segmentation

Semantic Segmentation Fine-tuning

Code for training semantic segmentation models on the Pannuke and Lizard datasets.

Requirements

• Python 3.8+
• pip install -r requirements.txt

Data Preparation

Pannuke

1. Download all three folds of the Pannuke dataset from here.
2. Extract all archives into one directory (e.g. pannuke/).
3. To convert the .npy files to images, run python scripts/process_pannuke -i path/to/pannuke.

• Note: The original masks.npy and types.npy files for each fold are used for evaluation.
  • i.e. --test_target_path fold1/masks/fold1/masks.npy --test_types_path fold1/images/types.npy

Lizard

1. Download the patch-level Lizard dataset from here (registration required) and label files from here.
2. Extract all archives into one directory (e.g. lizard/).
3. To convert the .npy files to images run python scripts/process_lizard.py -i path/to/lizard
4. To generate target files run python scripts/create_targets_lizard.py -i path/to/processed/lizard
   • These .npy files are used for evaluation (i.e. --test_target_path targets-fold1.npy).

Usage

Training

• To train from scratch on Pannuke run:

python train.py  --gpus 1 --precision 16  --name pannuke --model.n_classes 6 --max_epoch 100 
--data.train_fold "['data/pannuke/fold1/', 'data/pannuke/fold2']" 
--data.test_fold data/pannuke/fold3  --test pannuke 
--test_targets_path eval/targets/pannuke-f3.py --test_types_path eval/types/pannuke-f3.npy 

• configs/ contains example configuration files and can be run with python train.py --config path/to/config/file.
• Run python train.py --help for information on all options.
• This codebase uses the Segmentation Models PyTorch library and therefore --model.arch can be any of architectures from here and --model.encoder can be any of the encoders from here.

Mask Generation:

• To generate pseudo-label masks for a dataset using a trained segmentation model run:

python scripts/generate_masks.py -w path/to/segmentation/checkpoint -i path/to/image/directory
-o path/to/output/masks/directory --n_classes <number-of-classes>

Acknowledgements

Evaluation code is taken from the official Pannuke evaluation repo.