Initial commit

9em124t2-499968/clip_model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7d7b0548d12fa649370896982c2af9d03d43285b782bd47639c96e6e0b29473c
+size 1713067838

9em124t2-499968/config.yaml

@@ -0,0 +1,39 @@
+wandb_project: joy-caption-1
+device_batch_size: 2
+batch_size: 256
+learning_rate: 0.0002
+warmup_samples: 18000
+max_samples: 500000
+save_every: 50000
+test_every: 50000
+use_amp: true
+grad_scaler: true
+lr_scheduler_type: cosine
+min_lr_ratio: 0.0
+allow_tf32: true
+seed: 69
+num_workers: 8
+optimizer_type: adamw
+adam_beta1: 0.9
+adam_beta2: 0.999
+adam_eps: 1.0e-08
+adam_weight_decay: 0.0
+clip_grad_norm: 1.0
+dataset: fancyfeast/joy-captioning-20240917a
+clip_model: google/siglip-so400m-patch14-384
+text_model: meta-llama/Meta-Llama-3.1-8B
+resume: null
+gradient_checkpointing: false
+test_size: 2048
+grad_scaler_init: 65536.0
+max_caption_length: 257
+num_image_tokens: 32
+adapter_type: mlp
+text_model_dtype: bfloat16
+pre_test: false
+train_image_model: true
+image_model_lr: null
+train_lora: true
+lora_r: 64
+lora_alpha: 16
+lora_dropout: 0.1

9em124t2-499968/image_adapter.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e53c3bf8df745a3c19ae3c70dbf9bf23cfdc8f3fdb937000a4eafd2a36914661
+size 86067714

9em124t2-499968/text_model/README.md

@@ -0,0 +1,202 @@
+---
+base_model: meta-llama/Meta-Llama-3.1-8B
+library_name: peft
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.12.0

9em124t2-499968/text_model/adapter_config.json

@@ -0,0 +1,29 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "meta-llama/Meta-Llama-3.1-8B",
+  "bias": "none",
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 16,
+  "lora_dropout": 0.1,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 64,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "q_proj",
+    "v_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "use_dora": false,
+  "use_rslora": false
+}

9em124t2-499968/text_model/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b48221de174ab0db7b46b4833118c5c0a4c2bf0b51b77b4cc4ab04651bd06cca
+size 109069176

app.py

@@ -0,0 +1,242 @@
+import spaces
+import gradio as gr
+from huggingface_hub import InferenceClient
+from torch import nn
+from transformers import AutoModel, AutoProcessor, AutoTokenizer, PreTrainedTokenizer, PreTrainedTokenizerFast, AutoModelForCausalLM
+from pathlib import Path
+import torch
+import torch.amp.autocast_mode
+from PIL import Image
+import os
+import torchvision.transforms.functional as TVF
+
+
+CLIP_PATH = "google/siglip-so400m-patch14-384"
+MODEL_PATH = "meta-llama/Meta-Llama-3.1-8B"
+CHECKPOINT_PATH = Path("9em124t2-499968")
+TITLE = "<h1><center>JoyCaption Alpha One (2024-09-20a)</center></h1>"
+CAPTION_TYPE_MAP = {
+	("descriptive", "formal", False, False): ["Write a descriptive caption for this image in a formal tone."],
+	("descriptive", "formal", False, True): ["Write a descriptive caption for this image in a formal tone within {word_count} words."],
+	("descriptive", "formal", True, False): ["Write a {length} descriptive caption for this image in a formal tone."],
+	("descriptive", "informal", False, False): ["Write a descriptive caption for this image in a casual tone."],
+	("descriptive", "informal", False, True): ["Write a descriptive caption for this image in a casual tone within {word_count} words."],
+	("descriptive", "informal", True, False): ["Write a {length} descriptive caption for this image in a casual tone."],
+
+	("training_prompt", "formal", False, False): ["Write a stable diffusion prompt for this image."],
+	("training_prompt", "formal", False, True): ["Write a stable diffusion prompt for this image within {word_count} words."],
+	("training_prompt", "formal", True, False): ["Write a {length} stable diffusion prompt for this image."],
+
+	("rng-tags", "formal", False, False): ["Write a list of Booru tags for this image."],
+	("rng-tags", "formal", False, True): ["Write a list of Booru tags for this image within {word_count} words."],
+	("rng-tags", "formal", True, False): ["Write a {length} list of Booru tags for this image."],
+}
+
+HF_TOKEN = os.environ.get("HF_TOKEN", None)
+
+
+class ImageAdapter(nn.Module):
+	def __init__(self, input_features: int, output_features: int, ln1: bool, pos_emb: bool, num_image_tokens: int, deep_extract: bool):
+		super().__init__()
+		self.deep_extract = deep_extract
+
+		if self.deep_extract:
+			input_features = input_features * 5
+
+		self.linear1 = nn.Linear(input_features, output_features)
+		self.activation = nn.GELU()
+		self.linear2 = nn.Linear(output_features, output_features)
+		self.ln1 = nn.Identity() if not ln1 else nn.LayerNorm(input_features)
+		self.pos_emb = None if not pos_emb else nn.Parameter(torch.zeros(num_image_tokens, input_features))
+
+		# Mode token
+		#self.mode_token = nn.Embedding(n_modes, output_features)
+		#self.mode_token.weight.data.normal_(mean=0.0, std=0.02)   # Matches HF's implementation of llama3
+
+		# Other tokens (<|image_start|>, <|image_end|>, <|eot_id|>)
+		self.other_tokens = nn.Embedding(3, output_features)
+		self.other_tokens.weight.data.normal_(mean=0.0, std=0.02)   # Matches HF's implementation of llama3
+
+	def forward(self, vision_outputs: torch.Tensor):
+		if self.deep_extract:
+			x = torch.concat((
+				vision_outputs[-2],
+				vision_outputs[3],
+				vision_outputs[7],
+				vision_outputs[13],
+				vision_outputs[20],
+			), dim=-1)
+			assert len(x.shape) == 3, f"Expected 3, got {len(x.shape)}"  # batch, tokens, features
+			assert x.shape[-1] == vision_outputs[-2].shape[-1] * 5, f"Expected {vision_outputs[-2].shape[-1] * 5}, got {x.shape[-1]}"
+		else:
+			x = vision_outputs[-2]
+
+		x = self.ln1(x)
+
+		if self.pos_emb is not None:
+			assert x.shape[-2:] == self.pos_emb.shape, f"Expected {self.pos_emb.shape}, got {x.shape[-2:]}"
+			x = x + self.pos_emb
+
+		x = self.linear1(x)
+		x = self.activation(x)
+		x = self.linear2(x)
+
+		# Mode token
+		#mode_token = self.mode_token(mode)
+		#assert mode_token.shape == (x.shape[0], mode_token.shape[1], x.shape[2]), f"Expected {(x.shape[0], 1, x.shape[2])}, got {mode_token.shape}"
+		#x = torch.cat((x, mode_token), dim=1)
+
+		# <|image_start|>, IMAGE, <|image_end|>
+		other_tokens = self.other_tokens(torch.tensor([0, 1], device=self.other_tokens.weight.device).expand(x.shape[0], -1))
+		assert other_tokens.shape == (x.shape[0], 2, x.shape[2]), f"Expected {(x.shape[0], 2, x.shape[2])}, got {other_tokens.shape}"
+		x = torch.cat((other_tokens[:, 0:1], x, other_tokens[:, 1:2]), dim=1)
+
+		return x
+
+	def get_eot_embedding(self):
+		return self.other_tokens(torch.tensor([2], device=self.other_tokens.weight.device)).squeeze(0)
+
+
+
+# Load CLIP
+print("Loading CLIP")
+clip_processor = AutoProcessor.from_pretrained(CLIP_PATH)
+clip_model = AutoModel.from_pretrained(CLIP_PATH)
+clip_model = clip_model.vision_model
+
+if (CHECKPOINT_PATH / "clip_model.pt").exists():
+	print("Loading VLM's custom vision model")
+	checkpoint = torch.load(CHECKPOINT_PATH / "clip_model.pt", map_location='cpu')
+	checkpoint = {k.replace("_orig_mod.module.", ""): v for k, v in checkpoint.items()}
+	clip_model.load_state_dict(checkpoint)
+	del checkpoint
+
+clip_model.eval()
+clip_model.requires_grad_(False)
+clip_model.to("cuda")
+
+
+# Tokenizer
+print("Loading tokenizer")
+tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH, use_fast=False)
+assert isinstance(tokenizer, PreTrainedTokenizer) or isinstance(tokenizer, PreTrainedTokenizerFast), f"Tokenizer is of type {type(tokenizer)}"
+
+# LLM
+print("Loading LLM")
+if (CHECKPOINT_PATH / "text_model").exists:
+	print("Loading VLM's custom text model")
+	text_model = AutoModelForCausalLM.from_pretrained(CHECKPOINT_PATH / "text_model", device_map=0, torch_dtype=torch.bfloat16)
+else:
+	text_model = AutoModelForCausalLM.from_pretrained(MODEL_PATH, device_map="auto", torch_dtype=torch.bfloat16)
+
+text_model.eval()
+
+# Image Adapter
+print("Loading image adapter")
+image_adapter = ImageAdapter(clip_model.config.hidden_size, text_model.config.hidden_size, False, False, 38, False)
+image_adapter.load_state_dict(torch.load(CHECKPOINT_PATH / "image_adapter.pt", map_location="cpu"))
+image_adapter.eval()
+image_adapter.to("cuda")
+
+
+@spaces.GPU()
+@torch.no_grad()
+def stream_chat(input_image: Image.Image, caption_type: str, caption_tone: str, caption_length: str | int) -> str:
+	torch.cuda.empty_cache()
+
+	length = None if caption_length == "any" else caption_length
+	prompt_key = (caption_type, caption_tone, isinstance(length, str), isinstance(length, int))
+	if prompt_key not in CAPTION_TYPE_MAP:
+		raise ValueError(f"Invalid caption type: {prompt_key}")
+
+	prompt_str = CAPTION_TYPE_MAP[prompt_key][0]
+
+	# Preprocess image
+	#image = clip_processor(images=input_image, return_tensors='pt').pixel_values
+	image = input_image.resize((384, 384), Image.LANCZOS)
+	pixel_values = TVF.pil_to_tensor(image).unsqueeze(0) / 255.0
+	pixel_values = TVF.normalize(pixel_values, [0.5], [0.5])
+	pixel_values = pixel_values.to('cuda')
+
+	# Tokenize the prompt
+	prompt = tokenizer.encode(prompt_str, return_tensors='pt', padding=False, truncation=False, add_special_tokens=False)
+
+	# Embed image
+	with torch.amp.autocast_mode.autocast('cuda', enabled=True):
+		vision_outputs = clip_model(pixel_values=pixel_values, output_hidden_states=True)
+		image_features = vision_outputs.hidden_states
+		embedded_images = image_adapter(image_features)
+		embedded_images = embedded_images.to('cuda')
+	
+	# Embed prompt
+	prompt_embeds = text_model.model.embed_tokens(prompt.to('cuda'))
+	assert prompt_embeds.shape == (1, prompt.shape[1], text_model.config.hidden_size), f"Prompt shape is {prompt_embeds.shape}, expected {(1, prompt.shape[1], text_model.config.hidden_size)}"
+	embedded_bos = text_model.model.embed_tokens(torch.tensor([[tokenizer.bos_token_id]], device=text_model.device, dtype=torch.int64))
+	eot_embed = image_adapter.get_eot_embedding().unsqueeze(0).to(dtype=text_model.dtype)
+
+	# Construct prompts
+	inputs_embeds = torch.cat([
+		embedded_bos.expand(embedded_images.shape[0], -1, -1),
+		embedded_images.to(dtype=embedded_bos.dtype),
+		prompt_embeds.expand(embedded_images.shape[0], -1, -1),
+		eot_embed.expand(embedded_images.shape[0], -1, -1),
+	], dim=1)
+
+	input_ids = torch.cat([
+		torch.tensor([[tokenizer.bos_token_id]], dtype=torch.long),
+		torch.zeros((1, embedded_images.shape[1]), dtype=torch.long),
+		prompt,
+		torch.tensor([[tokenizer.convert_tokens_to_ids("<|eot_id|>")]], dtype=torch.long),
+	], dim=1).to('cuda')
+	attention_mask = torch.ones_like(input_ids)
+
+	#generate_ids = text_model.generate(input_ids, inputs_embeds=inputs_embeds, attention_mask=attention_mask, max_new_tokens=300, do_sample=False, suppress_tokens=None)
+	#generate_ids = text_model.generate(input_ids, inputs_embeds=inputs_embeds, attention_mask=attention_mask, max_new_tokens=300, do_sample=True, top_k=10, temperature=0.5, suppress_tokens=None)
+	generate_ids = text_model.generate(input_ids, inputs_embeds=inputs_embeds, attention_mask=attention_mask, max_new_tokens=300, do_sample=True, suppress_tokens=None)   # Uses the default which is temp=0.6, top_p=0.9
+
+	# Trim off the prompt
+	generate_ids = generate_ids[:, input_ids.shape[1]:]
+	if generate_ids[0][-1] == tokenizer.eos_token_id or generate_ids[0][-1] == tokenizer.convert_tokens_to_ids("<|eot_id|>"):
+		generate_ids = generate_ids[:, :-1]
+
+	caption = tokenizer.batch_decode(generate_ids, skip_special_tokens=False, clean_up_tokenization_spaces=False)[0]
+
+	return caption.strip()
+
+
+with gr.Blocks() as demo:
+	gr.HTML(TITLE)
+
+	with gr.Row():
+		with gr.Column():
+			input_image = gr.Image(type="pil", label="Input Image")
+
+			caption_type = gr.Dropdown(
+				choices=["descriptive", "training_prompt", "rng-tags"],
+				label="Caption Type",
+				value="descriptive",
+			)
+
+			caption_tone = gr.Dropdown(
+				choices=["formal", "informal"],
+				label="Caption Tone",
+				value="formal",
+			)
+
+			caption_length = gr.Dropdown(
+				choices=["any", "very short", "short", "medium-length", "long", "very long"] +
+						[str(i) for i in range(20, 261, 10)],
+				label="Caption Length",
+				value="any",
+			)
+
+			run_button = gr.Button("Caption")
+		
+		with gr.Column():
+			output_caption = gr.Textbox(label="Caption")
+	
+	run_button.click(fn=stream_chat, inputs=[input_image, caption_type, caption_tone, caption_length], outputs=[output_caption])
+
+
+if __name__ == "__main__":
+    demo.launch()

requirements.txt

@@ -0,0 +1,6 @@
+huggingface_hub==0.23.4
+accelerate
+torch
+transformers==4.44.0
+sentencepiece
+peft==0.12.0