phi4_detector.py

"""
Phi-4 model implementation for video anomaly detection using Hugging Face transformers.
"""

import cv2
import numpy as np
import base64
import os
import tempfile
from PIL import Image
import io
import re
import torch
import time
from transformers import AutoModelForCausalLM, AutoTokenizer, AutoProcessor
from huggingface_hub import snapshot_download

class Phi4AnomalyDetector:
    def __init__(self, model_name="microsoft/Phi-4-multimodal-instruct"):
        """
        Initialize the Phi4AnomalyDetector with the Phi-4 vision model.
        
        Args:
            model_name (str): Name of the Phi-4 vision model on Hugging Face
        """
        self.model_name = model_name
        self.model_dir = os.path.join(os.getcwd(), "phi4_model")
        self.device = "cuda" if torch.cuda.is_available() else "cpu"
        
        # Load or download the model
        self.load_model()
        
    def load_model(self):
        """
        Load the Phi-4 model from local directory or download from Hugging Face.
        """
        try:
            if not os.path.exists(self.model_dir):
                print(f"Downloading {self.model_name} model to {self.model_dir}...")
                snapshot_download(repo_id=self.model_name, local_dir=self.model_dir)
                print("Model downloaded successfully.")
            else:
                print(f"Using existing model from {self.model_dir}")
            
            # Load model components with trust_remote_code=True
            self.processor = AutoProcessor.from_pretrained(
                self.model_dir,
                trust_remote_code=True
            )
            
            self.tokenizer = AutoTokenizer.from_pretrained(
                self.model_dir,
                trust_remote_code=True
            )
            
            # Load model with appropriate dtype based on device
            if self.device == "cuda":
                self.model = AutoModelForCausalLM.from_pretrained(
                    self.model_dir,
                    torch_dtype=torch.float16,
                    device_map="auto",
                    trust_remote_code=True
                )
            else:
                self.model = AutoModelForCausalLM.from_pretrained(
                    self.model_dir,
                    device_map="auto",
                    trust_remote_code=True
                )
            
            print(f"Phi-4 model loaded successfully on {self.device}")
            
        except Exception as e:
            raise RuntimeError(f"Failed to load Phi-4 model: {str(e)}")
    
    def extract_frames(self, video_path, skip_frames):
        """
        Extract frames from a video file, skipping the specified number of frames.
        
        Args:
            video_path (str): Path to the video file
            skip_frames (int): Number of frames to skip between captures
            
        Returns:
            list: List of extracted frames as numpy arrays
        """
        frames = []
        # Use the default backend for video files
        # DirectShow can cause issues with some video files
        cap = cv2.VideoCapture(video_path)
        
        # Don't set MJPG format for video files as it can interfere with proper decoding
        # cap.set(cv2.CAP_PROP_FOURCC, cv2.VideoWriter_fourcc('M','J','P','G'))
        
        if not cap.isOpened():
            raise ValueError(f"Could not open video file: {video_path}")
        
        frame_count = 0
        while True:
            ret, frame = cap.read()
            if not ret:
                break
                
            if frame_count % (skip_frames + 1) == 0:
                # Convert from BGR to RGB
                rgb_frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
                frames.append(rgb_frame)
                
            frame_count += 1
            
        cap.release()
        return frames
    
    def process_live_stream(self, stream_source, skip_frames, prompt, analysis_depth="granular", max_frames=None, callback=None, time_interval=None):
        """
        Process frames from a live video stream.
        
        Args:
            stream_source: Stream source (0 for webcam, URL for IP camera or RTSP stream)
            skip_frames (int): Number of frames to skip between captures
            prompt (str): Prompt describing what anomaly to look for
            analysis_depth (str): "granular" for frame-by-frame analysis or "cumulative" for overall analysis
            max_frames (int, optional): Maximum number of frames to process (None for unlimited)
            callback (function, optional): Callback function to report progress
            time_interval (int, optional): If set, capture one frame every X seconds instead of using skip_frames
            
        Returns:
            list or dict: List of analysis results for each processed frame (granular) or dict with cumulative analysis (cumulative)
        """
        # Open the video stream with appropriate backend
        # Only use DirectShow for local webcams (0 or 1) on Windows
        if os.name == 'nt' and (stream_source == 0 or stream_source == 1):
            # This is a local webcam on Windows, use DirectShow
            cap = cv2.VideoCapture(stream_source, cv2.CAP_DSHOW)
            
            # For webcams, MJPG format can be more stable
            cap.set(cv2.CAP_PROP_FOURCC, cv2.VideoWriter_fourcc('M','J','P','G'))
        else:
            # For IP cameras, RTSP streams, or non-Windows systems, use default backend
            cap = cv2.VideoCapture(stream_source)
        
        if not cap.isOpened():
            raise ValueError(f"Could not open video stream: {stream_source}")
        
        frames = []
        frame_count = 0
        processed_count = 0
        last_capture_time = time.time()
        
        try:
            while True:
                ret, frame = cap.read()
                if not ret:
                    break
                
                current_time = time.time()
                
                # Determine if we should capture this frame
                should_capture = False
                
                if time_interval is not None:
                    # Time-based interval mode
                    if current_time - last_capture_time >= time_interval:
                        should_capture = True
                        last_capture_time = current_time
                else:
                    # Frame-skip mode
                    if frame_count % (skip_frames + 1) == 0:
                        should_capture = True
                
                if should_capture:
                    # Convert from BGR to RGB
                    rgb_frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
                    frames.append(rgb_frame)
                    processed_count += 1
                    
                    # Process the frame immediately in time interval mode
                    if time_interval is not None:
                        # Process the frame immediately
                        result = self.analyze_frame(rgb_frame, prompt)
                        
                        # Make sure the frame is included in the result
                        if "frame" not in result:
                            result["frame"] = rgb_frame
                        
                        # If we have a callback, call it with -1 for total to indicate continuous mode
                        if callback:
                            callback(processed_count, -1)
                        
                        # In time interval mode, we yield results one by one
                        if analysis_depth == "granular":
                            yield result
                        else:
                            # For cumulative analysis, we need to keep all frames
                            # We'll handle this outside the loop
                            pass
                    else:
                        # Update progress if callback is provided (frame count mode)
                        if callback and max_frames:
                            callback(processed_count, max_frames)
                        
                        # Break if we've reached the maximum number of frames
                        if max_frames and processed_count >= max_frames:
                            break
                
                frame_count += 1
        finally:
            cap.release()
        
        # If we're in time interval mode with cumulative analysis, we don't return here
        # as we're yielding results above
        if time_interval is not None and analysis_depth == "cumulative":
            # This is a special case - we need to periodically do cumulative analysis
            # For simplicity, we'll just return the current cumulative analysis
            result = self.analyze_frames_cumulatively(frames, prompt, callback)
            yield result
            return
            
        # Process the collected frames for non-time interval mode
        if time_interval is None:
            if analysis_depth == "cumulative":
                return self.analyze_frames_cumulatively(frames, prompt, callback)
            else:  # granular (default)
                results = []
                
                for i, frame in enumerate(frames):
                    if callback:
                        callback(i, len(frames))
                        
                    result = self.analyze_frame(frame, prompt)
                    results.append(result)
                    
                return results
    
    def analyze_frame(self, frame, prompt):
        """
        Analyze a frame using the Phi-4 vision model.
        
        Args:
            frame (numpy.ndarray): Frame to analyze
            prompt (str): Prompt describing what anomaly to look for
            
        Returns:
            dict: Analysis result from the model
        """
        # Convert numpy array to PIL Image
        pil_image = Image.fromarray(frame)
        
        # Enhanced prompt to get structured information about anomalies
        enhanced_prompt = f"""
{prompt}

After your analysis, please include a structured assessment at the end of your response in this exact format:
ANOMALY_DETECTED: [Yes/No]
ANOMALY_TYPE: [Human/Non-human/None]
CONFIDENCE: [0-100]

For ANOMALY_DETECTED, answer "Yes" if you detect any anomaly, otherwise "No".
For ANOMALY_TYPE, if an anomaly is detected, classify it as either "Human" (if it involves people or human activities) or "Non-human" (if it involves objects, animals, or environmental factors). If no anomaly is detected, use "None".
For CONFIDENCE, provide a number from 0 to 100 indicating your confidence level in the assessment.
"""
        
        try:
            # Process the image and prompt with the Phi-4 model
            inputs = self.processor(text=enhanced_prompt, images=pil_image, return_tensors="pt")
            
            # Move inputs to the same device as the model
            for key in inputs:
                if torch.is_tensor(inputs[key]):
                    inputs[key] = inputs[key].to(self.model.device)
            
            # Generate response
            with torch.no_grad():
                outputs = self.model.generate(
                    **inputs,
                    max_new_tokens=500,
                    do_sample=False
                )
            
            # Decode the response
            response_text = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
            
            # Extract the part after the prompt
            if enhanced_prompt in response_text:
                response_text = response_text.split(enhanced_prompt)[-1].strip()
            
            # Extract anomaly detection information using regex
            anomaly_detected = False
            anomaly_type = "None"
            confidence = 0
            
            # Look for the structured format
            anomaly_match = re.search(r'ANOMALY_DETECTED:\s*(Yes|No)', response_text, re.IGNORECASE)
            if anomaly_match and anomaly_match.group(1).lower() == 'yes':
                anomaly_detected = True
                confidence = 90  # Default high confidence when anomaly is detected
                
                # If anomaly detected, look for the type
                type_match = re.search(r'ANOMALY_TYPE:\s*(Human|Non-human|None)', response_text, re.IGNORECASE)
                if type_match:
                    anomaly_type = type_match.group(1)
                    
                # Look for confidence information
                conf_match = re.search(r'CONFIDENCE:\s*(\d+)', response_text, re.IGNORECASE)
                if conf_match:
                    try:
                        confidence = int(conf_match.group(1))
                    except:
                        pass  # Keep default confidence if parsing fails
            
            return {
                "text": response_text,
                "analysis": response_text,  # Add analysis field as an alias for text
                "frame": frame,
                "anomaly_detected": anomaly_detected,
                "anomaly_type": anomaly_type,
                "confidence": confidence,  # Add confidence field
                "timestamp": time.time()  # Add timestamp for live stream analysis
            }
        except Exception as e:
            return {
                "error": str(e),
                "frame": frame,
                "anomaly_detected": False,
                "anomaly_type": "None",
                "confidence": 0,  # Add default confidence for error
                "timestamp": time.time()  # Add timestamp for live stream analysis
            }
    
    def analyze_frames_cumulatively(self, frames, prompt, callback=None):
        """
        Analyze all frames together and provide a cumulative analysis.
        
        Args:
            frames (list): List of frames to analyze
            prompt (str): Prompt describing what anomaly to look for
            callback (function, optional): Callback function to report progress
            
        Returns:
            dict: Cumulative analysis result
        """
        # First, analyze each frame individually to identify potential anomalies
        individual_results = []
        for i, frame in enumerate(frames):
            if callback:
                callback(i, len(frames) * 2)  # First half of progress for individual analysis
                
            result = self.analyze_frame(frame, f"{prompt} Provide a brief analysis of this frame only.")
            individual_results.append(result)
        
        # Identify frames with potential anomalies
        anomaly_frames = []
        anomaly_descriptions = []
        anomaly_types = []
        
        for i, result in enumerate(individual_results):
            if "error" not in result and result["anomaly_detected"]:
                anomaly_frames.append(result["frame"])
                anomaly_descriptions.append(f"Frame {i+1}: {result['text']}")
                anomaly_types.append(result["anomaly_type"])
                
                # Limit to 3 anomaly frames
                if len(anomaly_frames) >= 3:
                    break
        
        # If no anomalies were detected, use the first, middle, and last frames
        if not anomaly_frames and len(frames) > 0:
            if len(frames) == 1:
                anomaly_frames = [frames[0]]
            elif len(frames) == 2:
                anomaly_frames = [frames[0], frames[1]]
            else:
                anomaly_frames = [
                    frames[0],
                    frames[len(frames) // 2],
                    frames[-1]
                ]
        
        # Limit to max 3 frames
        anomaly_frames = anomaly_frames[:3]
        
        # Create a cumulative analysis prompt with the anomaly descriptions
        cumulative_prompt = f"""
{prompt}

Based on the analysis of all frames, provide a comprehensive summary of any anomalies detected in the video. Focus on patterns or recurring issues. Here are some notable observations from individual frames:

{chr(10).join(anomaly_descriptions[:5])}

After your analysis, please include a structured assessment at the end of your response in this exact format:
ANOMALY_DETECTED: [Yes/No]
ANOMALY_TYPE: [Human/Non-human/None]

For ANOMALY_DETECTED, answer "Yes" if you detect any anomaly across the video, otherwise "No".
For ANOMALY_TYPE, if an anomaly is detected, classify the predominant type as either "Human" (if it involves people or human activities) or "Non-human" (if it involves objects, animals, or environmental factors). If no anomaly is detected, use "None".
"""
        
        try:
            if callback:
                callback(len(frames), len(frames) * 2)  # Second half of progress for cumulative analysis
            
            # For cumulative analysis, we'll use the first anomaly frame (or first frame if no anomalies)
            representative_frame = anomaly_frames[0] if anomaly_frames else frames[0]
            pil_image = Image.fromarray(representative_frame)
            
            # Process with Phi-4
            inputs = self.processor(text=cumulative_prompt, images=pil_image, return_tensors="pt")
            
            # Move inputs to the same device as the model
            for key in inputs:
                if torch.is_tensor(inputs[key]):
                    inputs[key] = inputs[key].to(self.model.device)
            
            # Generate response
            with torch.no_grad():
                outputs = self.model.generate(
                    **inputs,
                    max_new_tokens=1000,
                    do_sample=False
                )
            
            # Decode the response
            response_text = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
            
            # Extract the part after the prompt
            if cumulative_prompt in response_text:
                response_text = response_text.split(cumulative_prompt)[-1].strip()
            
            # Extract anomaly detection information using regex
            anomaly_detected = False
            anomaly_type = "None"
            confidence = 0
            
            # Look for the structured format
            anomaly_match = re.search(r'ANOMALY_DETECTED:\s*(Yes|No)', response_text, re.IGNORECASE)
            if anomaly_match and anomaly_match.group(1).lower() == 'yes':
                anomaly_detected = True
                confidence = 90  # Default high confidence when anomaly is detected
                
                # If anomaly detected, look for the type
                type_match = re.search(r'ANOMALY_TYPE:\s*(Human|Non-human|None)', response_text, re.IGNORECASE)
                if type_match:
                    anomaly_type = type_match.group(1)
                    
                # Look for confidence information
                conf_match = re.search(r'CONFIDENCE:\s*(\d+)', response_text, re.IGNORECASE)
                if conf_match:
                    try:
                        confidence = int(conf_match.group(1))
                    except:
                        pass  # Keep default confidence if parsing fails
            
            return {
                "text": response_text,
                "analysis": response_text,  # Add analysis field as an alias for text
                "frames": anomaly_frames,
                "anomaly_detected": anomaly_detected,
                "anomaly_type": anomaly_type,
                "confidence": confidence,  # Add confidence field
                "timestamp": time.time()  # Add timestamp for live stream analysis
            }
        except Exception as e:
            return {
                "error": str(e),
                "frames": anomaly_frames,
                "anomaly_detected": False,
                "anomaly_type": "None",
                "confidence": 0,  # Add default confidence for error
                "timestamp": time.time()  # Add timestamp for live stream analysis
            }
    
    def process_video(self, video_path, skip_frames, prompt, analysis_depth="granular", callback=None):
        """
        Process a video file, extracting frames and analyzing them for anomalies.
        
        Args:
            video_path (str): Path to the video file
            skip_frames (int): Number of frames to skip between captures
            prompt (str): Prompt describing what anomaly to look for
            analysis_depth (str): "granular" for frame-by-frame analysis or "cumulative" for overall analysis
            callback (function, optional): Callback function to report progress
            
        Returns:
            list or dict: List of analysis results for each processed frame (granular) or dict with cumulative analysis (cumulative)
        """
        frames = self.extract_frames(video_path, skip_frames)
        
        if analysis_depth == "cumulative":
            return self.analyze_frames_cumulatively(frames, prompt, callback)
        else:  # granular (default)
            results = []
            
            for i, frame in enumerate(frames):
                if callback:
                    callback(i, len(frames))
                    
                result = self.analyze_frame(frame, prompt)
                results.append(result)
                
            return results