Working notes from applied deep-learning practice.
August 6, 2026
One line enables mixed precision — and three known failure modes decide whether it helps. float16 vs bfloat16, loss scaling, the float32 output layer, and the A/B test that settles it.
Read moreJuly 16, 2026
A trained model is not a deliverable — a verified inference artifact is. Export contracts, op coverage, the quantization ladder, and the parity tests that make conversions safe.
Read moreJune 24, 2026
Honest splits, a linear-probe baseline, gradual unfreezing with frozen BatchNorm, domain-respecting augmentation, and one change at a time — the fine-tuning recipe we use in Keras.
Read moreJune 3, 2026
When GPU utilization is low, the input pipeline is usually the culprit. Seven checks, ordered by cost and payoff: prefetch, parallel map, cache, small-files fixes, and more.
Read moreMay 12, 2026
Most model definitions port to Keras 3 with minor edits — the pain concentrates in custom ops, keras.backend usage, serialization, and training loops. A field checklist for the migration.
Read more