git clone https://github.com/jinkos/nuke_kaggle

will create a nuke_kaggle directory with all the files in it.

edit the config file and add and entry for your machine_name

python3 image_wrapper.py your_machine_name

Create the aug files.

Used to take 10 hours but I have added multiprocessing. My LINUX box manages it in 2 hours, now. I had to use pillow to save the .png files because skimage wasn't working. So you need to install pillow

python3 train.py your_machine_name --save

should train the model and save the weights

python3 train.py your_machine_name --load --save

will load the weights and then update them

python3 submit.py your_machine_name will enable you to analyse your results or make a submission depending on how you have set the flags in the code.

arg_config.cfg - set up a section for your own machine and tell it where to find your data files

arg_config.py - decode the .cfg file and parse the command line

image_wrapper.py - generates all the augmentation files

metrics.py - everything to do with goddamn intersection over union (I don't understand the half of it!)

nuclear_gen.py - data generator functions

our_models.py - to keep track of what all the models are and what their weight files are called - trust me we'll need it

submit.py - makes a submission or you can set it up to analyse the results of your model on the validation set (lots of fun graphs)

unet.py - the unet model