torch.full((batch_size, max_len), IGNORE_INDEX, dtype=new_labels[0].dtype, device=new_labels[0].device)
+        attention_mask = torch.zeros((batch_size, max_len), dtype=attention_mask.dtype, device=attention_mask.device)
+        position_ids = torch.zeros((batch_size, max_len), dtype=position_ids.dtype, device=position_ids.device)
+
+        for i, (cur_new_embed, cur_new_labels) in enumerate(zip(new_input_embeds, new_labels)):
+            cur_len = cur_new_embed.shape[0]
+            if getattr(self.config, 'tokenizer_padding_side', 'right') == "left":
+                new_input_embeds_padded.append(torch.cat((
+                    torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device),
+                    cur_new_embed
+                ), dim=0))
+                if cur_len > 0:
+                    new_labels_padded[i, -cur_len:] = cur_new_labels
+                    attention_mask[i, -cur_len:] = True
+                    position_ids[i, -cur_len:] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)  
+            else:
+                new_input_embeds_padded.append(torch.cat((
+                    cur_new_embed,
+                    torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device)
+                ), dim=0))
+                if cur_len > 0:
+                    new_labels_padded[i, :cur_len] = cur_new_labels
+                    attention_mask[i, :cur_len] = True
+                    position_ids[i, :cur_len] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)
+                    
+        new_input_embeds = torch.stack(new_input_embeds_padded, dim=0)
+
+        if _labels is None:
+            new_labels = None
+        else:
+            new_labels = new_labels_padded
+
+        if _attention_mask is None:
+            attention_mask = None
+        else:
+            attention_mask = attention_mask.to(dtype=_attention_mask.dtype)
+
+        if _position_ids is None:
+            position_ids = None
+
+        return None, position_ids, attention_mask, past_key_values, new_input_embeds, new_labels
+
+    def initialize_vision_tokenizer(self, model_args, tokenizer):
+        if model_args.mm_use_im_patch_token:
+            tokenizer.add_tokens([DEFAULT_IMAGE_PATCH_TOKEN], special_tokens=True)
+            self.resize_token_embeddings(len(tokenizer))
+
+        if model_args.mm_use_im_start_end:
+            num_new_tokens = tokenizer.add_tokens([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN], special_tokens=True)
+            self.resize_token_embeddings(len(tokenizer))
+
+            if num_new_tokens > 0:
+                input_embeddings = self.get_input_embeddings().weight.data
+                output_embeddings = self.get_output_embeddings().weight.data
+
+                input_embeddings_avg = input_embeddings[:-num_new_tokens].mean(
+                    dim=0, keepdim=True)
+                output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(
+                    dim=0, keepdim=True)
+
+                input_embeddings[-num_new_tokens:] = input_embeddings_avg
+                output_embeddings[-num_new_tokens:] = output_embeddings_avg
+
+            if model_args.tune_mm_mlp_adapter:
+                for p in self.get_input_embeddings().parameters():
+                    p.requires_grad = True
+                for p in self.get_output_embeddings().parameters():
+                    p.requires_grad = False
+
+            if model_args.pretrain_mm_mlp_adapter:
+                mm_projector_weights = torch.load(model_args.pretrain_mm_mlp_adapter, map_location='cpu')
+                embed_tokens_weight = mm_projector_weights['model.embed_tokens.weight']
+                assert num_new_tokens == 2
+                if input_embeddings.shape == embed_tokens_weight.shape:
+                    input_embeddings[-num_new_tokens:] = embed_tokens_weight[-num_new_tokens:]
+                elif embed_tokens_weight.shape[0] == num_new_tokens:
+                    input_embeddings[-num_new_tokens:] = embed_tokens_weight
+                else:
+                    raise ValueError(f"Unexpected embed_tokens_weight shape. Pretrained: {embed_tokens_weight.shape}. Current: {input_embeddings.shape}. Numer of new tokens: {num_new_tokens}.")
+        elif model_args.mm_use_im_patch_token:
+            if model_args.tune_mm_mlp_adapter:
+                for p in self.get_input_embeddings().parameters():
+                    p.requires_grad = False
+                for p in self.get_output_embeddings().parameters():
+                    p.requires_grad = False

minigemini/model/multimodal_encoder/builder.py

@@ -0,0 +1,34 @@
+import os
+from .clip_encoder import CLIPVisionTower
+from .openclip_encoder import OpenCLIPVisionTower
+
+
+def build_vision_tower(vision_tower_cfg, **kwargs):
+    vision_tower = getattr(vision_tower_cfg, 'mm_vision_tower', getattr(vision_tower_cfg, 'vision_tower', None))
+    image_processor = getattr(vision_tower_cfg, 'image_processor', getattr(vision_tower_cfg, 'image_processor', "../processor/clip-patch14-224"))
+    
+    # if not os.path.exists(vision_tower):
+    #     raise ValueError(f'Not find vision tower: {vision_tower}')
+
+    if "openai" in vision_tower.lower() or "ShareGPT4V" in vision_tower:
+        vision_tower = 'openai/clip-vit-large-patch14-336'
+        # vision_tower = '/dataset/chengyaowang/official/MiniGemini/model_zoo/OpenAI/clip-vit-large-patch14-336'
+        return CLIPVisionTower(vision_tower, args=vision_tower_cfg, **kwargs)
+    else:
+        raise ValueError(f'Unknown vision tower: {vision_tower}')
+
+
+def build_vision_tower_aux(vision_tower_cfg, **kwargs):
+    vision_tower_aux = getattr(vision_tower_cfg, 'mm_vision_tower_aux', getattr(vision_tower_cfg, 'vision_tower_aux', None))
+    
+    # if not os.path.exists(vision_tower_aux):
+    #     raise ValueError(f'Not find vision tower: {vision_tower_aux}')
+
+    if "openclip" in vision_tower_aux.lower():
+        vision_tower_aux = './checkpoints/CLIP-convnext_large_d_320.laion2B-s29B-b131K-ft-soup'
+        # vision_tower_aux = '/dataset/chengyaowang/official/MiniGemini/model_zoo/OpenAI/openclip-convnext-large-d-320-laion2B-s29B-b131K-ft-soup'
+        return OpenCLIPVisionTower(vision_tower_aux, args=vision_tower_cfg, **kwargs)
+    elif "openai" in vision_tower_aux.lower():
+        return CLIPVisionTower(vision_tower_aux, args=vision_tower_cfg, **kwargs)
+    else:
+        raise ValueError(f'Unknown vision tower: {vision_tower_aux}')

minigemini/model/multimodal_encoder/clip_encoder.py

@@ -0,0 +1,89 @@
+import torch
+import torch.nn as nn
+
+from transformers import CLIPVisionModel, CLIPImageProcessor, CLIPVisionConfig
+from ..processor.video_processor import VideoFramesProcessor
+
+class CLIPVisionTower(nn.Module):
+    def __init__(self, vision_tower, args, delay_load=False):
+        super().__init__()
+
+        self.is_loaded = False
+
+        self.vision_tower_name = vision_tower
+        self.select_layer = args.mm_vision_select_layer
+        self.select_feature = getattr(args, 'mm_vision_select_feature', 'patch')
+        self.is_optimize = getattr(args, 'optimize_vision_tower', False)
+        
+        if not delay_load:
+            self.load_model()
+        elif getattr(args, 'unfreeze_mm_vision_tower', False):
+            self.load_model()
+        else:
+            self.cfg_only = CLIPVisionConfig.from_pretrained(self.vision_tower_name)
+
+    def load_model(self):
+        self.image_processor = VideoFramesProcessor.from_pretrained(self.vision_tower_name)
+        self.vision_tower = CLIPVisionModel.from_pretrained(self.vision_tower_name)
+        self.vision_tower.requires_grad_(False)
+
+        self.is_loaded = True
+
+    def feature_select(self, image_forward_outs):
+        image_features = image_forward_outs.hidden_states[self.select_layer]
+        if self.select_feature == 'patch':
+            image_features = image_features[:, 1:]
+        elif self.select_feature == 'cls_patch':
+            image_features = image_features
+        else:
+            raise ValueError(f'Unexpected select feature: {self.select_feature}')
+        return image_features
+
+    def image_forward(self, images):
+        if type(images) is list:
+            image_features = []
+            for image in images:
+                image_forward_out = self.vision_tower(image.to(device=self.device, dtype=self.dtype).unsqueeze(0), output_hidden_states=True)
+                image_feature = self.feature_select(image_forward_out).to(image.dtype)
+                image_features.append(image_feature)
+        else:
+            image_forward_outs = self.vision_tower(images.to(device=self.device, dtype=self.dtype), output_hidden_states=True)
+            image_features = self.feature_select(image_forward_outs).to(images.dtype)
+        
+        return image_features
+
+    def forward(self, images):
+        if not self.is_optimize:
+            with torch.no_grad():
+                image_features = self.image_forward(images)
+        else:
+            image_features = self.image_forward(images)
+
+        return image_features
+
+    @property
+    def dummy_feature(self):
+        return torch.zeros(1, self.hidden_size, device=self.device, dtype=self.dtype)
+
+    @property
+    def dtype(self):
+        return self.vision_tower.dtype
+
+    @property
+    def device(self):
+        return self.vision_tower.device
+
+    @property
+    def config(self):
+        if self.is_loaded:
+            return self.vision_tower.config
+        else:
+            return self.cfg_only
+
+    @property
+    def hidden_size(self):
+        return self.config.hidden_size
+
+    @property
+    def num_patches(self):
+        return (self.config.image_size // self.config.patch_size) ** 2

minigemini/model/multimodal_encoder/eva_encoder.py

@@ -0,0 +1,551 @@
+# Based on EVA, BEIT, timm and DeiT code bases
+# https://github.com/baaivision/EVA
+# https://github.com/rwightman/pytorch-image-models/tree/master/timm
+# https://github.com/microsoft/unilm/tree/master/beit
+# https://github.com/facebookresearch/deit/
+# https://github.com/facebookresearch/dino
+# --------------------------------------------------------'
+import math
+from functools import partial
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint as checkpoint
+from timm.models.layers import drop_path, to_2tuple, trunc_normal_
+from timm.models.registry import register_model
+from transformers import CLIPImageProcessor, CLIPVisionConfig
+from ..processor.video_processor import VideoFramesProcessor
+
+def _cfg(url='', **kwargs):
+    return {
+        'url': url,
+        'num_classes': 1000, 'input_size': (3, 224, 224), 'pool_size': None,
+        'crop_pct': .9, 'interpolation': 'bicubic',
+        'mean': (0.5, 0.5, 0.5), 'std': (0.5, 0.5, 0.5),
+        **kwargs
+    }
+
+class DropPath(nn.Module):
+    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).
+    """
+    def __init__(self, drop_prob=None):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+
+    def forward(self, x):
+        return drop_path(x, self.drop_prob, self.training)
+    
+    def extra_repr(self) -> str:
+        return 'p={}'.format(self.drop_prob)
+
+
+class Mlp(nn.Module):
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        # x = self.drop(x)
+        # commit this for the orignal BERT implement 
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+class Attention(nn.Module):
+    def __init__(
+            self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0.,
+            proj_drop=0., window_size=None, attn_head_dim=None):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        if attn_head_dim is not None:
+            head_dim = attn_head_dim
+        all_head_dim = head_dim * self.num_heads
+        self.scale = qk_scale or head_dim ** -0.5
+
+        self.qkv = nn.Linear(dim, all_head_dim * 3, bias=False)
+        if qkv_bias:
+            self.q_bias = nn.Parameter(torch.zeros(all_head_dim))
+            self.v_bias = nn.Parameter(torch.zeros(all_head_dim))
+        else:
+            self.q_bias = None
+            self.v_bias = None
+
+        if window_size:
+            self.window_size = window_size
+            self.num_relative_distance = (2 * window_size[0] - 1) * (2 * window_size[1] - 1) + 3
+            self.relative_position_bias_table = nn.Parameter(
+                torch.zeros(self.num_relative_distance, num_heads))  # 2*Wh-1 * 2*Ww-1, nH
+            # cls to token & token 2 cls & cls to cls
+
+            # get pair-wise relative position index for each token inside the window
+            coords_h = torch.arange(window_size[0])
+            coords_w = torch.arange(window_size[1])
+            coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
+            coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+            relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
+            relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
+            relative_coords[:, :, 0] += window_size[0] - 1  # shift to start from 0
+            relative_coords[:, :, 1] += window_size[1] - 1
+            relative_coords[:, :, 0] *= 2 * window_size[1] - 1
+            relative_position_index = \
+                torch.zeros(size=(window_size[0] * window_size[1] + 1, ) * 2, dtype=relative_coords.dtype)
+            relative_position_index[1:, 1:] = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+            relative_position_index[0, 0:] = self.num_relative_distance - 3
+            relative_position_index[0:, 0] = self.num_relative_distance - 2
+            relative_position_index[0, 0] = self.num_relative_distance - 1
+
+            self.register_buffer("relative_position_index", relative_position_index)
+        else:
+            self.window_size = None
+            self.relative_position_bias_table = None
+            self.relative_position_index = None
+
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(all_head_dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+    def forward(self, x, rel_pos_bias=None):
+        B, N, C = x.shape
+        qkv_bias = None
+        if self.q_bias is not None:
+            qkv_bias = torch.cat((self.q_bias, torch.zeros_like(self.v_bias, requires_grad=False), self.v_bias))
+        # qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        qkv = F.linear(input=x, weight=self.qkv.weight, bias=qkv_bias)
+        qkv = qkv.reshape(B, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]   # make torchscript happy (cannot use tensor as tuple)
+
+        q = q * self.scale
+        attn = (q @ k.transpose(-2, -1))
+
+        if self.relative_position_bias_table is not None:
+            relative_position_bias = \
+                self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
+                    self.window_size[0] * self.window_size[1] + 1,
+                    self.window_size[0] * self.window_size[1] + 1, -1)  # Wh*Ww,Wh*Ww,nH
+            relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+            attn = attn + relative_position_bias.unsqueeze(0)
+
+        if rel_pos_bias is not None:
+            attn = attn + rel_pos_bias
+        
+        attn = attn.softmax(dim=-1)
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B, N, -1)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+
+class Block(nn.Module):
+
+    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
+                 drop_path=0., init_values=None, act_layer=nn.GELU, norm_layer=nn.LayerNorm,
+                 window_size=None, attn_head_dim=None):
+        super().__init__()
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
+            attn_drop=attn_drop, proj_drop=drop, window_size=window_size, attn_head_dim=attn_head_dim)
+        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+        if init_values is not None and init_values > 0:
+            self.gamma_1 = nn.Parameter(init_values * torch.ones((dim)),requires_grad=True)
+            self.gamma_2 = nn.Parameter(init_values * torch.ones((dim)),requires_grad=True)
+        else:
+            self.gamma_1, self.gamma_2 = None, None
+
+    def forward(self, x, rel_pos_bias=None):
+        if self.gamma_1 is None:
+            x = x + self.drop_path(self.attn(self.norm1(x), rel_pos_bias=rel_pos_bias))
+            x = x + self.drop_path(self.mlp(self.norm2(x)))
+        else:
+            x = x + self.drop_path(self.gamma_1 * self.attn(self.norm1(x), rel_pos_bias=rel_pos_bias))
+            x = x + self.drop_path(self.gamma_2 * self.mlp(self.norm2(x)))
+        return x
+
+
+class PatchEmbed(nn.Module):
+    """ Image to Patch Embedding
+    """
+    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+        num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])
+        self.patch_shape = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.num_patches = num_patches
+
+        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
+
+    def forward(self, x, **kwargs):
+        B, C, H, W = x.shape
+        # FIXME look at relaxing size constraints
+        assert H == self.img_size[0] and W == self.img_size[1], \
+            f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
+        x = self.proj(x).flatten(2).transpose(1, 2)
+        return x
+
+
+class RelativePositionBias(nn.Module):
+
+    def __init__(self, window_size, num_heads):
+        super().__init__()
+        self.window_size = window_size
+        self.num_relative_distance = (2 * window_size[0] - 1) * (2 * window_size[1] - 1) + 3
+        self.relative_position_bias_table = nn.Parameter(
+            torch.zeros(self.num_relative_distance, num_heads))  # 2*Wh-1 * 2*Ww-1, nH
+        # cls to token & token 2 cls & cls to cls
+
+        # get pair-wise relative position index for each token inside the window
+        coords_h = torch.arange(window_size[0])
+        coords_w = torch.arange(window_size[1])
+        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
+        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
+        relative_coords[:, :, 0] += window_size[0] - 1  # shift to start from 0
+        relative_coords[:, :, 1] += window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * window_size[1] - 1
+        relative_position_index = \
+            torch.zeros(size=(window_size[0] * window_size[1] + 1,) * 2, dtype=relative_coords.dtype)
+        relative_position_index[1:, 1:] = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+        relative_position_index[0, 0:] = self.num_relative_distance - 3
+        relative_position_index[0:, 0] = self.num_relative_distance - 2
+        relative_position_index[0, 0] = self.num_relative_distance - 1
+
+        self.register_buffer("relative_position_index", relative_position_index)
+
+        # trunc_normal_(self.relative_position_bias_table, std=.02)
+
+    def forward(self):
+        relative_position_bias = \
+            self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
+                self.window_size[0] * self.window_size[1] + 1,
+                self.window_size[0] * self.window_size[1] + 1, -1)  # Wh*Ww,Wh*Ww,nH
+        return relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+
+
+class VisionTransformer(nn.Module):
+    """ Vision Transformer with support for patch or hybrid CNN input stage
+    """
+    def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, embed_dim=768, depth=12,
+                 num_heads=12, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop_rate=0., attn_drop_rate=0.,
+                 drop_path_rate=0., norm_layer=nn.LayerNorm, init_values=None,
+                 use_abs_pos_emb=True, use_rel_pos_bias=False, use_shared_rel_pos_bias=False,
+                 use_mean_pooling=True, init_scale=0.001, use_checkpoint=False):
+        super().__init__()
+        self.image_size = img_size
+        self.num_classes = num_classes
+        self.num_features = self.embed_dim = embed_dim  # num_features for consistency with other models
+
+        self.patch_embed = PatchEmbed(
+            img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim)
+        num_patches = self.patch_embed.num_patches
+
+        self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
+        if use_abs_pos_emb:
+            self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim))
+        else:
+            self.pos_embed = None
+        self.pos_drop = nn.Dropout(p=drop_rate)
+
+        if use_shared_rel_pos_bias:
+            self.rel_pos_bias = RelativePositionBias(window_size=self.patch_embed.patch_shape, num_heads=num_heads)
+        else:
+            self.rel_pos_bias = None
+        self.use_checkpoint = use_checkpoint
+        
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]  # stochastic depth decay rule
+        self.use_rel_pos_bias = use_rel_pos_bias
+        self.blocks = nn.ModuleList([
+            Block(
+                dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
+                drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer,
+                init_values=init_values, window_size=self.patch_embed.patch_shape if use_rel_pos_bias else None)
+            for i in range(depth)])
+#         self.norm = nn.Identity() if use_mean_pooling else norm_layer(embed_dim)
+#         self.fc_norm = norm_layer(embed_dim) if use_mean_pooling else None
+#         self.head = nn.Linear(embed_dim, num_classes) if num_classes > 0 else nn.Identity()
+
+        if self.pos_embed is not None:
+            trunc_normal_(self.pos_embed, std=.02)
+        trunc_normal_(self.cls_token, std=.02)
+        # trunc_normal_(self.mask_token, std=.02)
+#         if isinstance(self.head, nn.Linear):
+#             trunc_normal_(self.head.weight, std=.02)
+        self.apply(self._init_weights)
+        self.fix_init_weight()
+#         if isinstance(self.head, nn.Linear):
+#             self.head.weight.data.mul_(init_scale)
+#             self.head.bias.data.mul_(init_scale)
+
+    def fix_init_weight(self):
+        def rescale(param, layer_id):
+            param.div_(math.sqrt(2.0 * layer_id))
+
+        for layer_id, layer in enumerate(self.blocks):
+            rescale(layer.attn.proj.weight.data, layer_id + 1)
+            rescale(layer.mlp.fc2.weight.data, layer_id + 1)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    def get_classifier(self):
+        return self.head
+
+    def reset_classifier(self, num_classes, global_pool=''):
+        self.num_classes = num_classes
+        self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
+
+    def forward_features(self, x):
+        x = self.patch_embed(x)
+        batch_size, seq_len, _ = x.size()
+
+        cls_tokens = self.cls_token.expand(batch_size, -1, -1)  # stole cls_tokens impl from Phil Wang, thanks
+        x = torch.cat((cls_tokens, x), dim=1)
+        if self.pos_embed is not None:
+            x = x + self.pos_embed
+        x = self.pos_drop(x)
+
+        rel_pos_bias = self.rel_pos_bias() if self.rel_pos_bias is not None else None
+        for blk in self.blocks:
+            if self.use_checkpoint:
+                x = checkpoint.checkpoint(blk, x, rel_pos_bias)
+            else:
+                x = blk(x, rel_pos_bias)
+        return x
+#         x = self.norm(x)
+
+#         if self.fc_norm is not None:
+#             t = x[:, 1:, :]
+#             return self.fc_norm(t.mean(1))
+#         else:
+#             return x[:, 0]
+
+    def forward(self, x):
+        x = self.forward_features(x)
+#         x = self.head(x)
+        return x
+
+    def get_intermediate_layers(self, x):
+        x = self.patch_embed(x)
+        batch_size, seq_len, _ = x.size()
+
+        cls_tokens = self.cls_token.expand(batch_size, -1, -1)  # stole cls_tokens impl from Phil Wang, thanks
+        x = torch.cat((cls_tokens, x), dim=1)
+        if self.pos_embed is not None:
+            x = x + self.pos_embed
+        x = self.pos_drop(x)
+
+        features = []
+        rel_pos_bias = self.rel_pos_bias() if self.rel_pos_bias is not None else None
+        for blk in self.blocks:
+            x = blk(x, rel_pos_bias)
+            features.append(x)
+
+        return features
+    
+    @property
+    def dtype(self):
+        return self.cls_token.dtype
+    
+    @property
+    def device(self):
+        return self.cls_token.device
+    
+    def get_num_layer(self, var_name=""):
+        if var_name in ("cls_token", "mask_token", "pos_embed"):
+            return 0
+        elif var_name.startswith("patch_embed"):
+            return 0
+        elif var_name.startswith("rel_pos_bias"):
+            return len(self.blocks) - 1
+        elif var_name.startswith("blocks"):
+            layer_id = int(var_name.split('.')[1])
+            return layer_id + 1
+        else:
+            return len(self.blocks)
+        
+            
+def interpolate_pos_embed(model, checkpoint_model):
+    if 'pos_embed' in checkpoint_model:
+        pos_embed_checkpoint = checkpoint_model['pos_embed'].float()
+        embedding_size = pos_embed_checkpoint.shape[-1]
+        num_patches = model.patch_embed.num_patches
+        num_extra_tokens = model.pos_embed.shape[-2] - num_patches
+        # height (== width) for the checkpoint position embedding
+        orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5)
+        # height (== width) for the new position embedding
+        new_size = int(num_patches ** 0.5)
+        # class_token and dist_token are kept unchanged
+        if orig_size != new_size:
+            print("Position interpolate from %dx%d to %dx%d" % (orig_size, orig_size, new_size, new_size))
+            extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens]
+            # only the position tokens are interpolated
+            pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:]
+            pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute(0, 3, 1, 2)
+            pos_tokens = torch.nn.functional.interpolate(
+                pos_tokens, size=(new_size, new_size), mode='bicubic', align_corners=False)
+            pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)
+            new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)
+            checkpoint_model['pos_embed'] = new_pos_embed
+            
+            
+def convert_weights_to_fp16(model: nn.Module):
+    """Convert applicable model parameters to fp16"""
+
+    def _convert_weights_to_fp16(l):
+        if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Linear)):
+            l.weight.data = l.weight.data.half()
+            if l.bias is not None:
+                l.bias.data = l.bias.data.half()
+
+#         if isinstance(l, (nn.MultiheadAttention, Attention)):
+#             for attr in [*[f"{s}_proj_weight" for s in ["in", "q", "k", "v"]], "in_proj_bias", "bias_k", "bias_v"]:
+#                 tensor = getattr(l, attr)
+#                 if tensor is not None:
+#                     tensor.data = tensor.data.half()
+
+    model.apply(_convert_weights_to_fp16)
+
+class EVAVisionTower(nn.Module):
+    def __init__(self, vision_tower, image_processor, args, use_checkpoint=False, drop_path_rate=0.0, delay_load=False, dtype=torch.float32):
+        super().__init__()
+        
+        self.is_loaded = False
+        self.use_checkpoint = use_checkpoint
+        self.vision_tower_name = vision_tower
+        self.image_processor_name = image_processor
+        self.drop_path_rate = drop_path_rate
+        self.patch_size = 14
+        self.out_channel = 1408
+        if not delay_load:
+            self.load_model()
+        
+        self.vision_config = CLIPVisionConfig.from_pretrained(image_processor)
+        
+    def load_model(self):
+        # self.image_processor = CLIPImageProcessor.from_pretrained(self.image_processor_name)
+        self.image_processor = VideoFramesProcessor.from_pretrained(self.image_processor_name)
+        self.vision_tower = VisionTransformer(
+            img_size=self.image_processor.size['shortest_edge'],
+            patch_size=self.patch_size,
+            use_mean_pooling=False,
+            embed_dim=self.out_channel,
+            depth=39,
+            num_heads=self.out_channel//88,
+            mlp_ratio=4.3637,
+            qkv_bias=True,
+            drop_path_rate=self.drop_path_rate,
+            norm_layer=partial(nn.LayerNorm, eps=1e-6),
+            use_checkpoint=self.use_checkpoint,
+        )  
+        
+        state_dict = torch.load(self.vision_tower_name, map_location="cpu")    
+        interpolate_pos_embed(self.vision_tower, state_dict)
+        incompatible_keys = self.vision_tower.load_state_dict(state_dict, strict=False)
+        print(incompatible_keys)
+        self.vision_tower.requires_grad_(False)
+
+        self.is_loaded = True
+
+    @torch.no_grad()
+    def forward(self, images):
+        if type(images) is list:
+            image_features = []
+            for image in images:
+                image_forward_out = self.vision_tower(image.to(device=self.device, dtype=self.dtype).unsqueeze(0))
+                image_feature = image_forward_out.to(image.dtype)
+                image_features.append(image_feature)
+        else:
+            image_forward_outs = self.vision_tower(images.to(device=self.device, dtype=self.dtype))
+            image_features = image_forward_outs.to(images.dtype)
+
+        return image_features
+
+    def feature_select(self, image_features):
+        # image_features = image_features.hidden_states[self.select_layer]
+        if self.select_feature == 'patch':
+            image_features = image_features[:, 1:]
+        elif self.select_feature == 'cls_patch':
+            image_features = image_features
+        else:
+            raise ValueError(f'Unexpected select feature: {self.select_feature}')
+        return image_features
+
+    @property
+    def dummy_feature(self):
+        return torch.zeros(1, self.hidden_size, device=self.device, dtype=self.dtype)
+
+    @property
+    def dtype(self):
+        return self.vision_tower.dtype
+
+    @property
+    def device(self):
+        return self.vision_tower.device
+
+    @property
+    def config(self):
+        return self.vision_config
+
+    @property
+    def hidden_size(self):
+        return self.out_channel
+
+    @property
+    def num_patches(self):
+        return (self.image_processor.size['shortest_edge'] // self.patch_size) ** 2
+    
+
+    
+def create_eva_vit_g(img_size=224,drop_path_rate=0.4,use_checkpoint=False,model_path=None,precision="fp16"):
+    model = VisionTransformer(
+        img_size=img_size,
+        patch_size=14,
+        use_mean_pooling=False,
+        embed_dim=1408,
+        depth=39,
+        num_heads=1408//88,
+        mlp_ratio=4.3637,
+        qkv_bias=True,
+        drop_path_rate=drop_path_rate,
+        norm_layer=partial(nn.LayerNorm, eps=1e-6),
+        use_checkpoint=use_checkpoint,
+    )  
+    # url = "https://storage.googleapis.com/sfr-vision-language-research/LAVIS/models/BLIP2/eva_vit_g.pth"
+    # cached_file = download_cached_file(
+    #     url, check_hash=False, progress=True
+    # )
+    state_dict = torch.load(model_path, map_location="cpu")    
+    interpolate_pos_embed(model,state_dict)
+    
+    incompatible_keys = model.load_state_dict(state_dict, strict=False)
+    print(incompatible_keys)
+    
+    if precision == "fp16":
+        convert_weights_to_fp16(model)
+    return model

minigemini/model/multimodal_encoder/openclip_encoder.py

@@ -0,0 +1,225 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import os
+import json
+import logging
+import deepspeed
+from pathlib import Path
+from open_clip.factory import load_state_dict, get_model_config
+from open_clip.model import CLIPVisionCfg, CLIPTextCfg, _build_vision_tower, convert_to_custom_text_state_dict, resize_pos_embed
+from typing import Dict, Optional
+from transformers.deepspeed import deepspeed_config, is_deepspeed_zero3_enabled
+
+open_clip_config = {
+  "model_cfg": {
+    "embed_dim": 768,
+    "vision_cfg": {
+      "timm_model_name": "convnext_large",
+      "timm_model_pretrained": False,
+      "timm_pool": "",
+      "timm_proj": "mlp",
+      "timm_drop": 0.0,
+      "timm_drop_path": 0.1,
+      "image_size": 320
+    },
+    "text_cfg": {
+      "context_length": 77,
+      "vocab_size": 49408,
+      "width": 768,
+      "heads": 12,
+      "layers": 16
+    }
+  },
+  "preprocess_cfg": {
+    "mean": [
+      0.48145466,
+      0.4578275,
+      0.40821073
+    ],
+    "std": [
+      0.26862954,
+      0.26130258,
+      0.27577711
+    ]
+  }
+}
+
+# xxx
+class OpenCLIPVisionTower(nn.Module):
+    def __init__(self, vision_tower, args, delay_load=False):
+        super().__init__()
+
+        self.is_loaded = False
+        self.vision_tower_name = vision_tower
+        self.vision_config = open_clip_config
+        # json.load(open(os.path.join(vision_tower,'open_clip_config.json'), 'r'))
+        self.is_optimize = getattr(args, 'optimize_vision_tower_aux', False)
+
+        if not delay_load:
+            self.load_model()
+
+    def load_model(self):
+        # print(self.vision_tower_name)
+
+        ckpt_path = os.path.join(self.vision_tower_name, 'open_clip_pytorch_model.bin')
+        if 'convnext' in self.vision_tower_name:
+            if 'large' in self.vision_tower_name and 'd_320' in self.vision_tower_name:
+                self.model_type = 'convnext_large_d_320'
+                self.model_channel = [192, 384, 768, 1536] # stage 0-3
+            elif 'base' in self.vision_tower_name and 'w_320' in self.vision_tower_name:
+                self.model_type = 'convnext_base_w_320'
+                self.model_channel = [128, 256, 512, 1024]
+            elif 'xxlarge' in self.vision_tower_name:
+                self.model_type = 'convnext_xxlarge'
+                self.model_channel = [384, 768, 1536, 3072]
+
+        clip_model = CLIP(**get_model_config(self.model_type))
+        clip_model.visual.trunk.norm_pre = None
+        clip_model.visual.trunk.head = None
+        clip_model.visual.head = None
+        print(f'Loading pretrained weights ({self.model_type}).')
+        load_checkpoint(clip_model, ckpt_path, strict=False)
+
+        self.is_loaded = True
+        # decompose stem and stages blocks in vision tower
+        self.vision_stem = clip_model.visual.trunk.stem
+        self.vision_stages = clip_model.visual.trunk.stages
+        self.vision_stem.requires_grad_(False)
+        self.vision_stages.requires_grad_(False)
+    
+    def forward(self, images):
+        if type(images) is list:
+            image_features = []
+            for image in images:
+                image_feature = self.backbone(image.to(device=self.device, dtype=self.dtype).unsqueeze(0))
+                image_features.append(image_feature)
+        else:
+            image_features = self.backbone(images.to(device=self.device, dtype=self.dtype))
+
+        return image_features
+
+    def backbone(self, images):
+        if not self.is_optimize:
+            with torch.no_grad():
+                results = self.basic_forward(images)
+        else:
+            results = self.basic_forward(images)
+
+        target_size = (results['stage_0'].shape[-2], results['stage_0'].shape[-1])
+        result_cat = []
+        for _stage in results:
+            if _stage == 'stage_0':
+                result_cat.append(results[_stage].contiguous())
+            else:
+                result_cat.append(F.interpolate(results[_stage].float().contiguous() , 
+                                                size=target_size, 
+                                                mode='bilinear', 
+                                                align_corners=False).to(dtype=results[_stage].dtype))
+        result_cat = torch.cat(result_cat, dim=1)
+
+        return result_cat.contiguous()
+
+    def basic_forward(self, images):
+        results = {}    
+        x = self.vision_stem(images)
+        for _idx in range(len(self.vision_stages)):
+            x = self.vision_stages[_idx](x)
+            results[f'stage_{_idx}'] = x
+        return results
+
+    @property
+    def dummy_feature(self):
+        return torch.zeros(1, self.hidden_size, device=self.device, dtype=self.dtype)
+
+    @property
+    def dtype(self):
+        return self.vision_stem[0].weight.dtype
+
+    @property
+    def device(self):
+        return self.vision_stem[0].weight.device
+
+    @property
+    def config(self):
+        return self.vision_config
+
+    @property
+    def hidden_size(self):
+        return sum(self.model_channel)
+
+# modified function from open_clip to support zero3 stage
+def load_checkpoint(model, checkpoint_path, strict=True):
+    if Path(checkpoint_path).suffix in ('.npz', '.npy'):
+        from open_clip.big_vision import load_big_vision_weights
+        load_big_vision_weights(model, checkpoint_path)
+        return {}
+
+    state_dict = load_state_dict(checkpoint_path)
+    # detect old format and make compatible with new format
+    if 'positional_embedding' in state_dict and not hasattr(model, 'positional_embedding'):
+        state_dict = convert_to_custom_text_state_dict(state_dict)
+    # If loading a non-SigLIP model for SigLIP training. See https://github.com/mlfoundations/open_clip/issues/712
+    # if 'logit_bias' not in state_dict and model.logit_bias is not None:
+    #     state_dict["logit_bias"] = torch.zeros_like(state_dict["logit_scale"])
+    # Certain text transformers no longer expect position_ids after transformers==4.31
+    position_id_key = 'text.transformer.embeddings.position_ids'
+    if position_id_key in state_dict and not hasattr(model, position_id_key):
+        del state_dict[position_id_key]
+    resize_pos_embed(state_dict, model)
+    # resize_text_pos_embed(state_dict, model)
+    #incompatible_keys = model.load_state_dict(state_dict, strict=strict)
+    if is_deepspeed_zero3_enabled():
+
+        error_msgs = []
+
+        def load(module: nn.Module, state_dict, prefix=""):
+            metadata = None
+
+            local_metadata = {} if metadata is None else metadata.get(prefix[:-1], {})
+            args = (state_dict, prefix, local_metadata, True, [], [], error_msgs)
+            # Parameters of module and children will start with prefix. We can exit early if there are none in this
+            # state_dict
+            if len([key for key in state_dict if key.startswith(prefix)]) > 0:
+                if is_deepspeed_zero3_enabled():
+                    # In sharded models, each shard has only part of the full state_dict, so only gather
+                    # parameters that are in the current state_dict.
+                    named_parameters = dict(module.named_parameters(prefix=prefix[:-1], recurse=False))
+                    params_to_gather = [named_parameters[k] for k in state_dict.keys() if k in named_parameters]
+                    if len(params_to_gather) > 0:
+                        # because zero3 puts placeholders in model params, this context
+                        # manager gathers (unpartitions) the params of the current layer, then loads from
+                        # the state dict and then re-partitions them again
+                        with deepspeed.zero.GatheredParameters(params_to_gather, modifier_rank=0):
+                            if torch.distributed.get_rank() == 0:
+                                module._load_from_state_dict(*args)
+                else:
+                    module._load_from_state_dict(*args)
+
+            for name, child in module._modules.items():
+                if child is not None:
+                    load(child, state_dict, prefix + name + ".")
+
+        load(model, state_dict)
+        incompatible_keys = []
+    else:
+        incompatible_keys = model.load_state_dict(state_dict, strict=strict)
+        logging.info(f"incompatible_keys.missing_keys: {incompatible_keys.missing_keys}")
+    return incompatible_keys
+
+class CLIP(nn.Module):
+    output_dict: torch.jit.Final[bool]
+
+    def __init__(
+            self,
+            embed_dim: int,
+            vision_cfg: CLIPVisionCfg,
+            text_cfg: CLIPTextCfg,
+            quick_gelu: bool = False,
+            cast_dtype: Optional[torch.dtype] = None,
+            output_dict: bool = False,
+    ):
+        super().__init__()
+        self.output_dict = output_dict
+
+        self.visual = _build_vision_tower(embed_dim, vision_cfg, quick_gelu, cast_dtype)

minigemini/model/multimodal_projector/builder.py

@@ -0,0 +1,50 @@
+import torch
+import torch.nn as nn
+import re
+
+class IdentityMap(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x, *args, **kwargs):
+        return x
+
+    @property
+    def config(self):
+        return {"mm_projector_type": 'identity'}
+
+
+class SimpleResBlock(nn.Module):
+    def __init__(self, channels):
+        super().__init__()
+        self.pre_norm = nn.LayerNorm(channels)
+
+        self.proj = nn.Sequential(
+            nn.Linear(channels, channels),
+            nn.GELU(),
+            nn.Linear(channels, channels)
+        )
+    def forward(self, x):
+        x = self.pre_norm(x)
+        return x + self.proj(x)
+
+
+def build_vision_projector(config, delay_load=False, **kwargs):
+    projector_type = getattr(config, 'mm_projector_type', 'linear')
+
+    if projector_type == 'linear':
+        return nn.Linear(config.mm_hidden_size, config.hidden_size)
+
+    mlp_gelu_match = re.match(r'^mlp(\d+)x_gelu$', projector_type)
+    if mlp_gelu_match:
+        mlp_depth = int(mlp_gelu_match.group(1))
+        modules = [nn.Linear(config.mm_hidden_size, config.hidden_size)]
+        for _ in range(1, mlp_depth):
+            modules.append(nn.GELU())
+            modules.append(nn.Linear(config.hidden_size, config.hidden_size))
+        return nn.Sequential(*modules)
+
+    if projector_type == 'identity':
+        return IdentityMap()
+
+    raise ValueError(f'Unknown projector type: {projector_type}')

minigemini/model/processor/video_processor.py

@@ -0,0 +1,74 @@
+from transformers import CLIPImageProcessor
+from transformers.image_processing_utils import BatchFeature, get_size_dict
+from transformers.image_transforms import get_resize_output_image_size
+
+import torch
+import torch.nn.functional as F
+
+import numpy as np
+
+
+class VideoFramesProcessor(CLIPImageProcessor):
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+
+    def preprocess(self, images, **kwargs):
+        if not isinstance(images, np.ndarray):
+            return super().preprocess(images=images, **kwargs)
+        
+        do_resize = kwargs.get('do_resize', self.do_resize)
+        size = kwargs.get('size', self.size)
+        size = get_size_dict(size, param_name="size", default_to_square=False)
+        do_center_crop = kwargs.get('do_center_crop', self.do_center_crop)
+        crop_size = kwargs.get('crop_size', self.crop_size)
+        crop_size = get_size_dict(crop_size, param_name="crop_size", default_to_square=True)
+        do_rescale = kwargs.get('do_rescale', self.do_rescale)
+        rescale_factor = kwargs.get('rescale_factor', self.rescale_factor)
+        do_normalize = kwargs.get('do_normalize', self.do_normalize)
+        image_mean = kwargs.get('image_mean', self.image_mean)
+        image_std = kwargs.get('image_std', self.image_std)
+        return_tensors = kwargs.get('return_tensors', None)
+
+        def resize(images, output_size):
+            images = images.permute((0, 3, 1, 2))
+            images = F.interpolate(images, size=output_size, mode='bicubic')
+            images = images.permute((0, 2, 3, 1))
+            return images
+
+        def center_crop(images, crop_size):
+            crop_width, crop_height = crop_size["width"], crop_size["height"]
+            img_width, img_height = images.shape[1:3]
+            x = (img_width - crop_width) // 2
+            y = (img_height - crop_height) // 2
+            images = images[:, x:x+crop_width, y:y+crop_height]
+            return images
+        
+        def rescale(images, rescale_factor):
+            images = images * rescale_factor
+            return images
+        
+        def normalize(images, mean, std):
+            mean = torch.tensor(mean)
+            std = torch.tensor(std)
+            images = (images - mean) / std
+            return images
+
+        images = torch.from_numpy(images).float()
+
+        if do_resize:
+            output_size = get_resize_output_image_size(images[0], size=size["shortest_edge"], default_to_square=False)
+            images = resize(images, output_size)
+        
+        if do_center_crop:
+            images = center_crop(images, crop_size)
+        
+        if do_rescale:
+            images = rescale(images, rescale_factor)
+        
+        if do_normalize:
+            images = normalize(images, image_mean, image_std)
+
+        images = images.permute((0, 3, 1, 2))
+        data = {"pixel_values": images}
+        return BatchFeature(data=data, tensor_type=return_tensors)

minigemini/serve/cli.py

@@ -0,0 +1,237 @@
+import argparse
+import torch
+
+from minigemini.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+from minigemini.conversation import conv_templates, SeparatorStyle
+from minigemini.model.builder import load_pretrained_model
+from minigemini.utils import disable_torch_init
+from minigemini.mm_utils import process_images, tokenizer_image_token, get_model_name_from_path, KeywordsStoppingCriteria
+
+from PIL import Image
+
+import requests
+from PIL import Image
+from io import BytesIO
+from transformers import TextStreamer
+try:
+    from diffusers import StableDiffusionXLPipeline
+except:
+    print('please install diffusers==0.26.3')
+
+try:
+    from paddleocr import PaddleOCR
+except:
+    print('please install paddleocr following https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.7/README_en.md')
+
+
+def load_image(image_file):
+    if image_file.startswith('http://') or image_file.startswith('https://'):
+        response = requests.get(image_file)
+        image = Image.open(BytesIO(response.content)).convert('RGB')
+    else:
+        image = Image.open(image_file).convert('RGB')
+    return image
+
+
+def main(args):
+    # Model
+    disable_torch_init()
+    
+    if args.ocr and args.image_file is not None:
+        ocr = PaddleOCR(use_angle_cls=True, use_gpu=True, lang="ch")
+        result = ocr.ocr(args.image_file)   
+        str_in_image = ''
+        if result[0] is not None:
+            result = [res[1][0] for res in result[0] if res[1][1] > 0.1]
+            if len(result) > 0:
+                str_in_image = ', '.join(result)
+                print('OCR Token: ' + str_in_image)
+    
+    if args.gen:
+        pipe = StableDiffusionXLPipeline.from_pretrained(
+            "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16, use_safetensors=True, variant="fp16"
+        ).to("cuda")
+
+    model_name = get_model_name_from_path(args.model_path)
+    tokenizer, model, image_processor, context_len = load_pretrained_model(args.model_path, args.model_base, model_name, args.load_8bit, args.load_4bit, device=args.device)
+    
+    if '8x7b' in model_name.lower():
+        conv_mode = "mistral_instruct"
+    elif '34b' in model_name.lower():
+        conv_mode = "chatml_direct"
+    elif '2b' in model_name.lower():
+        conv_mode = "gemma"
+    else:
+        conv_mode = "vicuna_v1"
+
+    if args.conv_mode is not None and conv_mode != args.conv_mode:
+        print('[WARNING] the auto inferred conversation mode is {}, while `--conv-mode` is {}, using {}'.format(conv_mode, args.conv_mode, args.conv_mode))
+    else:
+        args.conv_mode = conv_mode
+
+    conv = conv_templates[args.conv_mode].copy()
+    if "mpt" in model_name.lower():
+        roles = ('user', 'assistant')
+    else:
+        roles = conv.roles
+
+    if args.image_file is not None:
+        images = []
+        if ',' in args.image_file:
+            images = args.image_file.split(',')
+        else:
+            images = [args.image_file]
+        
+        image_convert = []
+        for _image in images:
+            image_convert.append(load_image(_image))
+    
+        if hasattr(model.config, 'image_size_aux'):
+            if not hasattr(image_processor, 'image_size_raw'):
+                image_processor.image_size_raw = image_processor.crop_size.copy()
+            image_processor.crop_size['height'] = model.config.image_size_aux
+            image_processor.crop_size['width'] = model.config.image_size_aux
+            image_processor.size['shortest_edge'] = model.config.image_size_aux
+        
+        # Similar operation in model_worker.py
+        image_tensor = process_images(image_convert, image_processor, model.config)
+    
+        image_grid = getattr(model.config, 'image_grid', 1)
+        if hasattr(model.config, 'image_size_aux'):
+            raw_shape = [image_processor.image_size_raw['height'] * image_grid,
+                        image_processor.image_size_raw['width'] * image_grid]
+            image_tensor_aux = image_tensor 
+            image_tensor = torch.nn.functional.interpolate(image_tensor,
+                                                        size=raw_shape,
+                                                        mode='bilinear',
+                                                        align_corners=False)
+        else:
+            image_tensor_aux = []
+
+        if image_grid >= 2:            
+            raw_image = image_tensor.reshape(3, 
+                                            image_grid,
+                                            image_processor.image_size_raw['height'],
+                                            image_grid,
+                                            image_processor.image_size_raw['width'])
+            raw_image = raw_image.permute(1, 3, 0, 2, 4)
+            raw_image = raw_image.reshape(-1, 3,
+                                        image_processor.image_size_raw['height'],
+                                        image_processor.image_size_raw['width'])
+                    
+            if getattr(model.config, 'image_global', False):
+                global_image = image_tensor
+                if len(global_image.shape) == 3:
+                    global_image = global_image[None]
+                global_image = torch.nn.functional.interpolate(global_image, 
+                                                            size=[image_processor.image_size_raw['height'],
+                                                                    image_processor.image_size_raw['width']], 
+                                                            mode='bilinear', 
+                                                            align_corners=False)
+                # [image_crops, image_global]
+                raw_image = torch.cat([raw_image, global_image], dim=0)
+            image_tensor = raw_image.contiguous()
+            image_tensor = image_tensor.unsqueeze(0)
+    
+        if type(image_tensor) is list:
+            image_tensor = [image.to(model.device, dtype=torch.float16) for image in image_tensor]
+            image_tensor_aux = [image.to(model.device, dtype=torch.float16) for image in image_tensor_aux]
+        else:
+            image_tensor = image_tensor.to(model.device, dtype=torch.float16)
+            image_tensor_aux = image_tensor_aux.to(model.device, dtype=torch.float16)
+    else:
+        images = None
+        image_tensor = None
+        image_tensor_aux = []
+
+
+    while True:
+        try:
+            inp = input(f"{roles[0]}: ")
+        except EOFError:
+            inp = ""
+        if not inp:
+            print("exit...")
+            break
+
+        print(f"{roles[1]}: ", end="")
+
+        if args.ocr and len(str_in_image) > 0:
+            inp = inp + '\nReference OCR Token: ' + str_in_image + '\n'
+        if args.gen:
+            inp = inp + ' <GEN>'
+        # print(inp, '====')
+
+        if images is not None:
+            # first message
+            if model.config.mm_use_im_start_end:
+                inp = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\n' + inp
+            else:
+                inp = (DEFAULT_IMAGE_TOKEN + '\n')*len(images) + inp
+            conv.append_message(conv.roles[0], inp)
+            images = None
+        else:
+            # later messages
+            conv.append_message(conv.roles[0], inp)
+        conv.append_message(conv.roles[1], None)
+        prompt = conv.get_prompt()
+        
+        # add image split string
+        if prompt.count(DEFAULT_IMAGE_TOKEN) >= 2:
+            final_str = ''
+            sent_split = prompt.split(DEFAULT_IMAGE_TOKEN)
+            for _idx, _sub_sent in enumerate(sent_split):
+                if _idx == len(sent_split) - 1:
+                    final_str = final_str + _sub_sent
+                else:
+                    final_str = final_str + _sub_sent + f'Image {_idx+1}:' + DEFAULT_IMAGE_TOKEN
+            prompt = final_str
+        
+        input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).to(model.device)
+        streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
+
+        with torch.inference_mode():
+            output_ids = model.generate(
+                input_ids,
+                images=image_tensor,
+                images_aux=image_tensor_aux if len(image_tensor_aux)>0 else None,
+                do_sample=True if args.temperature > 0 else False,
+                temperature=args.temperature,
+                max_new_tokens=args.max_new_tokens,
+                bos_token_id=tokenizer.bos_token_id,  # Begin of sequence token
+                eos_token_id=tokenizer.eos_token_id,  # End of sequence token
+                pad_token_id=tokenizer.pad_token_id,  # Pad token
+                streamer=streamer,
+                use_cache=True)
+
+        outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0].strip()
+        conv.messages[-1][-1] = outputs
+        
+        if args.gen and '<h>' in outputs and '</h>' in outputs:
+            common_neg_prompt = "out of frame, lowres, text, error, cropped, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, out of frame, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck, username, watermark, signature"
+            prompt = outputs.split("</h>")[-2].split("<h>")[-1]
+            output_img = pipe(prompt, negative_prompt=common_neg_prompt).images[0]
+            output_img.save(args.output_file)
+            print(f'Generate an image, save at {args.output_file}')
+
+        if args.debug:
+            print("\n", {"prompt": prompt, "outputs": outputs}, "\n")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model-path", type=str, default="facebook/opt-350m")
+    parser.add_argument("--model-base", type=str, default=None)
+    parser.add_argument("--image-file", type=str, default=None) # file_0.jpg,file_1.jpg for multi image
+    parser.add_argument("--device", type=str, default="cuda")
+    parser.add_argument("--conv-mode", type=str, default=None)
+    parser.add_argument("--temperature", type=float, default=0.2)
+    parser.add_argument("--max-new-tokens", type=int, default=512)
+    parser.add_argument("--load-8bit", action="store_true")
+    parser.add_argument("--load-4bit", action="store_true")
+    parser.add_argument("--ocr", action="store_true")
+    parser.add_argument("--gen", action="store_true")
+    parser.add_argument("--output-file", type=str, default='generate.png')
+    parser.add_argument("--debug", action="store_true")
+    args = parser.parse_args()
+    main(args)

minigemini/serve/controller.py

@@ -0,0 +1,298 @@
+"""
+A controller manages distributed workers.
+It sends worker addresses to clients.
+"""
+import argparse
+import asyncio
+import dataclasses
+from enum import Enum, auto
+import json
+import logging
+import time
+from typing import List, Union
+import threading
+
+from fastapi import FastAPI, Request
+from fastapi.responses import StreamingResponse
+import numpy as np
+import requests
+import uvicorn
+
+from minigemini.constants import CONTROLLER_HEART_BEAT_EXPIRATION
+from minigemini.utils import build_logger, server_error_msg
+
+
+logger = build_logger("controller", "controller.log")
+
+
+class DispatchMethod(Enum):
+    LOTTERY = auto()
+    SHORTEST_QUEUE = auto()
+
+    @classmethod
+    def from_str(cls, name):
+        if name == "lottery":
+            return cls.LOTTERY
+        elif name == "shortest_queue":
+            return cls.SHORTEST_QUEUE
+        else:
+            raise ValueError(f"Invalid dispatch method")
+
+
+@dataclasses.dataclass
+class WorkerInfo:
+    model_names: List[str]
+    speed: int
+    queue_length: int
+    check_heart_beat: bool
+    last_heart_beat: str
+
+
+def heart_beat_controller(controller):
+    while True:
+        time.sleep(CONTROLLER_HEART_BEAT_EXPIRATION)
+        controller.remove_stable_workers_by_expiration()
+
+
+class Controller:
+    def __init__(self, dispatch_method: str):
+        # Dict[str -> WorkerInfo]
+        self.worker_info = {}
+        self.dispatch_method = DispatchMethod.from_str(dispatch_method)
+
+        self.heart_beat_thread = threading.Thread(
+            target=heart_beat_controller, args=(self,))
+        self.heart_beat_thread.start()
+
+        logger.info("Init controller")
+
+    def register_worker(self, worker_name: str, check_heart_beat: bool,
+                        worker_status: dict):
+        if worker_name not in self.worker_info:
+            logger.info(f"Register a new worker: {worker_name}")
+        else:
+            logger.info(f"Register an existing worker: {worker_name}")
+
+        if not worker_status:
+            worker_status = self.get_worker_status(worker_name)
+        if not worker_status:
+            return False
+
+        self.worker_info[worker_name] = WorkerInfo(
+            worker_status["model_names"], worker_status["speed"], worker_status["queue_length"],
+            check_heart_beat, time.time())
+
+        logger.info(f"Register done: {worker_name}, {worker_status}")
+        return True
+
+    def get_worker_status(self, worker_name: str):
+        try:
+            r = requests.post(worker_name + "/worker_get_status", timeout=5)
+        except requests.exceptions.RequestException as e:
+            logger.error(f"Get status fails: {worker_name}, {e}")
+            return None
+
+        if r.status_code != 200:
+            logger.error(f"Get status fails: {worker_name}, {r}")
+            return None
+
+        return r.json()
+
+    def remove_worker(self, worker_name: str):
+        del self.worker_info[worker_name]
+
+    def refresh_all_workers(self):
+        old_info = dict(self.worker_info)
+        self.worker_info = {}
+
+        for w_name, w_info in old_info.items():
+            if not self.register_worker(w_name, w_info.check_heart_beat, None):
+                logger.info(f"Remove stale worker: {w_name}")
+
+    def list_models(self):
+        model_names = set()
+
+        for w_name, w_info in self.worker_info.items():
+            model_names.update(w_info.model_names)
+
+        return list(model_names)
+
+    def get_worker_address(self, model_name: str):
+        if self.dispatch_method == DispatchMethod.LOTTERY:
+            worker_names = []
+            worker_speeds = []
+            for w_name, w_info in self.worker_info.items():
+                if model_name in w_info.model_names:
+                    worker_names.append(w_name)
+                    worker_speeds.append(w_info.speed)
+            worker_speeds = np.array(worker_speeds, dtype=np.float32)
+            norm = np.sum(worker_speeds)
+            if norm < 1e-4:
+                return ""
+            worker_speeds = worker_speeds / norm
+            if True:  # Directly return address
+                pt = np.random.choice(np.arange(len(worker_names)),
+                    p=worker_speeds)
+                worker_name = worker_names[pt]
+                return worker_name
+
+            # Check status before returning
+            while True:
+                pt = np.random.choice(np.arange(len(worker_names)),
+                    p=worker_speeds)
+                worker_name = worker_names[pt]
+
+                if self.get_worker_status(worker_name):
+                    break
+                else:
+                    self.remove_worker(worker_name)
+                    worker_speeds[pt] = 0
+                    norm = np.sum(worker_speeds)
+                    if norm < 1e-4:
+                        return ""
+                    worker_speeds = worker_speeds / norm
+                    continue
+            return worker_name
+        elif self.dispatch_method == DispatchMethod.SHORTEST_QUEUE:
+            worker_names = []
+            worker_qlen = []
+            for w_name, w_info in self.worker_info.items():
+                if model_name in w_info.model_names:
+                    worker_names.append(w_name)
+                    worker_qlen.append(w_info.queue_length / w_info.speed)
+            if len(worker_names) == 0:
+                return ""
+            min_index = np.argmin(worker_qlen)
+            w_name = worker_names[min_index]
+            self.worker_info[w_name].queue_length += 1
+            logger.info(f"names: {worker_names}, queue_lens: {worker_qlen}, ret: {w_name}")
+            return w_name
+        else:
+            raise ValueError(f"Invalid dispatch method: {self.dispatch_method}")
+
+    def receive_heart_beat(self, worker_name: str, queue_length: int):
+        if worker_name not in self.worker_info:
+            logger.info(f"Receive unknown heart beat. {worker_name}")
+            return False
+
+        self.worker_info[worker_name].queue_length = queue_length
+        self.worker_info[worker_name].last_heart_beat = time.time()
+        logger.info(f"Receive heart beat. {worker_name}")
+        return True
+
+    def remove_stable_workers_by_expiration(self):
+        expire = time.time() - CONTROLLER_HEART_BEAT_EXPIRATION
+        to_delete = []
+        for worker_name, w_info in self.worker_info.items():
+            if w_info.check_heart_beat and w_info.last_heart_beat < expire:
+                to_delete.append(worker_name)
+
+        for worker_name in to_delete:
+            self.remove_worker(worker_name)
+
+    def worker_api_generate_stream(self, params):
+        worker_addr = self.get_worker_address(params["model"])
+        if not worker_addr:
+            logger.info(f"no worker: {params['model']}")
+            ret = {
+                "text": server_error_msg,
+                "error_code": 2,
+            }
+            yield json.dumps(ret).encode() + b"\0"
+
+        try:
+            response = requests.post(worker_addr + "/worker_generate_stream",
+                json=params, stream=True, timeout=5)
+            for chunk in response.iter_lines(decode_unicode=False, delimiter=b"\0"):
+                if chunk:
+                    yield chunk + b"\0"
+        except requests.exceptions.RequestException as e:
+            logger.info(f"worker timeout: {worker_addr}")
+            ret = {
+                "text": server_error_msg,
+                "error_code": 3,
+            }
+            yield json.dumps(ret).encode() + b"\0"
+
+
+    # Let the controller act as a worker to achieve hierarchical
+    # management. This can be used to connect isolated sub networks.
+    def worker_api_get_status(self):
+        model_names = set()
+        speed = 0
+        queue_length = 0
+
+        for w_name in self.worker_info:
+            worker_status = self.get_worker_status(w_name)
+            if worker_status is not None:
+                model_names.update(worker_status["model_names"])
+                speed += worker_status["speed"]
+                queue_length += worker_status["queue_length"]
+
+        return {
+            "model_names": list(model_names),
+            "speed": speed,
+            "queue_length": queue_length,
+        }
+
+
+app = FastAPI()
+
+
+@app.post("/register_worker")
+async def register_worker(request: Request):
+    data = await request.json()
+    controller.register_worker(
+        data["worker_name"], data["check_heart_beat"],
+        data.get("worker_status", None))
+
+
+@app.post("/refresh_all_workers")
+async def refresh_all_workers():
+    models = controller.refresh_all_workers()
+
+
+@app.post("/list_models")
+async def list_models():
+    models = controller.list_models()
+    return {"models": models}
+
+
+@app.post("/get_worker_address")
+async def get_worker_address(request: Request):
+    data = await request.json()
+    addr = controller.get_worker_address(data["model"])
+    return {"address": addr}
+
+
+@app.post("/receive_heart_beat")
+async def receive_heart_beat(request: Request):
+    data = await request.json()
+    exist = controller.receive_heart_beat(
+        data["worker_name"], data["queue_length"])
+    return {"exist": exist}
+
+
+@app.post("/worker_generate_stream")
+async def worker_api_generate_stream(request: Request):
+    params = await request.json()
+    generator = controller.worker_api_generate_stream(params)
+    return StreamingResponse(generator)
+
+
+@app.post("/worker_get_status")
+async def worker_api_get_status(request: Request):
+    return controller.worker_api_get_status()
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--host", type=str, default="localhost")
+    parser.add_argument("--port", type=int, default=21001)
+    parser.add_argument("--dispatch-method", type=str, choices=[
+        "lottery", "shortest_queue"], default="shortest_queue")
+    args = parser.parse_args()
+    logger.info(f"args: {args}")
+
+    controller = Controller(args.dispatch_method)
+    uvicorn.run(app, host=args.host, port=args.port, log_level="info")