bryanwong17/graph-transformer-with-self-supervised-pre-training-in-digital-histopathology

Graph Transformer with Self-Supervised Pre-training in Digital Histopathology Documentation

1. Patch Tiling

How to run:

• python patch_tiler.py

Note:

• check arg parser for the configurations

2. Clustering (optional)

How to run:

• python preprocessing/clustering.py

Note:

• number of clusters is equal to batch size (32)

3. Training Patch Feature Extractor

How to run:

• python feature_extractor/feature_extract.py

Notes:

• check config.yaml for configuration
• change either want to use SimCLR, SimSiam or MoCoV3 in feature_extract.py (simply comment/uncomment it)
• if want to use Pytorch Lightning, run feature_extract_lightning.py instead (only MoCoV3 is ready)
• model path: ../../graph_transformer/runs/(simclr or mocov3)/runs/(name of SSL model)

4. Build Graph

How to run:

• python feature_extractor/build_graphs.py

Notes:

• check arg parser for configuration
• choose between loading SimCLR, SimSiam, and MoCoV3 model in build_graphs.py (simply comment/uncomment it)
• graph path: ../../graph_transformer/build_graphs/(simclr, mocov3, or simsiam)/(name of graph)

5. Training Graph Transformer

How to run:

• cd ..
• python main.py (check arg parser at option.py)

Notes:

• graph path: ../graph_transformer/build_graphs/(simclr, mocov3, or simsiam)/(name of graph)/(simclr_files or simsiam_files)
• when testing: train=False, test=True, val_set=(test data), resume=(graph VIT model path)
• result path: ../graph_transformer/results_with_graph/(simclr, mocov3, or simsiam)/(name of result folder)

6. Training without graph (VIT)

How to run:

• change load model in main.py to "from models.OnlyVisionTransformer import Classifier"
• python main.py (check arg parser at option.py)

Note:

• result path: ../graph_transformer/results_without_graph/(simclr, mocov3, or simsiam)/(name of result folder)