import torch import clip import cv2, youtube_dl from PIL import Image,ImageDraw, ImageFont import os from functools import partial from multiprocessing.pool import Pool import shutil from pathlib import Path import numpy as np import datetime import gradio as gr # load model and preprocess device = "cuda" if torch.cuda.is_available() else "cpu" model, preprocess = clip.load("ViT-B/32") def select_video_format(url, format_note='480p', ext='mp4'): defaults = ['480p', '360p','240p','144p'] ydl_opts = {} ydl = youtube_dl.YoutubeDL(ydl_opts) info_dict = ydl.extract_info(url, download=False) formats = info_dict.get('formats', None) available_format_notes = set([f['format_note'] for f in formats]) if format_note not in available_format_notes: format_note = [d for d in defaults if d in available_format_notes][0] formats = [f for f in formats if f['format_note'] == format_note and f['ext'] == ext] format = formats[0] format_id = format.get('format_id', None) fps = format.get('fps', None) print(f'format selected: {format}') return(format_id, fps) def download_video(url,format_id): ydl_opts = { 'format':format_id, 'outtmpl': "%(id)s.%(ext)s"} meta = youtube_dl.YoutubeDL(ydl_opts).extract_info(url) save_location = meta['id'] + '.' + meta['ext'] return(save_location) def read_frames(dest_path): original_images = [] images = [] for filename in sorted(dest_path.glob('*.jpg'),key=lambda p: int(p.stem)): image = Image.open(filename).convert("RGB") original_images.append(image) images.append(preprocess(image)) return original_images, images def process_video_parallel(video, skip_frames, dest_path, num_processes, process_number): cap = cv2.VideoCapture(video) chunks_per_process = int(cap.get(cv2.CAP_PROP_FRAME_COUNT)) // (num_processes * skip_frames) count = skip_frames * chunks_per_process * process_number print(f"worker: {process_number}, process frames {count} ~ {skip_frames * chunks_per_process * (process_number + 1)}") while count < skip_frames * chunks_per_process * (process_number + 1) : cap.set(cv2.CAP_PROP_POS_FRAMES, count) ret, frame = cap.read() if not ret: break filename =f"{dest_path}/{count}.jpg" cv2.imwrite(filename, frame) count += skip_frames # Skip 300 frames i.e. 10 seconds for 30 fps cap.release() def vid2frames(url, sampling_interval=1, ext='mp4'): # create folder for extracted frames - if folder exists, delete and create a new one dest_path = Path('frames') try: dest_path.mkdir(parents=True) except FileExistsError: shutil.rmtree(dest_path) dest_path.mkdir(parents=True) # figure out the format for download, # by default select 480p, if not available, choose the best format available # mp4 format_id, fps = select_video_format(url, format_note='480p', ext='mp4') # download the video video = download_video(url,format_id) # calculate skip_frames try: skip_frames = int(fps * sampling_interval) except: skip_frames = int(30 * sampling_interval) print(f'video saved at: {video}, fps:{fps}, skip_frames: {skip_frames}') # extract video frames at given sampling interval with multiprocessing - print('extracting frames...') n_workers = os.cpu_count() with Pool(n_workers) as pool: pool.map(partial(process_video_parallel, video, skip_frames, dest_path, n_workers), range(n_workers)) return dest_path def captioned_strip(images, caption=None, times=None, rows=1): increased_h = 0 if caption is None else 30 w, h = images[0].size[0], images[0].size[1] img = Image.new("RGB", (len(images) * w // rows, h * rows + increased_h)) for i, img_ in enumerate(images): img.paste(img_, (i // rows * w, increased_h + (i % rows) * h)) if caption is not None: draw = ImageDraw.Draw(img) font = ImageFont.truetype( "/usr/share/fonts/truetype/liberation2/LiberationMono-Bold.ttf", 16 ) font_small = ImageFont.truetype( "/usr/share/fonts/truetype/liberation2/LiberationMono-Bold.ttf", 12 ) draw.text((20, 3), caption, (255, 255, 255), font=font) for i,ts in enumerate(times): draw.text(( (i % rows) * w + 40 , #column poistion i // rows * h + 33) # row position , ts, (255, 255, 255), font=font_small) return img def run_inference(url, sampling_interval, search_query): path_frames = vid2frames(url,sampling_interval) original_images, images = read_frames(path_frames) image_input = torch.tensor(np.stack(images)).to(device) with torch.no_grad(): image_features = model.encode_image(image_input) text_features = model.encode_text(clip.tokenize(search_query).to(device)) image_features /= image_features.norm(dim=-1, keepdim=True) text_features /= text_features.norm(dim=-1, keepdim=True) similarity = (100.0 * image_features @ text_features.T) values, indices = similarity.topk(4, dim=0) best_frames = [original_images[ind] for ind in indices] times = [f'{datetime.timedelta(seconds = ind[0].item() * sampling_interval)}' for ind in indices] image_output = captioned_strip(best_frames,search_query, times,2) title = search_query return(title, image_output) inputs = [gr.inputs.Textbox(label="Give us the link to your youtube video!"), gr.Number(5,label='sampling interval (seconds)'), gr.inputs.Textbox(label="What do you want to search?")] outputs = [ gr.outputs.HTML(label=""), # To be used as title gr.outputs.Image(label=""), ] gr.Interface( run_inference, inputs=inputs, outputs=outputs, title="It Happened One Frame", description='A CLIP-based app that search video frame based on text', examples=[ ['https://youtu.be/v1rkzUIL8oc', 1, "James Cagney dancing down the stairs"], ['https://youtu.be/k4R5wZs8cxI', 1, "James Cagney smashes a grapefruit into Mae Clarke's face"] ] ).launch(debug=True,enable_queue=True)