ZoneTwelve CIFAR-10 model suite
This repository contains checkpoints from the MPS-trained CIFAR-10 model
comparison suite. Each .pt file preserves model weights, optimizer state,
epoch history, architecture metadata, and validation-selection metadata.
Fairness note: DenseCNN was trained for 200 epochs; the other headline models were trained for 30 epochs. DenseCNN's 92.53% is not a matched-budget comparison. The best completed 30-epoch result is CNN-C + BN at 84.27%.
Performance benchmarks
The following charts show MPS inference performance for every completed checkpoint benchmark. They measure execution cost only; they do not rank model accuracy.
Raw per-run JSON results and the combined summary are in the
performance/ directory. Benchmarks used synthetic CIFAR
inputs, batch size 128, 10 warmup iterations, and 50 measured iterations.
Final results
| Model | Run epochs | Selected epoch | Best validation | Test accuracy |
|---|---|---|---|---|
| DenseCNN | 200* | 192 | 93.30% | 92.53% |
| CNN-C + BN | 30 | 27 | 86.24% | 84.27% |
| CNN-C + LN | 30 | 30 | 84.00% | 82.74% |
| R-CNN-B + BN | 30 | 29 | 82.40% | 81.22% |
| MobileNet-A + BN | 30 | 29 | 62.24% | 62.68% |
| MobileNet-A | 30 | 27 | 53.74% | 53.35% |
| CNN-B + BN | 30 | 29 | 83.02% | 81.55% |
| ViT | 30 | 28 | 70.66% | 69.08% |
The full per-variant report is available in reports/ in the source repository.
Selected epoch is the epoch with the highest validation accuracy; its checkpoint state is used for the reported test accuracy. Run epochs is the total training budget.
* DenseCNN used a longer training budget; rerun all models for 200 epochs for
a fair comparison.
| Checkpoint family | Notes |
|---|---|
cnn_A/B/C_{def,bn,ln}.pt |
CNN capacity and normalization variants |
rcnn_B_bn.pt |
recurrent convolutional classifier |
vit.pt |
Vision Transformer |
densecnn.pt |
DenseNet-inspired CNN; 92.53% test accuracy |
dynamic_cnn.pt |
Dynamic capacity CNN |
Source code and synchronized reports are available at https://github.com/ZoneTwelve/cifar-baselines.