Update app.py

app.py

@@ -1,17 +1,171 @@
 import gradio as gr
+import io
+import numpy as np
 import torch
-from PIL import Image
+from decord import cpu, VideoReader, bridge
 from transformers import AutoModelForCausalLM, AutoTokenizer
 from transformers import BitsAndBytesConfig
-import torchvision.transforms as transforms
 
-# Model configuration
 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
 
-# Load Model and Tokenizer
+# 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 = {
+        1: {
+            "Name": "Bead Insertion",
+            "Standard Time": "4 seconds",
+            "Video_substeps_expected": {
+                "0-1 second": "Machine starts bead insertion process.",
+                "1-3 seconds": "Beads are aligned and positioned.",
+                "3-4 seconds": "Final adjustment and confirmation of bead placement."
+            },
+            "Potential_Delay_Reasons": [
+                "Delay in bead insertion",
+                "Lack of raw material",
+                "Machine malfunction during bead alignment"
+            ]
+        },
+        2: {
+            "Name": "Inner Liner Apply",
+            "Standard Time": "4 seconds",
+            "Video_substeps_expected": {
+                "0-1 second": "Machine applies the first layer of the liner.",
+                "1-3 seconds": "Technician checks alignment and adjusts if needed.",
+                "3-4 seconds": "Final inspection and confirmation."
+            },
+            "Potential_Delay_Reasons": [
+                "Technician adjusting inner liner alignment",
+                "Person rebuilding defective tire sections",
+                "Machine alignment issues"
+            ]
+        },
+        3: {
+            "Name": "Ply1 Apply",
+            "Standard Time": "4 seconds",
+            "Video_substeps_expected": {
+                "0-2 seconds": "First ply is loaded onto the machine.",
+                "2-4 seconds": "Technician inspects and adjusts ply placement."
+            },
+            "Potential_Delay_Reasons": [
+                "Manual adjustment of ply placement",
+                "Technician repairing defective ply sections",
+                "Ply loading issues"
+            ]
+        },
+        4: {
+            "Name": "Bead Set",
+            "Standard Time": "8 seconds",
+            "Video_substeps_expected": {
+                "0-3 seconds": "Bead is positioned and pre-set.",
+                "3-6 seconds": "Machine secures the bead in place.",
+                "6-8 seconds": "Technician confirms the bead alignment."
+            },
+            "Potential_Delay_Reasons": [
+                "Delay in bead positioning",
+                "Lack of raw material",
+                "Machine securing process failure"
+            ]
+        },
+        5: {
+            "Name": "Turnup",
+            "Standard Time": "4 seconds",
+            "Video_substeps_expected": {
+                "0-2 seconds": "Turnup process begins with machine handling.",
+                "2-4 seconds": "Technician inspects the turnup and makes adjustments if necessary."
+            },
+            "Potential_Delay_Reasons": [
+                "Delay in turnup handling",
+                "Lack of raw material",
+                "Technician adjustment delays"
+            ]
+        },
+        6: {
+            "Name": "Sidewall Apply",
+            "Standard Time": "14 seconds",
+            "Video_substeps_expected": {
+                "0-5 seconds": "Sidewall material is positioned by the machine.",
+                "5-10 seconds": "Technician checks for alignment and begins application.",
+                "10-14 seconds": "Final adjustments and confirmation of sidewall placement."
+            },
+            "Potential_Delay_Reasons": [
+                "Person repairing sidewall",
+                "Person rebuilding defective tire sections",
+                "Sidewall positioning issues"
+            ]
+        },
+        7: {
+            "Name": "Sidewall Stitching",
+            "Standard Time": "5 seconds",
+            "Video_substeps_expected": {
+                "0-2 seconds": "Stitching process begins automatically.",
+                "2-4 seconds": "Technician inspects stitching for any irregularities.",
+                "4-5 seconds": "Machine completes stitching process."
+            },
+            "Potential_Delay_Reasons": [
+                "Delay in stitching process",
+                "Technician repairing stitching irregularities",
+                "Machine stitching malfunction"
+            ]
+        },
+        8: {
+            "Name": "Carcass Unload",
+            "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)
+    timestamps = [i[0] for i in decord_vr.get_frame_timestamp(np.arange(total_frames))]
+    max_second = round(max(timestamps)) + 1
+    
+    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:
+            break
+
+    video_data = decord_vr.get_batch(frame_id_list)
+    video_data = video_data.permute(3, 0, 1, 2)
+    return video_data
+
 def load_model():
+    """Loads the pre-trained model and tokenizer with quantization configurations."""
     quantization_config = BitsAndBytesConfig(
         load_in_4bit=True,
         bnb_4bit_compute_dtype=TORCH_TYPE,
@@ -27,50 +181,17 @@ def load_model():
         quantization_config=quantization_config,
         device_map="auto"
     ).eval()
+    
     return model, tokenizer
 
-model, tokenizer = load_model()
-
-# 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 load_image(image_data):
-    """Preprocess the input image for model compatibility."""
-    transform = transforms.Compose([
-        transforms.Resize((224, 224)),
-        transforms.ToTensor(),
-        transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
-    ])
-    image = Image.open(image_data).convert("RGB")
-    image_tensor = transform(image).unsqueeze(0).to(DEVICE, dtype=TORCH_TYPE)
-    return image_tensor
-
-def get_analysis_prompt(step_number):
-    """Generates the analysis prompt for the given step."""
-    delay_reasons = DELAY_REASONS.get(step_number, [])
-    prompt = f"""
-    You are an AI expert analyzing tire manufacturing steps. 
-    This is Step {step_number}. Identify the most likely cause of delay based on visual evidence.
-    Possible Delay Reasons: {', '.join(delay_reasons)}
-    Provide the reason and supporting evidence.
-    """
-    return prompt
-
-def predict(prompt, image_tensor, temperature=0.3):
-    """Generates predictions based on the image and textual prompt."""
+def predict(prompt, video_data, temperature, model, tokenizer):
+    """Generates predictions based on the video and textual prompt."""
+    video = load_video(video_data, strategy='chat')
+    
     inputs = model.build_conversation_input_ids(
         tokenizer=tokenizer,
         query=prompt,
-        images=[image_tensor],
+        images=[video],
         history=[],
         template_version='chat'
     )
@@ -79,11 +200,11 @@ def predict(prompt, image_tensor, temperature=0.3):
         'input_ids': inputs['input_ids'].unsqueeze(0).to(DEVICE),
         'token_type_ids': inputs['token_type_ids'].unsqueeze(0).to(DEVICE),
         'attention_mask': inputs['attention_mask'].unsqueeze(0).to(DEVICE),
-        'images': [[inputs['images'][0]]],
+        'images': [[inputs['images'][0].to(DEVICE).to(TORCH_TYPE)]],
     }
     
     gen_kwargs = {
-        "max_new_tokens": 1024,
+        "max_new_tokens": 2048,
         "pad_token_id": 128002,
         "top_k": 1,
         "do_sample": False,
@@ -93,33 +214,84 @@ def predict(prompt, image_tensor, temperature=0.3):
     
     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)
+    
     return response
 
-def inference(image, step_number):
-    """Handles the inference process."""
+def get_analysis_prompt(step_number):
+    """Constructs the prompt for analyzing delay reasons based on the selected step."""
+    step_info = get_step_info(step_number)
+    
+    if "Error" in step_info:
+        return step_info["Error"]
+    
+    step_name = step_info["Name"]
+    standard_time = step_info["Standard Time"]
+    analysis = step_info["Analysis"]
+
+    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 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 image:
-            return "Please upload an image."
+        if not video:
+            return "Please upload a video first."
         
-        image_tensor = load_image(image)
         prompt = get_analysis_prompt(step_number)
-        response = predict(prompt, image_tensor)
+        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)}"
 
-# Gradio Interface
 def create_interface():
+    """Creates the Gradio interface for the Manufacturing Analysis System."""
     with gr.Blocks() as demo:
         gr.Markdown("""
-        # Manufacturing Step Analysis System (Image Input)
-        Upload an image and select the manufacturing step to analyze potential delays.
+        # 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():
-                image_input = gr.Image(label="Upload Manufacturing Image", type="file")
+                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"
@@ -129,13 +301,26 @@ def create_interface():
             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=[image_input, step_number],
+            inputs=[video, step_number],
             outputs=[output]
         )
+    
     return demo
 
 if __name__ == "__main__":
     demo = create_interface()
-    demo.queue().launch(share=True)
+    demo.queue().launch(share=True)