Skip to content

Inference Guide

Use this guide alongside Workflow Step 8. The workflow covers the basic commands; here we compare inference strategies and highlight when to apply each.

Quick Commands

# Evaluate a single checkpoint on the test split
ml-inference --checkpoint_path runs/my_dataset_fold_0/weights/best.pt

# Evaluate on validation data instead of test
ml-inference --checkpoint_path runs/my_dataset_fold_0/weights/best.pt --split val

# Supply a different config (rare)
ml-inference --checkpoint_path runs/my_dataset_fold_0/weights/best.pt --config configs/alt.yaml

Outputs land under the run directory (logs/classification_report_test.txt, TensorBoard metrics, confusion matrices).

Strategy Overview

Strategy When to choose it CLI shortcut
standard Baseline evaluation, CPU runs, simplicity ml-inference --checkpoint_path ...
mixed_precision Faster GPU inference with minimal memory Enable via config (inference.strategy: mixed_precision) or matching CLI flags if exposed
tta Single model, boost robustness with augmentations ml-inference --checkpoint_path ... --tta [--tta-augmentations ...]
ensemble Combine multiple folds/models ml-inference --ensemble run1/best.pt run2/best.pt ...
tta_ensemble Maximum accuracy regardless of cost Add --tta to the ensemble command

For in-depth tuning of TTA parameters or ensemble weighting, see the dedicated guides below.

Practical Tips

  • Always evaluate the checkpoint saved as best.pt; it reflects the highest validation score.
  • When comparing multiple runs, write results to a table (accuracy, precision, recall) using the generated classification reports.
  • Keep inference configuration aligned with training (transforms, class order). If you need to override, provide the exact config via --config.