README.md

torchvision-tinyimagenet

Dataset class for PyTorch and the TinyImageNet dataset.

Installation

pip install tinyimagenet

How to use

from tinyimagenet import TinyImageNet
from pathlib import Path
import logging

logging.basicConfig(level=logging.INFO)

split ="val"
dataset = TinyImageNet(Path("~/.torchvision/tinyimagenet/"),split=split,imagenet_idx=False)
n = len(dataset)
print(f"TinyImageNet, split {split}, has  {n} samples.")
n_samples = 5
print(f"Showing info of {n_samples} samples...")
for i in range(0,n,n//n_samples):
    image,klass = dataset[i]
    print(f"Sample of class {klass:3d}, image {image}, words {dataset.idx_to_words[klass]}")

The imagenet_idx indicates if the dataset's labels correspond to those in the full ImageNet dataset. By default (imagenet_idx=False) the labels are renumbered sequentially so that the 200 classes are named 0, 1, 2, ..., 199.

You can also check the quickstart notebook to peruse the dataset.

Finally, we also provide some example notebooks that use TinyImageNet with PyTorch models:

• Evaluate a pretrained EfficientNet model
• Train a simple CNN on the dataset
• Finetune an EfficientNet model pretrained on the full ImageNet to classify only the 200 classes of TinyImageNet