Update app.py

app.py

@@ -4,13 +4,12 @@ import gradio as gr
 import tempfile
 import secrets
 from pathlib import Path
-from transformers import AutoModelForCausalLM, AutoTokenizer, BlipForConditionalGeneration, AutoProcessor, Qwen2VLForConditionalGeneration
-from qwen_vl_utils import process_vision_info
+from transformers import AutoModelForCausalLM, AutoTokenizer, AutoProcessor, Qwen2VLForConditionalGeneration
 from PIL import Image
 
 # Load Vision-Language Model
 vl_model = Qwen2VLForConditionalGeneration.from_pretrained(
-    "Qwen/Qwen2-VL-2B-Instruct", torch_dtype="auto", device_map="auto"
+    "Qwen/Qwen2-VL-2B-Instruct", torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32, device_map="auto"
 )
 vl_processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-2B-Instruct")
 
@@ -34,18 +33,17 @@ def process_image(image, shouldConvert=False):
         new_img.paste(image, (0, 0), mask=image)
         image = new_img
 
-    # Convert the image to tensor
-    inputs = vl_processor(images=image, return_tensors="pt").to(device)
-    generated_ids = vl_model.generate(**inputs)
-    generated_ids_trimmed = [
-        out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
-    ]
-    output = vl_processor.batch_decode(
-        generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
-    )
-    description = output[0] if output else ""
-    
-    return f"Math-related content detected: {description}"
+    try:
+        inputs = vl_processor(images=image, return_tensors="pt").to(device)
+        if inputs is None:
+            return "Error processing image."
+
+        generated_ids = vl_model.generate(**inputs)
+        output = vl_processor.batch_decode(generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)
+        description = output[0] if output else ""
+        return f"Math-related content detected: {description}"
+    except Exception as e:
+        return f"Error in processing image: {str(e)}"
 
 def get_math_response(image_description, user_question):
     global math_messages
@@ -55,6 +53,7 @@ def get_math_response(image_description, user_question):
     content = f'Image description: {image_description}\n\n' if image_description else ''
     query = f"{content}User question: {user_question}"
     math_messages.append({'role': 'user', 'content': query})
+    
     model_inputs = tokenizer(query, return_tensors="pt").to(device)
     output = model.generate(**model_inputs, max_new_tokens=512)
     answer = tokenizer.decode(output[0], skip_special_tokens=True)
@@ -62,14 +61,12 @@ def get_math_response(image_description, user_question):
     math_messages.append({'role': 'assistant', 'content': answer})
 
 def math_chat_bot(image, sketchpad, question, state):
-    current_tab_index = state["tab_index"]
+    current_tab_index = state.get("tab_index", 0)
     image_description = None
-    if current_tab_index == 0:
-        if image is not None:
-            image_description = process_image(image)
-    elif current_tab_index == 1:
-        if sketchpad and sketchpad.get("composite"):
-            image_description = process_image(sketchpad["composite"], True)
+    if current_tab_index == 0 and image is not None:
+        image_description = process_image(image)
+    elif current_tab_index == 1 and sketchpad and sketchpad.get("composite"):
+        image_description = process_image(sketchpad["composite"], True)
     yield from get_math_response(image_description, question)
 
 css = """
@@ -81,13 +78,12 @@ css = """
 def tabs_select(e: gr.SelectData, _state):
     _state["tab_index"] = e.index
 
-# 创建Gradio接口
 with gr.Blocks(css=css) as demo:
     gr.HTML("""
-<p align="center"><img src="https://modelscope.oss-cn-beijing.aliyuncs.com/resource/qwen.png" style="height: 60px"/><p>"""
-            """<center><font size=8>📖 Qwen2-Math Demo</center>"""
-            """
-<center><font size=3>This WebUI is based on Qwen2-VL for OCR and Qwen2-Math for mathematical reasoning. You can input either images or texts of mathematical or arithmetic problems.</center>""")
+    <p align="center"><img src="https://modelscope.oss-cn-beijing.aliyuncs.com/resource/qwen.png" style="height: 60px"/></p>
+    <center><font size=8>📖 Qwen2-Math Demo</font></center>
+    <center><font size=3>This WebUI uses Qwen2-VL for OCR and Qwen2-Math for mathematical reasoning.</font></center>
+    """)
     state = gr.State({"tab_index": 0})
     with gr.Row():
         with gr.Column():
@@ -97,18 +93,18 @@ with gr.Blocks(css=css) as demo:
                 with gr.Tab("Sketch"):
                     input_sketchpad = gr.Sketchpad(type="pil", label="Sketch", layers=False)
             input_tabs.select(fn=tabs_select, inputs=[state])
-            input_text = gr.Textbox(label="input your question")
+            input_text = gr.Textbox(label="Enter your question")
             with gr.Row():
                 with gr.Column():
                     clear_btn = gr.ClearButton([input_image, input_sketchpad, input_text])
                 with gr.Column():
                     submit_btn = gr.Button("Submit", variant="primary")
         with gr.Column():
-            output_md = gr.Markdown(label="answer", elem_id="qwen-md")
+            output_md = gr.Markdown(label="Answer", elem_id="qwen-md")
         submit_btn.click(
             fn=math_chat_bot,
             inputs=[input_image, input_sketchpad, input_text, state],
             outputs=output_md)
 
 if __name__ == "__main__":
-    demo.launch()
+    demo.launch()