benoitkoenig / pneumothorax-segmentation Goto Github PK

This repository is my solution to this Kaggle competition

Author: Benoît Koenig

• Run the whole training process from scratch: whole_training_from_scratch.sh

• Training the segmentation model: python -m segmentation.train

• Training the hydra classifier's body: python -m hydra_classifier.train_hydra_body [resnet50|densenet169]

• Training the hydra classifier's heads: python -m hydra_classifier.train_hydra_head [resnet50|densenet169] [none|resize|flip_rotate|filter]

• Evaluating segmentation: python -m evaluate.segmentation

• Evaluating classification: python -m evaluate.classification

• Exporting final predictions to results.csv: python -m export_predictions

• Check data augmentation: python -m visualization.show_data_augment [index]

• Visualize prediction: python -m visualization.show_prediction [train|test] [index]

• Show raw data from dicom: python -m visualization.show_raw_data [train|test] [index]

• Show ground truth mask: python -m visualization.show_true_mask [index]

tensorflow-gpu is not included in requirements.txt as it is not always relevant. To use the GPU, install tensorflow-gpu via "pip install tensorflow-gpu" Currently, the heads of the hydra_classifier are not used for predictions due to a bug. To investigate