dataloader.py

import torch
from datasets import load_dataset
from torch.utils.data import DataLoader
from transformers import BertTokenizer
import decord
import numpy as np
from tqdm import tqdm


FRAMES = 50
H, W = 128, 128
BATCH_SIZE = 8
TEXT_MAX_LEN = 3000


dataset = load_dataset("gaussalgo/webvid-10m", split="train")  # 10M samples


tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")


class VideoDataset(torch.utils.data.Dataset):
    def __init__(self, dataset):
        self.dataset = dataset
        self.decord_ctx = decord.cpu(0)  # CPU decoding

    def __len__(self):
        return len(self.dataset)

    def __getitem__(self, idx):
        item = self.dataset[idx]
        

        vr = decord.VideoReader(item["video_path"], ctx=self.decord_ctx)
        frame_indices = np.linspace(0, len(vr)-1, FRAMES, dtype=int)
        video = vr.get_batch(frame_indices).numpy()  # (FRAMES, H, W, 3)
        video = torch.from_numpy(video).permute(3, 0, 1, 2).float()  # (3, FRAMES, H, W)
        
        
        video = F.interpolate(video, size=(H, W), mode="bilinear")
        video = (video / 255.0) * 2 - 1  # [-1, 1]
        
        
        text = tokenizer(
            item["caption"],
            padding="max_length",
            truncation=True,
            max_length=TEXT_MAX_LEN,
            return_tensors="pt"
        ).input_ids.squeeze(0)
        
        return {"video": video, "text": text}

# DataLoader
dataset = VideoDataset(dataset)
dataloader = DataLoader(dataset, batch_size=BATCH_SIZE, shuffle=True, num_workers=4)