EndermanPC/ffhq512-processed-pt

FFHQ-512 Processed Chunks (PyTorch Tensors)

This dataset contains 7,000 images sourced from Ryan-sjtu/ffhq512-caption, preprocessed and saved as chunked PyTorch tensor (.pt) files.

Preprocessing Details

Images are preprocessed using a standardized pipeline:

• Color Format: Converted to RGB.
• Resize: Resized to (512, 512) using BICUBIC interpolation.
• Normalization: Normalized to the range [-1.0, 1.0] with mean=[0.5, 0.5, 0.5] and std=[0.5, 0.5, 0.5].
• Tensor Format: Stored as torch.float32. Each chunk has the shape (chunk_size, 3, 512, 512).

Dataset Structure

• Train: 5,000 images (train/train_xxxx.pt)
• Validation: 1,000 images (val/val_xxxx.pt)
• Test: 1,000 images (test/test_xxxx.pt)

ffhq_processed_pt/
├── train/
│   ├── train_0000.pt
│   └── ...
├── val/
│   ├── val_0000.pt
│   └── ...
└── test/
    ├── test_0000.pt
    └── ...

Quick Start

1. Download the Dataset

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="your_username/your_repo_name",
    repo_type="dataset",
    local_dir="./ffhq_processed_pt",
)

2. Load Data in PyTorch

import torch
from torch.utils.data import TensorDataset, DataLoader

# Load a single chunk (e.g., 250 images)
tensor_chunk = torch.load("ffhq_processed_pt/train/train_0000.pt")

# Slice a smaller subset (e.g., 50 images) for quick testing/sanity check
small_train = tensor_chunk[:50]

dataset = TensorDataset(small_train)
dataloader = DataLoader(dataset, batch_size=8, shuffle=True)

for (batch,) in dataloader:
    print(batch.shape)  # torch.Size([8, 3, 512, 512])
    break