TGS Salt Segmentation competition

This is a project containing source code to the TGS image segmentation competition.

(!I misspelled TGS, and only realized two months after. You will see TSG at many places in the code and in file names.)

Installing

To run training on multiple GPUs, first follow the guides to install Horovod. Horovod allows you to run training in a distributed ring-allreduce fashion. Here is a short summary, about ring-allreduce and other distributed architectures.

Stratified k-fold cross validation

The script below will split the dataset into k groups, and create symlinks in the 'folds' folder. Stratification by coverage is used to ensure each group has a good representation of the whole dataset.

To experiment with other qualities to stratify by, you can extend the TGSDatasetPreprocessor class. Stratified k-fold cross validation helps to better estimate the performance of the trained model on the unseen test set.

python tgs_preprocess.py

Further experiments: Coverage class may not be the best quality to stratify by and create a good representation of the overall dataset. It may ignore other important features.

Run in distributed mode

(!Under construction)

Horovod uses MPI - Message Passing Interface under the hood, to communicate between GPUs. A basic intro can be found here.

mpirun -np 3 -bind-to none -map-by slot -x NCCL_DEBUG=INFO -x LD_LIBRARY_PATH -x PATH -mca pml ob1 -mca btl ^openib /home/pkovacs/anaconda3/envs/exp/bin/python tsg_train_ring.py --hvd=True --data_path=folds/fold_0 --model=resnet34 --batch_size=16 --epochs=90

or alternatively run the bash script below: (set the path of your python env in the script file)

./train_multi_gpu_ring.sh {number_of_gpus}

Run in single-gpu mode

python tgs_train.py

Generate predictions

Submission file will be saved in the submission folder in a 'submission_weightFileName_dateTime.csv' file

python tgs_predict.py --file_name=path_to_the_model

To generate a weighted average prediction of folds, run:

python tgs_predict.py --use_folds=True

This picks up all the weights from the 'weights/model_name/' folder and generate an avaraged prediction.

License

This project is licensed under the MIT License - see the LICENSE file for details