Update app.py

app.py

@@ -1,81 +1,24 @@
-import spaces
-import torch
-import re
 import gradio as gr
-from threading import Thread
-from transformers import TextIteratorStreamer, AutoTokenizer, AutoModelForCausalLM
-from PIL import ImageDraw
-from torchvision.transforms.v2 import Resize
+import moondream as md
+import os
 
-import subprocess
-subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)
+moondream_api_key = os.getenv("MOONDREAM_API_KEY")
+model = md.vl(api_key=moondream_api_key)
 
-model_id = "vikhyatk/moondream2"
-revision = "2024-08-26"
-tokenizer = AutoTokenizer.from_pretrained(model_id, revision=revision)
-moondream = AutoModelForCausalLM.from_pretrained(
-    model_id, trust_remote_code=True, revision=revision,
-    torch_dtype=torch.bfloat16, device_map={"": "cuda"},
-    attn_implementation="flash_attention_2"
-)
-moondream.eval()
-
-
-@spaces.GPU(duration=10)
 def answer_question(img, prompt):
-    image_embeds = moondream.encode_image(img)
-    streamer = TextIteratorStreamer(tokenizer, skip_special_tokens=True)
-    thread = Thread(
-        target=moondream.answer_question,
-        kwargs={
-            "image_embeds": image_embeds,
-            "question": prompt,
-            "tokenizer": tokenizer,
-            "streamer": streamer,
-        },
-    )
-    thread.start()
-
     buffer = ""
-    for new_text in streamer:
-        buffer += new_text
-        yield buffer.strip()
-
-def extract_floats(text):
-    # Regular expression to match an array of four floating point numbers
-    pattern = r"\[\s*(-?\d+\.\d+)\s*,\s*(-?\d+\.\d+)\s*,\s*(-?\d+\.\d+)\s*,\s*(-?\d+\.\d+)\s*\]"
-    match = re.search(pattern, text)
-    if match:
-        # Extract the numbers and convert them to floats
-        return [float(num) for num in match.groups()]
-    return None  # Return None if no match is found
-
-
-def extract_bbox(text):
-    bbox = None
-    if extract_floats(text) is not None:
-        x1, y1, x2, y2 = extract_floats(text)
-        bbox = (x1, y1, x2, y2)
-    return bbox
+    for chunk in model.query(img, prompt, stream=True)["answer"]:
+        buffer += chunk
+        yield buffer
 
 def process_answer(img, answer):
-    if extract_bbox(answer) is not None:
-        x1, y1, x2, y2 = extract_bbox(answer)
-        draw_image = Resize(768)(img)
-        width, height = draw_image.size
-        x1, x2 = int(x1 * width), int(x2 * width)
-        y1, y2 = int(y1 * height), int(y2 * height)
-        bbox = (x1, y1, x2, y2)
-        ImageDraw.Draw(draw_image).rectangle(bbox, outline="red", width=3)
-        return gr.update(visible=True, value=draw_image)
-
     return gr.update(visible=False, value=None)
 
 with gr.Blocks() as demo:
     gr.Markdown(
         """
         # 🌔 moondream2
-        A tiny vision language model. [GitHub](https://github.com/vikhyat/moondream)
+        A tiny vision language model. Check out other capabilities (object detection, pointing etc.) in the [Moondream Playground](https://moondream.ai/playground).
         """
     )
     with gr.Row():