Deep learning CNN toolkit

This toolkit was used for the Kaggle Ultrasound nerve segmentation challenge.

Instructions

Data

The training data need to be stored in a 2D image format (e.g. png, tif) inside a train directory.

• For segmentation the train directory has to contain a mask for each image:
  if the image is stored as train/image1.tif the mask will be train/image1_mask.tif

• For autoencoders only the images will be inside the train directory (no masks).

The testing data are stored as images inside a test directory (no masks).

Training

Training is performed with the train.py with parameters:

-v [version] :	version of the experiment to be run (see below for configuration of the experiment)
-train [dir] :	train directory (default: train)
-cv [num_folds] :	cross-validation with a number of folds (default: 10)
-fold [fold_nr] :	run for a specific fold of the cross-validation
-seed [seed_nr] :	change the seed used for the cross-validation (default: 1234)

Testing

Testing is performed with the test.py. Test.py has the same parameters as above (apart from -train) and additionally

-test [dir] :	test directory (default: test)
-results [dir] :	results directory (default: submit)

Additional parameters can be viewed by running the test.py with the --help option.

Configuration

The config_default.py contains the default parameters used for training.
This can be overriden for the different experiments if required.
In order to setup a new experiment the user needs to:

• create a directory with the experiment name inside the "params" folder (e.g. params/unet-small)
• write a config.py file inside the experiment directory (e.g. params/unet-small/config.py) with the parameters he wants overriden (e.g. depth=4). Parameters that are not specified retain their original values from the config_default.py
• run the training/testing specifying the experiment as the version (e.g. python train.py -v unet-small )