xenopus-edge-pytorch

PyTorch-Lightning checkpoint trained with kapoorlabs_vollseg (models_edge_pytorch). Flat layout — the folder ships:

• last.ckpt (and optionally per-epoch <model_name>-epoch=NNN.ckpt)
• training_config.json — Hydra parameters block, what kapoorlabs_vollseg reads first to rebuild the architecture
• <model_name>.json — legacy CareInception fallback config

StarDist rays are regenerated deterministically from (conv_dims, n_rays, anisotropy) in the JSON; no rays.npy sidecar is needed.

Loading

# StarDist
from kapoorlabs_vollseg import StarDistSegmenter, ensure_model
folder = ensure_model("./local_models", "models_edge_pytorch",
                      repo_id="KapoorLabs/xenopus-edge-pytorch")
star = StarDistSegmenter.from_folder(folder)
labels = star.predict(volume).labels

# U-Net
from kapoorlabs_vollseg import UNetSegmenter, ensure_model
folder = ensure_model("./local_models", "models_edge_pytorch",
                      repo_id="KapoorLabs/xenopus-edge-pytorch")
unet = UNetSegmenter.from_folder(folder)
labels = unet.predict(volume).labels

# CARE
from kapoorlabs_vollseg import CAREDenoiser, ensure_model
folder = ensure_model("./local_models", "models_edge_pytorch",
                      repo_id="KapoorLabs/xenopus-edge-pytorch")
care = CAREDenoiser.from_folder(folder)
denoised = care.predict(volume).denoised

See https://github.com/Kapoorlabs-CAPED/KapoorLabs-VollSeg for the full segmentation pipeline.