mammadov7/cdpath21-gan

Self-Supervised Representation Learning (CDpath ICCV 2021)

cdpath21-gan

Code to train self-supervised adversarial autoencoder for visual field expansion of histopathology tiles [1].

• main.py contains training code.
• src/models.py defines GAN generator and discriminators.
• src/utils.py defines utility functions for training and graphing.
• config/ defines .yml configuration files to set experiment parameters.

Installation

The Anaconda environment is specified in environment.yml. The environment can be recreated using,

conda env create -f environment.yml

Tested with single NVIDIA P100 GPU, running Cuda 10.0.130, and PyTorch 1.9.0 with torchvision 0.10.0.

Usage

main.py is the training code, which requires two parameters

• job_number specifies a unique identifier for writing outputs
• config specifies configuration file path

See slurm_submit.sh for example.

Config files

See config/README.md for a description of configuration options.

Data

See data/README.md for data procurement instructions.

Reference

[1] Self-Supervised Representation Learning using Visual Field Expansion on Digital Pathology, Joseph Boyd, Mykola Liashuha, Eric Deutsch, Nikos Paragios, Stergios Christodoulidis, Maria Vakalopoulou, CDpath ICCV 2021 [PDF]

Reconstructions (CAMELYON17)

Samples (CAMELYON17, CRC)

Model extension - coordinates

Decoder can be trained with coordinates of crop input to control position of reconstruction. See coordinates branch for code.