![pre-commit-ci[bot]](https://avatars.githubusercontent.com/in/68672?v=4&size=40){kind=avatar}
This folder contains examples to run train and validate a deepgrow 2D/3D model. It also has notebooks to run inference over trained model.
implementation based on:
Sakinis et al., Interactive segmentation of medical images through fully convolutional neural networks. (2019) https://arxiv.org/abs/1903.08205
Training a deepgrow model requires data. Some public available datasets which are used in the examples can be downloaded from Medical Segmentation Decathlon or Synapse.
- For questions relating to the use of MONAI, please us our Discussions tab on the main repository of MONAI.
- For bugs relating to MONAI functionality, please create an issue on the main repository.
- For bugs relating to the running of a tutorial, please create an issue in this repository.
This example is a standard PyTorch program and helps user to prepare training input for 2D or 3D.
# Run to know all possible options
python ./prepare_dataset.py -h
# Prepare dataset to train a 2D Deepgrow model
python ./prepare_dataset.py
--dimensions 2 \
--dataset_root MSD_Task09_Spleen \
--dataset_json MSD_Task09_Spleen/dataset.json \
--output deepgrow/2D/MSD_Task09_Spleen
# Prepare dataset to train a 3D Deepgrow model
python ./prepare_dataset.py
--dimensions 3 \
--dataset_root MSD_Task09_Spleen \
--dataset_json MSD_Task09_Spleen/dataset.json \
--output deepgrow/3D/MSD_Task09_SpleenThis example is a standard PyTorch program and helps user to run training over pre-processed dataset for 2D.
# Run to know all possible options
python ./train.py -h
# Train a 2D Deepgrow model on single-gpu
python ./train.py
--input deepgrow/2D/MSD_Task09_Spleen/dataset.json \
--output models/2D \
--epochs 50
# Train a 2D Deepgrow model on multi-gpu (NVIDIA)
python -m torch.distributed.launch \
--nproc_per_node=`nvidia-smi -L | wc -l` \
--nnodes=1 \
--node_rank=0 \
--master_addr="localhost" \
--master_port=1234 \
-m train \
--multi_gpu true \
--input deepgrow/2D/MSD_Task09_Spleen/dataset.json \
--output models/2D \
--epochs 50
# After training to export/save as torch script model
python ./train.py
--input models/2D/model.pt \
--output models/2D/model.ts \
--export trueThis example is a standard PyTorch program and helps user to run evaluation for a trained 2D model.
# Run to know all possible options
python ./validate.py -h
# Evaluate a 2D Deepgrow model
python ./validate.py
--input deepgrow/2D/MSD_Task09_Spleen/dataset.json \
--output eval/2D \
--model_path models/2D/model.ptThis notebook helps to run pre-transforms before running inference over a Deepgrow 2D model. It also helps to run post-transforms to get the final label mask.
This is an extension for train.py that redefines basic default arguments to run 3D training.
# Run to know all possible options
python ./train_3d.py -h
# Train a 3D Deepgrow model on single-gpu
python ./train_3d.py
--input deepgrow/3D/MSD_Task09_Spleen/dataset.json \
--output models/3D \
--epochs 100
# Train a 3D Deepgrow model on multi-gpu (NVIDIA)
python -m torch.distributed.launch \
--nproc_per_node=`nvidia-smi -L | wc -l` \
--nnodes=1 \
--node_rank=0 \
--master_addr="localhost" \
--master_port=1234 \
-m train_3d \
--multi_gpu true \
--input deepgrow/3D/MSD_Task09_Spleen/dataset.json \
--output models/3D \
--epochs 100
# After training to export/save as torch script model
python ./train_3d.py
--input models/3D/model.pt \
--output models/3D/model.ts \
--export trueThis is an extension for validate.py that redefines basic default arguments to run 3D validation.
# Run to know all possible options
python ./validate_3d.py -h
# Evaluate a 3D Deepgrow model
python ./validate_3d.py
--input deepgrow/3D/MSD_Task09_Spleen/dataset.json \
--output eval/3D \
--model_path models/3D/model.ptThis notebook helps to run any pre-transforms before running inference over a Deepgrow 3D model. It also helps to run post-transforms to get the final label mask.
It contains basic ignite handler to capture region/organ-wise statistics, save snapshots, outputs while running train/validation over a dataset that has multi-label mask. By-default the handler is added as part of training/validation steps.
