uoft-cs/cifar10
Viewer β’ Updated β’ 60k β’ 292k β’ 142
A compact VGG-style convolutional network trained from scratch on CIFAR-10. It reaches 92.23% top-1 test accuracy with under 1.2M parameters and trains in about 8.5 minutes on a single T4 GPU.
Linear(256β192) β ReLU β Dropout(0.3) β Linear(192β10)Educational and portfolio use: a small, reproducible baseline for CIFAR-10 and for experiments such as ablations, augmentation or compression.
Not intended for real-world recognition tasks. It only knows 10 CIFAR classes at 32Γ32 resolution.
| Data | CIFAR-10: 45,000 train / 5,000 validation / 10,000 test |
| Augmentation | random crop (padding 4), horizontal flip |
| Optimiser | SGD, Nesterov momentum 0.9, weight decay 5e-4 |
| Schedule | one-cycle LR, max 0.1, 20% warm-up |
| Regularisation | label smoothing 0.1, dropout 0.3 |
| Epochs / batch | 30 / 128, mixed precision |
| Selection | best validation accuracy |
| Hardware | Tesla T4, about 8.5 minutes |
| Metric | Value |
|---|---|
| Test accuracy (top-1) | 92.23% |
| Test accuracy (top-3) | 98.77% |
| Best validation accuracy | 91.66% |
The test set was used once, after model selection on the validation split.
The class definition is in the training notebook (cnn_cifar10_classifier.ipynb).
ckpt = torch.load("cifar_cnn.pt", map_location="cpu")
model = CifarCNN(); model.load_state_dict(ckpt["state_dict"]); model.eval()
CIFAR-10 (Krizhevsky, Nair and Hinton) β see the dataset page for terms. Code and weights: MIT.