CIFAR-10 scratch baseline

Question

How well can a small residual CNN trained from scratch do on CIFAR-10 with a plain, reproducible recipe — and does the library reproduce the reference result without special-case code?

This is the reference baseline for the project. It is deliberately a standard recipe, not a state-of-the-art submission: the point is a trustworthy, config-driven training pipeline that reports its own behavior honestly.

Protocol

Component	Value
Data	CIFAR-10, 45,000 train / 5,000 validation / 10,000 test
Split	plain random, seed 42
Model	residual_cnn, 3→64 stem, four stages, widths 64/128/256/512
Parameters	7,827,786
Augmentation	flip, random crop (32, padding 4), mild color jitter, random erasing
Optimizer	Adam, lr 1e-3, weight decay 5e-4
Schedule	cosine, no warmup, no gradient clipping
Budget	50 epochs, batch 128, early stopping patience 10
Device	Apple M3 Max MPS

Model design

The reference preset is a 7.8M-parameter residual CNN: a 3→64 stem, four residual stages at widths 64/128/256/512, global average pooling, dropout 0.3, and a linear head. The schematic shows the stage shapes and parameter counts; the model card in the report lists every module.

Result

Field	Value
Best validation accuracy	92.78% (epoch 46)
Test accuracy	92.25%
Test macro F1	92.26%
Test top-5 accuracy	98.90%
Test accuracy 95% CI	91.76–92.77%
Test NLL	0.2894
Calibration ECE	0.0434
Latency	2.08 ms/image on MPS
Training time	1,338 s (22.3 min)

Reference: 92.42% validation / 92.61% test on MPS, seed 42. The test difference is −0.36 percentage points; the validation difference is +0.36. The ≥90% gate is met. No tuning was performed against the test split.

Artifacts

The run directory contains best.pt, last.pt, the resolved config, environment and git state, the data contract and fingerprints, history, metrics, test predictions, figures, and self-contained HTML/Markdown reports. Checkpoints stay local; committed result packaging lives under results/.

• Demo source
• Config
• Measured results

What to notice

• One YAML file describes the experiment. The model is a generic residual_cnn preset, not a CIFAR-specific class.
• The test split is isolated. Trainer.fit() never iterates the test loader; test is evaluated once from the best validation checkpoint.
• The result is reproducible. Seed, environment, git state, and data fingerprint are recorded in the run.
• The report is honest. It shows calibration, per-class metrics, and the error gallery alongside accuracy.

Caveat

This is a reference reproduction on one Apple Silicon machine, not a state-of-the-art benchmark. CIFAR-10 is saturated; the interesting part is the pipeline, the discipline, and the diagnostics.