Release Notes

Release Notes#

0.1.0#

Initial scmiracle package release focused on continual integration for single-cell multimodal data.

API#

  • Public high-level helper: scmiracle.model.MIRACLE.

  • Initial training: MIRACLE.setup_mudata(...).train(...).

  • Continual training: MIRACLE.setup_continual(...) with replay memory and a previous checkpoint.

  • Inference: get_latent_representation, load_model_from_checkpoint_for_mdata, and get_latent_from_checkpoint.

  • Replay utilities: build_replay, load_replay, load_replay_with_metadata, attach_replay_metadata, and export_mudata_metadata.

  • Data migration utilities: load_mtx_dir_as_mudata, prefix_batch_and_obs_names, attach_feature_and_label_metadata, and get_obs_series.

Continual Integration#

  • Replay-aware training alternates current-task and replay batches.

  • Original per-batch cell counts are stored in checkpoint and replay metadata for stable weighting across incremental updates.

  • Compatible checkpoint weights are transferred into expanded feature and batch spaces when possible.

  • Lazy continual setup aligns features without materializing a full merged expression matrix.

API Check Status#

  • from scmiracle.model import MIRACLE imports successfully in the current workspace.

  • scmiracle.download_data.download and list_available_datasets are importable.

  • scmiracle.MIRACLE is not currently available because scmiracle/__init__.py does not export the class.