Update app.py

app.py

@@ -10,13 +10,6 @@ MODEL_PATH = "THUDM/cogvlm2-llama3-caption"
 DEVICE = 'cuda' if torch.cuda.is_available() else 'cpu'
 TORCH_TYPE = torch.bfloat16 if torch.cuda.is_available() and torch.cuda.get_device_capability()[0] >= 8 else torch.float16
 
-# Configurable constants
-NUM_FRAMES = 24  # Default number of frames to extract
-MAX_NEW_TOKENS = 2048
-TOP_K = 1
-TOP_P = 0.1
-DEFAULT_TEMPERATURE = 1.0
-
 # Delay Reasons for Each Manufacturing Step
 DELAY_REASONS = {
     "Step 1": ["Delay in Bead Insertion", "Lack of raw material"],
@@ -98,34 +91,38 @@ def get_step_info(step_number):
             "Standard Time": "7 seconds",
             "Video_substeps_expected": {
                 "0-3 seconds": "Technician unloads(removes) carcass(tire) from the machine."
-            }
+            },
+            "Potential_Delay_reasons": [
+                "Person not available in time(in 3 sec) to remove carcass.",
+                "Person is doing bead(ring) insertion before carcass unload causing unload to be delayed by more than 3 sec"
+            ]
         }
     }
     
     return step_details.get(step_number, {"Error": "Invalid step number. Please provide a valid step number."})
 
+
+
 def load_video(video_data, strategy='chat'):
     """Loads and processes video data into a format suitable for model input."""
     bridge.set_bridge('torch')
-
+    num_frames = 24
+    
     if isinstance(video_data, str): 
         decord_vr = VideoReader(video_data, ctx=cpu(0))
     else:  
         decord_vr = VideoReader(io.BytesIO(video_data), ctx=cpu(0))
     
+    frame_id_list = []
     total_frames = len(decord_vr)
-    if total_frames < NUM_FRAMES:
-        raise ValueError("Uploaded video is too short for meaningful analysis.")
-    
     timestamps = [i[0] for i in decord_vr.get_frame_timestamp(np.arange(total_frames))]
     max_second = round(max(timestamps)) + 1
     
-    frame_id_list = []
     for second in range(max_second):
         closest_num = min(timestamps, key=lambda x: abs(x - second))
         index = timestamps.index(closest_num)
         frame_id_list.append(index)
-        if len(frame_id_list) >= NUM_FRAMES:
+        if len(frame_id_list) >= num_frames:
             break
 
     video_data = decord_vr.get_batch(frame_id_list)
@@ -154,10 +151,7 @@ def load_model():
 
 def predict(prompt, video_data, temperature, model, tokenizer):
     """Generates predictions based on the video and textual prompt."""
-    try:
-        video = load_video(video_data, strategy='chat')
-    except ValueError as e:
-        return f"Error loading video: {str(e)}"
+    video = load_video(video_data, strategy='chat')
     
     inputs = model.build_conversation_input_ids(
         tokenizer=tokenizer,
@@ -175,20 +169,20 @@ def predict(prompt, video_data, temperature, model, tokenizer):
     }
     
     gen_kwargs = {
-        "max_new_tokens": MAX_NEW_TOKENS,
-        "pad_token_id": tokenizer.pad_token_id,
-        "top_k": TOP_K,
+        "max_new_tokens": 2048,
+        "pad_token_id": 128002,
+        "top_k": 1,
         "do_sample": False,
-        "top_p": TOP_P,
-        "temperature": temperature or DEFAULT_TEMPERATURE,
+        "top_p": 0.1,
+        "temperature": temperature,
     }
     
     with torch.no_grad():
         outputs = model.generate(**inputs, **gen_kwargs)
         outputs = outputs[:, inputs['input_ids'].shape[1]:]
-        response = tokenizer.decode(outputs[0], skip_special_tokens=True).strip()
+        response = tokenizer.decode(outputs[0], skip_special_tokens=True)
     
-    return f"Analysis Result:\n{response}"
+    return response
 
 def get_analysis_prompt(step_number):
     """Constructs the prompt for analyzing delay reasons based on the selected step."""
@@ -199,23 +193,100 @@ def get_analysis_prompt(step_number):
     
     step_name = step_info["Name"]
     standard_time = step_info["Standard Time"]
-    substeps = step_info["Video_substeps_expected"]
-    delay_reasons = DELAY_REASONS.get(f"Step {step_number}", ["No specific reasons provided."])
+    analysis = step_info["Analysis"]
 
-    substeps_text = "\n".join([f"- {time}: {action}" for time, action in substeps.items()])
-    reasons_text = "\n".join([f"- {reason}" for reason in delay_reasons])
-    
     return f"""
-You are an AI expert system analyzing manufacturing delays in tire production. Below are the details:
-Step: {step_number} - {step_name}
-Standard Time: {standard_time}
-Substeps Expected in Video:
-{substeps_text}
-
-Potential Delay Reasons:
-{reasons_text}
-
-Task: Analyze the provided video to identify the delay reason. Use the following format:
-1. **Selected Reason:** [Choose the most likely reason from the list above]
-2. **Visual Evidence:** [Describe specific visual cues from the video that support your analysis.]
+You are an AI expert system specialized in analyzing manufacturing processes and identifying production delays in tire manufacturing. Your role is to accurately classify delay reasons based on visual evidence from production line footage.
+Task Context:
+You are analyzing video footage from Step {step_number} of a tire manufacturing process where a delay has been detected. The step is called {step_name}, and its standard time is {standard_time}.
+Required Analysis:
+Carefully observe the video for visual cues indicating production interruption.
+- If no person is visible in any of the frames, the reason probably might be due to their absence.
+- If a person is visible in the video and is observed touching and modifying the layers of the tire, it indicates an issue with tire patching, and the person might be repairing it.
+- Compare observed evidence against the following possible delay reasons: 
+  - {analysis}
+Following are the subactivities needs to happen in this step.
+
+{get_step_info(step_number)}
+
+
+Please provide your output in the following format:
+Output_Examples = {
+     ["Delay in Bead Insertion", "Lack of raw material"],
+     ["Inner Liner Adjustment by Technician", "Person rebuilding defective Tire Sections"],
+     ["Manual Adjustment in Ply1 Apply", "Technician repairing defective Tire Sections"],
+     ["Delay in Bead Set", "Lack of raw material"],
+     ["Delay in Turnup", "Lack of raw material"],
+     ["Person Repairing Sidewall", "Person rebuilding defective Tire Sections"],
+     ["Delay in Sidewall Stitching", "Lack of raw material"],
+     ["No person available to load Carcass", "No person available to collect tire"]
+}
+1. **Selected Reason:** [State the most likely reason from the given options]
+2. **Visual Evidence:** [Describe specific visual cues that support your selection]
+3. **Reasoning:** [Explain why this reason best matches the observed evidence]
+4. **Alternative Analysis:** [Brief explanation of why other possible reasons are less likely]
+Important: Base your analysis solely on visual evidence from the video. Focus on concrete, observable details rather than assumptions. Clearly state if no person or specific activity is observed.
 """
+
+
+
+model, tokenizer = load_model()
+
+def inference(video, step_number):
+    """Analyzes video to predict possible issues based on the manufacturing step."""
+    try:
+        if not video:
+            return "Please upload a video first."
+        
+        prompt = get_analysis_prompt(step_number)
+        temperature = 0.3
+        response = predict(prompt, video, temperature, model, tokenizer)
+        
+        return response
+    except Exception as e:
+        return f"An error occurred during analysis: {str(e)}"
+
+def create_interface():
+    """Creates the Gradio interface for the Manufacturing Analysis System."""
+    with gr.Blocks() as demo:
+        gr.Markdown("""
+        # Manufacturing Analysis System
+        Upload a video of the manufacturing step and select the step number. 
+        The system will analyze the video and provide observations.
+        """)
+        
+        with gr.Row():
+            with gr.Column():
+                video = gr.Video(label="Upload Manufacturing Video", sources=["upload"])
+                step_number = gr.Dropdown(
+                    choices=[f"Step {i}" for i in range(1, 9)],
+                    label="Manufacturing Step"
+                )
+                analyze_btn = gr.Button("Analyze", variant="primary")
+            
+            with gr.Column():
+                output = gr.Textbox(label="Analysis Result", lines=10)
+        
+        gr.Examples(
+            examples=[
+                ["7838_step2_2_eval.mp4", "Step 2"],
+                ["7838_step6_2_eval.mp4", "Step 6"],
+                ["7838_step8_1_eval.mp4", "Step 8"],
+                ["7993_step6_3_eval.mp4", "Step 6"],
+                ["7993_step8_3_eval.mp4", "Step 8"]    
+            ],
+            inputs=[video, step_number],
+            cache_examples=False
+        )
+        
+        analyze_btn.click(
+            fn=inference,
+            inputs=[video, step_number],
+            outputs=[output]
+        )
+    
+    return demo
+
+if __name__ == "__main__":
+    demo = create_interface()
+    demo.queue().launch(share=True)