app.py

import gradio as gr
import io
import numpy as np
import torch
from decord import cpu, VideoReader, bridge
from transformers import AutoModelForCausalLM, AutoTokenizer
from transformers import BitsAndBytesConfig


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 = {
    "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_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,
        bnb_4bit_use_double_quant=True,
        bnb_4bit_quant_type="nf4"
    )
    
    tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH, trust_remote_code=True)
    model = AutoModelForCausalLM.from_pretrained(
        MODEL_PATH,
        torch_dtype=TORCH_TYPE,
        trust_remote_code=True,
        quantization_config=quantization_config,
        device_map="auto"
    ).eval()
    
    return model, tokenizer

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=[video],
        history=[],
        template_version='chat'
    )
    
    inputs = {
        '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].to(DEVICE).to(TORCH_TYPE)]],
    }
    
    gen_kwargs = {
        "max_new_tokens": 2048,
        "pad_token_id": 128002,
        "top_k": 1,
        "do_sample": False,
        "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)
    
    return response

def get_analysis_prompt(step_number, possible_reasons):
    """
    Constructs the prompt for analyzing delay reasons based on the selected step.
    Args:
        step_number (int): The manufacturing step number being analyzed.
        possible_reasons (list): A list of possible delay reasons for this step.
    Returns:
        str: A robust analysis prompt tailored to the given step and reasons.
    """
    return f"""
    You are an advanced AI expert system specialized in analyzing manufacturing processes to diagnose production delays. Your task is to analyze video footage from Step {step_number} of a tire manufacturing process, where a delay has been identified. Based on visual evidence in the footage, determine the most accurate reason for the delay.

    ### Required Analysis:
    1. Carefully observe the video for any signs of production interruptions or anomalies.
    2. Compare observed evidence against the following potential delay reasons:
        - If no technician or worker is visible in any frame, consider absence as a probable cause.
        - If a technician is visible and interacting with materials, determine if:
          - They are repairing an inner liner or sidewall, suggesting material damage or alignment issues.
          - They are manually adjusting or repatching tire layers, indicating a procedural or equipment issue.
        - Check for machine pauses, material misalignment, or missing components that could indicate equipment or process failures.
        - Observe whether a technician is collecting or loading the carcass, as delays in this activity may signal inefficiencies or staffing issues.

    ### Output Requirements:
    Provide your analysis in the following structured format:
    1. **Selected Reason**: [State the most likely reason from the given options]
    2. **Visual Evidence**: [Describe key observations from the video that support the selected reason]
    3. **Reasoning**: [Explain why this reason aligns best with the observed evidence]
    4. **Alternative Analysis**: [Briefly outline why other possible reasons are less likely]

    ### Important:
    - Base your conclusions solely on observable evidence from the video.
    - Focus on specific visual details rather than assumptions.
    - Clearly state if no conclusive evidence is found and recommend further investigation.
    """




# Load model globally
model, tokenizer = load_model()

def inference(video, step_number):
    """Analyzes video to predict the most likely cause of delay in the selected manufacturing step."""
    try:
        if not video:
            return "Please upload a video first."
        
        possible_reasons = DELAY_REASONS[step_number]
        prompt = get_analysis_prompt(step_number, possible_reasons)
        temperature = 0.8
        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 Delay Analysis System with examples."""
    with gr.Blocks() as demo:
        gr.Markdown("""
        # Manufacturing Delay Analysis System
        Upload a video of the manufacturing step and select the step number. 
        The system will analyze the video and determine the most likely cause of delay.
        """)
        
        with gr.Row():
            with gr.Column():
                video = gr.Video(label="Upload Manufacturing Video", sources=["upload"])
                step_number = gr.Dropdown(
                    choices=list(DELAY_REASONS.keys()),
                    label="Manufacturing Step"
                )
                analyze_btn = gr.Button("Analyze Delay", variant="primary")
            
            with gr.Column():
                output = gr.Textbox(label="Analysis Result", lines=10)
        
        # Add 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"]
            
        ]
        
        gr.Examples(
            examples=examples,
            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)