Non-Adversarial Shared-Encoder Disentangled VAE for Unpaired Fluorescence Restoration

This small PyTorch project implements a shared-encoder VAE that disentangles structural (shape) and style latents for restoration of degraded fluorescence microscopy images without any GANs.

Files

• models.py - Encoder, Decoder with FiLM conditioning, Segmentation U-Net
• losses.py - KL, alignment, Dice, total loss aggregation
• dataset.py - Unpaired loader reading data/clean/ and data/degraded/
• train.py - train_model(...) function entrypoint (import and call from notebook)
• utils.py - helpers for checkpointing and visualization

Quick start (in a notebook):

from encoder_CARE.train import train_model

train_model(data_root='./data', epochs=50, batch_size=8, device='cuda')

Data layout

• data/clean/ : clean fluorescence images (png, jpg, tiff)
• data/degraded/ : degraded images

Notes

• This implementation is intentionally minimal and designed to be imported from Python code or notebooks. There is no CLI or argparse.
• Adjust image size, channels, and hyperparameters in the code as needed for your dataset.