preprocessing.py

import numpy as np
import torch
from torchvision import transforms
import av
import logging
import base64
import io

logging.basicConfig(filename='/mnt/data/uploads/logfile-video.log', level=logging.INFO)

def makeStack(video_base64):
    video_data = base64.b64decode(video_base64)

    container = av.open(io.BytesIO(video_data))
    frames = []
    for frame in container.decode(video=0):
        frames.append(frame.to_ndarray(format="rgb24").astype(np.uint8))
    
    return np.stack(frames, axis=0)

def read_video(video_base64, num_frames=24, target_size=(224, 224)):
    # video_data = base64.b64decode(video_base64)

    # container = av.open(io.BytesIO(video_data))
    # frames = []
    # for frame in container.decode(video=0):
    #     frames.append(frame.to_ndarray(format="rgb24").astype(np.uint8))
    
    # sampled_frames = sample_frames(frames, num_frames)
    # processed_frames = pad_and_resize(sampled_frames, target_size)
    # return processed_frames
    frames = makeStack(video_base64)
    frames = sample_frames(frames, num_frames)
    processed_frames = pad_and_resize(frames, target_size)
    return processed_frames

def sample_frames(frames, num_frames):
    total_frames = len(frames)
    sampled_frames = list(frames)
    if total_frames <= num_frames:
        # sampled_frames = frames
        if total_frames < num_frames:
            padding = [np.zeros_like(frames[0]) for _ in range(num_frames - total_frames)]
            sampled_frames.extend(padding)
    else:
        indices = np.linspace(0, total_frames - 1, num=num_frames, dtype=int)
        sampled_frames = [frames[i] for i in indices]

    return np.array(sampled_frames)


def pad_and_resize(frames, target_size):
    transform = transforms.Compose([
        transforms.ToPILImage(),
        transforms.Resize(target_size),
        transforms.ToTensor()
    ])
    processed_frames = [transform(frame) for frame in frames]
    processed_frames = torch.stack(processed_frames)
    # return processed_frames.permute(1, 0, 2, 3).unsqueeze(0)  # Add batch dimension and permute [3, 24, 224, 224]
    # return processed_frames.permute(0, 2, 3, 1).unsqueeze(0) # [24, 224, 224, 3]
    return processed_frames.permute(0, 1, 2, 3).unsqueeze(0)