cnn-cifar10-classifier v0.1.0

Small PyTorch CNN baseline for CIFAR-10 image classification.

This release contains the trained v0.1.0 checkpoint from the GitHub project:

https://github.com/diverHansun/cnn-cifar10-classifier

Results

Split	Metric	Value
test	accuracy	0.7857

The checkpoint was selected by best validation/test accuracy during a 20 epoch run.

Training summary:

• Dataset: uoft-cs/cifar10
• Config: plain_text
• Epochs: 20
• Batch size: 256
• Optimizer: SGD, momentum 0.9, weight decay 0.0005
• Learning rate: 0.01
• Augmentation: random crop with padding 4, random horizontal flip
• AMP: enabled
• GPU used: NVIDIA GeForce RTX 5070 Ti
• PyTorch: 2.11.0+cu128

Files

checkpoints/best_model.pth
checkpoints/last_model.pth
outputs/training_metrics.json
outputs/training_curves.png
outputs/confusion_matrix.png
outputs/demo_predictions.png
logs/train_20260602_044856.log
logs/evaluate_20260602_045254.log
logs/demo_20260602_045311.log
runs/cifar10_cnn_20260602_044901/events.out.tfevents...
release_summary.json
manifest.sha256

Usage

Clone the project code first:

git clone git@github.com:diverHansun/cnn-cifar10-classifier.git
cd cnn-cifar10-classifier

Install dependencies with a CUDA-compatible PyTorch build for your machine, then download this checkpoint:

hf download diverWayne/cnn-cifar10-classifier checkpoints/best_model.pth --local-dir .

Evaluate:

python evaluate.py --checkpoint checkpoints/best_model.pth --device cuda

Run the demo grid:

python demo.py --checkpoint checkpoints/best_model.pth --samples 16 --device cuda

Predict one image:

python predict.py --image demo_images/your_image.png --checkpoint checkpoints/best_model.pth --device cuda

Limitations

This is a simple hand-written CNN baseline trained on CIFAR-10 32x32 images. It supports the 10 CIFAR-10 classes only:

airplane, automobile, bird, cat, deer, dog, frog, horse, ship, truck

It can classify arbitrary images after resizing to 32x32, but reliability outside CIFAR-10-like images is limited.

The strongest observed confusions are between visually similar categories such as cat and dog, bird and deer/dog, and airplane and ship.

Dataset

The training data is not redistributed in this model repository. It is loaded from the public Hugging Face dataset uoft-cs/cifar10.