October 26, 2026 · LayerGenius
Most augmentation pipelines are copied from a tutorial and never measured. Here is how we pick transforms from the failure modes in the data, place them so they don't starve the GPU, and ablate them one at a time.
Read moreOctober 19, 2026 · LayerGenius
A single validation number tells you almost nothing about what to fix. How we run error analysis on a Keras classifier: per-slice metrics, a ranked confusion inventory, Grad-CAM as a hypothesis test, and the rule for deciding between a data fix and a model change.
Read moreOctober 12, 2026
Quantization is the first compression lever, not the only one. Structured pruning, magnitude sparsity, and knowledge distillation in Keras 3 — what each one actually buys on real hardware, and how to measure it.
Read moreOctober 5, 2026
Most Keras bugs are software bugs, not modeling bugs. A small pytest suite — shape and dtype contracts, overfit-one-batch, leakage checks, serialization round-trips — catches them before a GPU run does.
Read moreSeptember 28, 2026 · LayerGenius
Most stalled Keras projects are data problems wearing a modeling costume. Out-of-fold confident-learning scans, embedding-based duplicate detection, and a fix-the-test-set-first policy — with the numbers to justify the hours.
Read moreSeptember 21, 2026 · LayerGenius
A search will happily optimize your validation-set noise. How we run KerasTuner: baseline first, four hyperparameters instead of twelve, executions_per_trial against spurious winners, a three-way split that matches deployment, and reading the top ten trials rather than the top one.
Read moreSeptember 21, 2026 · LayerGenius
A sigmoid output is not a probability and 0.5 is not a threshold. How to measure calibration on a Keras classifier, fix it with temperature scaling or isotonic regression, pick an operating point from the cost of each error, and keep both honest after deploy.
Read moreSeptember 17, 2026 · LayerGenius
Structured data is where deep learning most often loses to a 40-line GBDT. Here is how we decide: a boosted-tree baseline first, then Keras FeatureSpace for preprocessing that travels with the model, and an honest accounting of when the neural net is worth its operating cost.
Read moreSeptember 16, 2026 · LayerGenius
Before you fine-tune a pretrained language model, price the alternatives. Then: LoRA ranks, sequence length, bfloat16, memory math, and how to serve an adapter you can actually roll back.
Read moreSeptember 15, 2026
Most time-series projects fail on evaluation, not architecture. Seasonal-naive baselines, leakage-safe windowing with tf.data, rolling-origin backtesting, and the small set of Keras models worth trying.
Read moreSeptember 14, 2026
Bit-identical reruns are a tooling problem, not a research one. Seeding across backends, GPU determinism and what it costs, pinning the data pipeline, and the run record that makes a result defensible six months later.
Read moreSeptember 10, 2026 · LayerGenius
Data parallelism in Keras 3 across the TensorFlow and JAX backends: global batch size, learning-rate scaling, input-pipeline limits, checkpointing, and the measurements that tell you whether the second GPU earned its keep.
Read moreAugust 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 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 more