video_processing.py

import os
import cv2
from scenedetect import VideoManager, SceneManager
from scenedetect.detectors import ContentDetector
from moviepy.editor import VideoFileClip
from transformers import CLIPProcessor, CLIPModel
import torch
import yt_dlp
from PIL import Image
import uuid
from torchvision import models, transforms
from torch.nn import functional as F
import numpy as np


categories = ["Joy", "Trust", "Fear", "Surprise", "Sadness", "Disgust", "Anger", "Anticipation"]


device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32").to(device)
processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")

# Load ResNet-50 model
resnet50 = models.resnet50(pretrained=True)
resnet50.eval().to(device)


def classify_frame(frame):
    # Preprocess the image
    preprocess = transforms.Compose([
        transforms.Resize(256),
        transforms.CenterCrop(224),
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    ])
    input_tensor = preprocess(Image.fromarray(frame))
    input_batch = input_tensor.unsqueeze(0).to(device)

    # Predict with ResNet-50
    with torch.no_grad():
        output = resnet50(input_batch)
        probabilities = F.softmax(output[0], dim=0)

    # Create a numpy array from the probabilities of the categories
    # This example assumes each category is mapped to a model output directly
    results_array = np.array([probabilities[i].item() for i in range(len(categories))])

    return results_array


def download_video(url):
    ydl_opts = {
        'format': 'bestvideo[height<=1440]+bestaudio/best[height<=1440]',
        'outtmpl': f'temp_videos/{uuid.uuid4()}_video.%(ext)s',
        'merge_output_format': 'mp4',
    }
    with yt_dlp.YoutubeDL(ydl_opts) as ydl:
        result = ydl.extract_info(url, download=True)
        video_filename = ydl.prepare_filename(result)
        safe_filename = sanitize_filename(video_filename)
        if os.path.exists(video_filename) and video_filename != safe_filename:
            os.rename(video_filename, safe_filename)
        return safe_filename

def sanitize_filename(filename):
    return "".join([c if c.isalnum() or c in " .-_()" else "_" for c in filename])

def find_scenes(video_path):
    video_manager = VideoManager([video_path])
    scene_manager = SceneManager()
    scene_manager.add_detector(ContentDetector(threshold=33))  # Adjusted threshold for finer segmentation
    video_manager.set_downscale_factor()
    video_manager.start()
    scene_manager.detect_scenes(frame_source=video_manager)
    scene_list = scene_manager.get_scene_list()
    video_manager.release()
    scenes = [(start.get_timecode(), end.get_timecode()) for start, end in scene_list]
    return scenes

def convert_timestamp_to_seconds(timestamp):
    h, m, s = map(float, timestamp.split(':'))
    return int(h) * 3600 + int(m) * 60 + s

def extract_frames(video, start_time, end_time):
    frames = []
    start_seconds = convert_timestamp_to_seconds(start_time)
    end_seconds = convert_timestamp_to_seconds(end_time)
    video_clip = video.subclip(start_seconds, end_seconds)

    for frame_time in range(0, int(video_clip.duration * video_clip.fps), int(video_clip.fps / 3)):
        frame = video_clip.get_frame(frame_time / video_clip.fps)
        frames.append(frame)
    return frames

def analyze_scenes(video_path, scenes, description):
    scene_scores = []
    negative_descriptions = [
        "black screen",
        "Intro text for a video",
        "dark scene without much contrast",
        "No people are in this scene",
        #"A still shot of natural scenery",
        #"Still-camera shot of a person's face"
    ]

    preprocess = transforms.Compose([
        transforms.ToTensor(),  # Directly convert numpy arrays to tensors
        transforms.Resize((224, 224)),  # Resize the tensor
        transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])  # Normalize the tensor
    ])

    text_inputs = processor(text=[description] + negative_descriptions, return_tensors="pt", padding=True).to(device)
    text_features = model.get_text_features(**text_inputs).detach()
    positive_feature, negative_features = text_features[0], text_features[1:]

    video = VideoFileClip(video_path)

    for scene_num, (start_time, end_time) in enumerate(scenes):
        frames = extract_frames(video, start_time, end_time)
        if not frames:
            print(f"Scene {scene_num + 1}: Start={start_time}, End={end_time} - No frames extracted")
            continue

        scene_prob = 0.0
        sentiment_distributions = np.zeros(8)  # Assuming there are 8 sentiments
        for frame in frames:
            # Directly preprocess the frame
            frame_tensor = preprocess(frame).unsqueeze(0).to(device)  # Add batch dimension and send to device
            with torch.no_grad():
                image_features = model.get_image_features(pixel_values=frame_tensor).detach()
                positive_similarity = torch.cosine_similarity(image_features, positive_feature.unsqueeze(0)).squeeze().item()
                negative_similarities = torch.cosine_similarity(image_features, negative_features).squeeze().mean().item()
                scene_prob += positive_similarity - negative_similarities
            
            frame_sentiments = classify_frame(frame)
            sentiment_distributions += np.array(frame_sentiments)

        sentiment_distributions /= len(frames)  # Normalize to get average probabilities
        sentiment_percentages = {category: round(prob * 100, 2) for category, prob in zip(categories, sentiment_distributions)}
        scene_prob /= len(frames)
        scene_duration = convert_timestamp_to_seconds(end_time) - convert_timestamp_to_seconds(start_time)
        print(f"Scene {scene_num + 1}: Start={start_time}, End={end_time}, Probability={scene_prob}, Duration={scene_duration}, Sentiments: {sentiment_percentages}")

        scene_scores.append((scene_prob, start_time, end_time, scene_duration, sentiment_percentages))

        # Sort scenes by confidence, highest first
        scene_scores.sort(reverse=True, key=lambda x: x[0])
        
        # Select the longest scene from the top 3 highest confidence scenes
        top_3_scenes = scene_scores[:3]  # Get the top 3 scenes
        best_scene = max(top_3_scenes, key=lambda x: x[3])  # Find the longest scene from these top 3

    if best_scene:
        print(f"Best Scene: Start={best_scene[1]}, End={best_scene[2]}, Probability={best_scene[0]}, Duration={best_scene[3]}, Sentiments: {best_scene[4]}")
        return (best_scene[1], best_scene[2]), best_scene[4]  # Returning a tuple with scene times and sentiments
    else:
        print("No suitable scene found")
        return None, {}


def extract_best_scene(video_path, scene):
    if scene is None:
        return None

    start_time, end_time = scene
    start_seconds = convert_timestamp_to_seconds(start_time)
    end_seconds = convert_timestamp_to_seconds(end_time)
    video_clip = VideoFileClip(video_path).subclip(start_seconds, end_seconds)
    return video_clip

def cleanup_temp_files():
    temp_dir = 'temp_videos'
    if os.path.exists(temp_dir):
        for file in os.listdir(temp_dir):
            file_path = os.path.join(temp_dir, file)
            try:
                if os.path.isfile(file_path):
                    os.unlink(file_path)
            except Exception as e:
                print(f"Error: {e}")