Update app.py

app.py

@@ -1,129 +1,60 @@
-import spaces
-import os
-import sys
-import subprocess
-
-def install_packages():
-    subprocess.check_call([sys.executable, "-m", "pip", "install", "unsloth-zoo"])
-    subprocess.check_call([sys.executable, "-m", "pip", "install", "--no-deps", "git+https://github.com/unslothai/unsloth.git"])
-
-try:
-    install_packages()
-except Exception as e:
-    print(f"Failed to install packages: {e}")
-
-import warnings
-import torch
-
-from transformers import TextStreamer
 import gradio as gr
-from huggingface_hub import login
+import torch
 from PIL import Image
-
-warnings.filterwarnings('ignore')
-
-model = None
-tokenizer = None
-
-###@spaces.GPU
-def load_model():
-    global model
-    print("กำลังโหลดโมเดล...")
-    try:
-        from transformers import AutoModelForVision2Seq
-        print("กำลังโหลดโมเดล fine-tuned...")
-        
-        device = "cuda" if torch.cuda.is_available() else "cpu"
-        model = AutoModelForVision2Seq.from_pretrained(
-            "0llheaven/Llama-3.2-11B-Vision-Radiology-mini",
-            # load_in_4bit = True,
-            device_map=device,
-            torch_dtype = torch.float16
-        )
-
-        print("โหลดโมเดลสำเร็จ!")
-        return True
-        
-    except Exception as e:
-        print(f"เกิดข้อผิดพลาดในการโหลดโมเดล: {str(e)}")
-        import traceback
-        traceback.print_exc()
-        return False
-
-@spaces.GPU(duration=120)
-def process_image(image):
-    global model
-
-    ### โหลด tokenizer จาก base model
-    from unsloth import FastVisionModel
-
-    FastVisionModel.for_inference(model) ###ลองแก้ไขปัญหา torch
-
-    from transformers import AutoTokenizer
-    print("กำลังโหลด tokenizer...")
-    base_model, tokenizer = FastVisionModel.from_pretrained(
-        "unsloth/Llama-3.2-11B-Vision-Instruct",
-        use_gradient_checkpointing = "unsloth",
-        ### device_map="auto"  ### เพิ่มตรงนี้
-    )
-    
-    print("\nใน process_image():")
-    print("Type of model:", type(model))
-    print("A. Type of tokenizer:", type(tokenizer))
-    if tokenizer is not None:
-        print("B. Available methods:", dir(tokenizer))
-    
-    if image is None:
-        return "กรุณาอัพโหลดรูปภาพ"
-    
-    try:
-        if not isinstance(image, Image.Image):
-            image = Image.fromarray(image)
-
-        print("0. Image info:", type(image), image.size)  # เพิ่ม debug ข้อมูลรูปภาพ
-        instruction = "You are an expert radiographer. Describe accurately what you see in this image."
-        messages = [
-            {"role": "user", "content": [
-                {"type": "image"},
-                {"type": "text", "text": instruction}
-            ]}
-        ]
-
-        print("1. Messages:", messages)  
-
-        print("2. Tokenizer type:", type(tokenizer))
-        input_text = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
-        print("3. Chat template success:", input_text[:100])
-        inputs = tokenizer(
-            image,
-            input_text,
-            add_special_tokens=False,
-            return_tensors="pt",
-        ).to("cuda")
-        print("3. Tokenizer inputs:", inputs.keys())  # Debug 3
-
-        text_streamer = TextStreamer(tokenizer, skip_prompt=True)
-        outputs = model.generate(
-            **inputs, 
-            streamer=text_streamer,
-            max_new_tokens=256,
-            use_cache=True,
-            temperature=1.5,
-            min_p=0.1
-        )
-        
-        return tokenizer.decode(outputs[0], skip_special_tokens=True).strip()
-        
-    except Exception as e:
-        return f"เกิดข้อผิดพลาด: {str(e)}"
-
-if load_model():
-    demo = gr.Interface(
-        fn=process_image,
-        inputs=gr.Image(type="pil", label="Upload Image"),
-        outputs=gr.Textbox(label="Generated Caption"),
-        title="Medical Vision Analysis"
-    )
-    
-    if __name__ == "__main__":
-        demo.launch()
+from transformers import MllamaForConditionalGeneration, AutoProcessor
+from transformers import TextStreamer
+from torchvision.transforms import Resize
+
+# Define the model and processor
+model_id = "0llheaven/Llama-3.2-11B-Vision-Radiology-mini"
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+model = MllamaForConditionalGeneration.from_pretrained(
+    model_id,
+    load_in_4bit=True,
+    torch_dtype=torch.bfloat16,
+    device_map=device,
+)
+
+model.gradient_checkpointing_enable()
+
+processor = AutoProcessor.from_pretrained(model_id)
+
+# Function to process the image and generate the description
+def generate_description(image: Image.Image, instruction: str):
+    image = image.convert("RGB")
+    # image = Resize((224, 224))(image)
+
+    # Create the message to pass to the model
+    instruction = "You are an expert radiographer. Describe accurately what you see in this image."
+    messages = [
+        {"role": "user", "content": [
+            {"type": "image"},
+            {"type": "text", "text": instruction}
+        ]}
+    ]
+
+    input_text = processor.apply_chat_template(messages, add_generation_prompt=True)
+    inputs = processor(
+        image,
+        input_text,
+        add_special_tokens=False,
+        return_tensors="pt"
+    ).to(model.device)
+
+    # Generate the output from the model
+    output = model.generate(**inputs, max_new_tokens=256)
+    return processor.decode(output[0])
+
+# Define Gradio interface
+interface = gr.Interface(
+    fn=generate_description,
+    inputs=gr.Image(type="pil", label="Upload an Image"),
+    outputs=gr.Textbox(label="Generated Description"),
+    live=True,
+    title="Radiology Image Description Generator",
+    description="Upload an image and provide an instruction to generate a description using a vision-language model."
+)
+
+# Launch the interface
+interface.launch()