The pipeline to do training and inference on segmentation data.

All you need before is that pytorch with the gpu support is installed. Afterwards you can clone the repository and install the package using pip.

pip install -e dl_pipeline

To read more about the kaggle competition, please read: https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation

The solution itself is described in the topic: https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/107687

The whole pipeline consist of segmentation and classification models.

For this all you need is the package itself: dl_pipeline and everything from the classification folder: script that runs the training and hyperparameters.

Example how to train the first fold of se_resnext101 classifier:

python classification/run.py classification/hyperparameters/hyperparameters_res101_0.json

Example how to train the first fold of senet154 classifier:

python classification/run.py classification/hyperparameters/hyperparameters_senet154_0.json

After running all training scripts for all folds, you will have all necessery classification models.

For this all you need is the package itself: dl_pipeline and everything from the segmentation folder: script that runs the training and hyperparameters.

Example how to train the first fold of dpn98 segmetation:

python segmentation/run.py segmentation/hyperparameters/hyperparameters_dpn98_512_0.json
python segmentation/run.py segmentation/hyperparameters/hyperparameters_dpn98_1024_0.json

Then you have to repeat this procedure 15 times: for each of three models dense121, dpn98 and resnet101 and each of 5 folds.

Afterwards you will be able to use this models to do th inference.

To do the inference please run the notebook submission/prepare_submission.ipynb with the specified models, that are trained in the previous step