Skip to content

tutorial.md

Latest commit

 

History

History
41 lines (37 loc) · 2.71 KB

tutorial.md

File metadata and controls

41 lines (37 loc) · 2.71 KB
Raw

Training a custom AutoMorph model

The current /custom workflow trains a YOLO oriented-box detector and one dlib landmark predictor per class. It differs from the older standalone shape-predictor workflow shown in the Drosophila recording.

  1. Create Project: open Train Custom Model Wizard from the landing page; enter a project name and organism.
  2. Add Data: upload a ZIP, or select one TPS/XML annotation file and every referenced image together. ZIPs may contain subdirectories; annotation paths must resolve unambiguously. Use at least two annotated images, consistent class names and landmark names/order, and decoded image dimensions matching the annotations. Retain specimen identifiers separately for study-level splitting.
  3. Check Data: inspect every displayed box and landmark overlay. Confirm image orientation, labels, and padding; do not proceed with reflected, displaced, or missing points. TPS coordinate conversion depends on the actual image height.
  4. Train Model: choose a preset or expert parameters, then Start Training Job. Defaults include nu 0.1, tree depth 4, cascade depth 15, oversampling 5, feature pool 400, candidate tree splits 20, and validation fraction 0.2. Candidate tree splits are a model hyperparameter, not cross-validation folds.
  5. Follow Progress: keep the job ID. Training runs in a separate process; concurrent submissions to the same orchestrator are rejected as busy, not queued. Cancelling or closing the backend interrupts training; retain the error/log if the job fails. Do not treat the presence of a model file as proof of completion.
  6. Review Results: inspect detector metrics and per-class landmark error. The landmark value is an internal crop-space error, not an end-to-end accuracy guarantee. Read the metric limitations.
  7. Publish & Use: select the registered model for prediction. Verify predictions on new images and export annotations through the UI. Preserve unedited outputs separately from human corrections.

The generic upload limit defaults to 1 GiB locally and 512 MiB with AUTOMORPH_HOSTED=true (legacy alias LIZARDMORPH_HOSTED). Override with TRAINING_MAX_BYTES. Archive defaults are 20,000 files and 1 GiB uncompressed, configured by TRAINING_MAX_ARCHIVE_FILES and TRAINING_MAX_UNCOMPRESSED_BYTES. Older /train_predictor endpoints have different limits; do not use their old 100 MB/500-file documentation for this wizard.

For reproducible work, archive the complete training job folder and follow the research handoff guide. The tiny sample dataset is for workflow checks, not an accuracy study.