You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
legacy reproduces OTU-Former 0.2.1 augmentation exactly, for old-run continuation and comparison.
Orientation policy
New --orientation-policy invariant|sensitive on both training commands. The new-run default is sensitive: rotation [-15°, 15°] and no horizontal reflection for every global/local pretraining view and for fine-tuning conservative. invariant is the opt-in broad-rotation + horizontal-reflection policy for orientation-insensitive markers. Neither policy is inferred from the image or taxon.
Reproducibility and resume
Checkpoints store config.augmentation_profile and a fully expanded config.augmentation_config (including orientation_policy).
Resume inherits omitted profile/policy/local-view values and rejects conflicts. Fine-tuning initialized from --checkpoint never inherits the pretraining profile, and inherits the saved orientation policy when omitted.
Checkpoint metadata is validated for self-consistency: a top-level/nested profile mismatch, a missing nested profile, or an unknown saved policy is rejected as malformed augmentation metadata.
CLI
Detailed augmentation contract in pretrain --help / finetune --help.
Shell completion for --augmentation and --orientation-policy values.
--local-crop-size / --local-crops inherit the checkpoint values on --resume.
Platform fix
PYTORCH_ENABLE_MPS_FALLBACK is enabled before augmentation validation, so the new default profile works on Apple Silicon with --device auto.
Breaking changes
New pretraining runs default to global-barcode + orientation-policy=sensitive instead of the historical augmentation. Resuming pre-0.4.0 checkpoints keeps the compatible legacy/invariant behavior.