Multinucleation VAE — Local Notes

This folder is the development workspace for the Darth VAEder project applied to myotube multinucleation imaging data.

What this is

A β-VAE trained on single-cell patches (256×256 px) from WGA-stained myotubes. Each cell contributes two image channels (membrane WGA + nuclear NLS) plus a crop mask. The model learns a compact 10-dimensional representation that we hope captures multinucleation state and morphology.

Repo

GitHub: Cryaaa/Darth_VAEder, branch JS_training

Quick start (server training)

ssh S-JS@3.17.63.55
conda activate darth-vaeder
cd /home/S-JS/Darth_VAEder
git pull origin JS_training

python "Joaquin'scripts/train.py" \
    --zarr  /mnt/efs/dl_jrc/student_data/S-JS/multinucleation.zarr \
    --table outputs/cell_table.csv \
    --out   outputs \
    --epochs 50

Data

Dataset	Location	Channels	Replicates	Notes
WGA + NLS	zarr on server	membrane, nuclei	N1–N3	16,678 cells in zarr
Phalloidin + DAPI	/Users/joaco/Documents/Janelia/Phalloidin and DAPI/	actin, nuclei	IDM09–IDM24	local TIFFs, not yet in zarr

Pipeline stages

1. Raw TIFFs → migrate_to_zarr.py → multinucleation.zarr (on server EFS)
2. zarr enrichment → add_pCellmask.py → adds pCellmask per cell group
3. Training → train.py → checkpoints + TensorBoard logs in outputs/

See also

CLAUDE.md — detailed notes for Claude Code sessions (architecture, gotchas, decisions)