SK-AutoD diagnoses your ML training in seconds. Overfitting, exploding gradients, learning rate issues — detected automatically with confidence scores and fix recommendations.
No configuration. No ML model needed to diagnose your ML model. Pure Python signal analysis, offline, instant.
BaseDetector, implement one method, pass it to diagnose(). Your custom pattern in three minutes.Each detector returns a severity level, confidence score, root cause, and a concrete fix recommendation.
| Detector | Severity | Signal detected | Fix hint |
|---|---|---|---|
| Classic overfitting | critical | Val loss rises while train loss falls | Dropout, L2 reg, ↓ capacity |
| Exploding gradient | critical | Loss spikes >300% in one epoch | Gradient clipping, ↓ LR |
| LR too high | high | Loss oscillates without downtrend | Reduce LR 5–10×, add warmup |
| Underfitting | high | Both losses plateau at high values | ↑ capacity, train longer |
| Dying ReLU proxy | high | Loss flatlines early at high value | He init, Leaky ReLU, BatchNorm |
| Data leakage proxy | high | Val loss consistently < train loss | Audit train/val split |
| LR too low | medium | Loss decreases extremely slowly | ↑ LR, cosine decay schedule |
| Noisy training | medium | Jagged loss, frequent direction flips | ↑ batch size, gradient clipping |
| Label noise floor | medium | Loss cannot drop below a high floor | Label smoothing, audit data |
| Missed early stopping | warning | Val minimum not used as checkpoint | EarlyStopping, patience=3 |
Use the full report, the one-liner, or embed a live callback in your training loop — whichever fits your workflow.
SK-AutoD is under active development. The API is stabilising fast.
Requires Python 3.10+. NumPy and SciPy are the only dependencies.
Use the homepage for the overview, docs for depth, and tutorial for the shortest path to first value.