xenopus-cellfate-inception-pytorch

PyTorch-Lightning InceptionNet checkpoint trained with kapoorlabs_lightning (inception_nmorpho_fate). Classifies cell-fate labels from 25-timepoint sliding-window tracklets of 18 morphodynamic features (6 shape + 12 dynamic channels: SHAPE_DYNAMIC_FEATURES in kapoorlabs_lightning.tracking.track_features).

The folder ships:

• epoch=N-step=M.ckpt — Lightning checkpoint with optimizer + scheduler state.
• training_config.json — Hydra parameters block, what predict-cellfate.py reads first to rebuild the architecture (architecture knobs from the JSON win over the prediction yaml so the same predict script handles checkpoints trained with different arches).
• <experiment_name>.json / <experiment_name>.npz — CustomNPZLogger side-cars (per-epoch metrics).

Loading

from kapoorlabs_lightning.hub import ensure_model

# Auto-downloads to ./local_models/inception_nmorpho_fate/ if missing
folder = ensure_model("./local_models", "inception_nmorpho_fate",
                      repo_id="KapoorLabs/xenopus-cellfate-inception-pytorch")

# Then point predict-cellfate.py at it:
#   python predict-cellfate.py \
#       experiment_data_paths.checkpoint_path=<folder>

Or simply invoke predict-cellfate.py without a checkpoint_path override — it falls back to this model automatically:

python scripts/model_prediction/predict-cellfate.py \
    experiment_data_paths.xml_file=<your_xml> \
    experiment_data_paths.output_dir=<results_dir>

See https://github.com/Kapoorlabs-CAPED/KapoorLabs-Lightning for the full cell-fate pipeline (training, prediction, transition-time determination, GT confusion-matrix).