In this example, a simple CNN model is trained from scratch on the CIFAR-10 dataset by leveraging the ImageFolderDataset struct to retrieve images from a folder structure on disk.

Since the original source is in binary format, the data is downloaded from a fastai mirror in a folder structure with .png images.

cifar10
├── labels.txt
├── test
│   ├── airplane
│   ├── automobile
│   ├── bird
│   ├── cat
│   ├── deer
│   ├── dog
│   ├── frog
│   ├── horse
│   ├── ship
│   └── truck
└── train
    ├── airplane
    ├── automobile
    ├── bird
    ├── cat
    ├── deer
    ├── dog
    ├── frog
    ├── horse
    ├── ship
    └── truck

To load the training and test dataset splits, it is as simple as providing the root path to both folders

let train_ds = ImageFolderDataset::new_classification("/path/to/cifar10/train").unwrap();
let test_ds = ImageFolderDataset::new_classification("/path/to/cifar10/test").unwrap();

as is done in CIFAR10Loader for this example.

The CNN model and training recipe used in this example are fairly simple since the objective is to demonstrate how to load a custom image classification dataset from disk. Nonetheless, it still achieves 70-80% accuracy on the test set after just 30 epochs.

Run it with the Torch GPU backend:

export TORCH_CUDA_VERSION=cu124
cargo run --example custom-image-dataset --release --features tch-gpu

Run it with our WGPU backend:

cargo run --example custom-image-dataset --release --features wgpu