Update app.py

app.py

@@ -6,10 +6,22 @@ from decord import cpu, VideoReader, bridge
 from transformers import AutoModelForCausalLM, AutoTokenizer
 from transformers import BitsAndBytesConfig
 
-MODEL_PATH = "THUDM/cogvlm2-video-llama3-chat"
+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
 
+# Delay Reasons for Each Manufacturing Step
+DELAY_REASONS = {
+    "Step 1": ["Delay in Bead Insertion", "Lack of raw material"],
+    "Step 2": ["Inner Liner Adjustment by Technician", "Person rebuilding defective Tire Sections"],
+    "Step 3": ["Manual Adjustment in Ply1 apply", "Technician repairing defective Tire Sections"],
+    "Step 4": ["Delay in Bead set", "Lack of raw material"],
+    "Step 5": ["Delay in Turnup", "Lack of raw material"],
+    "Step 6": ["Person Repairing sidewall", "Person rebuilding defective Tire Sections"],
+    "Step 7": ["Delay in sidewall stitching", "Lack of raw material"],
+    "Step 8": ["No person available to load Carcass", "No person available to collect tire"]
+}
+
 def get_step_info(step_number):
     """Returns detailed information about a manufacturing step."""
     step_details = {
@@ -124,6 +136,8 @@ def get_step_info(step_number):
     
     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')
@@ -171,10 +185,9 @@ def load_model():
     return model, tokenizer
 
 def predict(prompt, video_data, temperature, model, tokenizer):
-    """Generates predictions based on the video and textual prompt, focusing on one delay reason."""
+    """Generates predictions based on the video and textual prompt."""
     video = load_video(video_data, strategy='chat')
     
-    # Prepare the inputs for the model
     inputs = model.build_conversation_input_ids(
         tokenizer=tokenizer,
         query=prompt,
@@ -204,10 +217,7 @@ def predict(prompt, video_data, temperature, model, tokenizer):
         outputs = outputs[:, inputs['input_ids'].shape[1]:]
         response = tokenizer.decode(outputs[0], skip_special_tokens=True)
     
-    # Extract the output in the desired format (return just one delay reason)
-    # We can assume the model will return one reason in the format "Output: <delay reason>"
-    return response.strip().split('\n')[-1]  # Assuming the reason is on the last line
-
+    return response
 
 def get_analysis_prompt(step_number):
     """Constructs the prompt for analyzing delay reasons based on the selected step."""
@@ -218,25 +228,42 @@ def get_analysis_prompt(step_number):
     
     step_name = step_info["Name"]
     standard_time = step_info["Standard Time"]
-    potential_delay_reasons = step_info.get("Potential_Delay_Reasons", [])
+    analysis = step_info["Analysis"]
 
-    # Constructing the prompt dynamically with potential delay reasons
-    potential_reasons_text = "\n    ".join(potential_delay_reasons)
-    
     return f"""
-You are an AI expert system specialized in analyzing manufacturing processes and identifying production delays in tire manufacturing. Your role is to accurately classify the correct delay reason based on visual evidence from production line footage.
+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. The possible delay reasons are:
-    {potential_reasons_text}
-Please provide only the most likely delay reason in the following format:
+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)}
 
-Output:
-    <One specific delay reason>
+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):
@@ -296,4 +323,4 @@ def create_interface():
 
 if __name__ == "__main__":
     demo = create_interface()
-    demo.queue().launch(share=True)
+    demo.queue().launch(share=True)