tsilva/mnist-cnn-classifier

MNIST CNN Classifier

This is a compact convolutional MNIST digit classifier trained as part of the dlab deep-learning experimentation roadmap.

The model was selected by validation loss, then evaluated once on the held-out MNIST test set across five random seeds. The uploaded checkpoint is the seed-1 run, which had the best validation loss in the final 5-seed evaluation sweep.

Results

5-seed held-out test evaluation:

metric	result
test accuracy	99.6140% ± 0.0802 pp
test loss	0.14677 ± 0.00272
best validation loss	0.14282 ± 0.00138

Per-seed held-out test results:

seed	test accuracy	test loss	best validation loss
1	99.6100%	0.14772	0.14124
2	99.6800%	0.14437	0.14317
3	99.4800%	0.15107	0.14403
4	99.6300%	0.14573	0.14150
5	99.6700%	0.14494	0.14418

Architecture

• Dataset: MNIST
• Model: CNN
• Channels: 32, 64, 128
• Dropout: 0.1
• Optimizer: Adam
• Learning rate: 0.001
• Scheduler: OneCycleLR
• Weight decay: 0.0001
• Label smoothing: 0.02
• Weight averaging: EMA
• Batch size: 512
• Train augmentation: random affine, 10 degrees, translate 0.05, scale 0.9-1.1
• Early stopping: validation loss, patience 8, min_delta 0.0005

Files

• model.ckpt: PyTorch Lightning checkpoint from the validation-selected seed-1 run.
• model.onnx: ONNX export of the checkpoint for inference.
• config.yaml: resolved Hydra training config.
• metrics.csv: training metrics from the uploaded checkpoint run.
• metrics_summary.csv: compact 5-seed final evaluation summary.
• metadata.json: compact metadata for inference and provenance.

Preprocessing

Inputs should be MNIST grayscale images converted to tensors and normalized with:

mean = [0.1307]
std = [0.3081]

The ONNX model expects float tensors shaped [batch, 1, 28, 28] under the input name images, and returns class logits under the output name logits.

Provenance

• Uploaded checkpoint run: https://wandb.ai/tsilva/dlab/runs/5um57rnu
• Final 5-seed test sweep: https://wandb.ai/tsilva/dlab/sweeps/ikfs5ox8