README.md

Train and test on CVPPP dataset

In this example we use zero_padding convolutional UNet, that means that the coloring procedure will use the knowledge about borders to do coloring.

Run training

python train_cvppp.py path_to_A1_dataset

Notes on code

Creates a batch generator

generator = train_data.create_batch_generator(30, transforms=transforms)

---

Creates a halo region function

mask_builder = dc.build_halo_mask(fixed_depth=30, margin=21, min_fragment=10)

1. fixed_depth - maximum number of object in a training batch
2. margin - size of margin (dilatation) around the object sould be odd
3. min_fragment - minimal size of an object in pixels

---

Training

model, errors = dc.train(generator=generator,
                             model=net,
                             mask_builder=mask_builder,
                             niter=10000,
                             k_neg=7.,
                             lr=1e-3,                             
                             caption=join(directory, "model"))

1. generator - batch generator
2. model - segmentation network
3. niter - number of iterations
4. k_neg - balance between positive and negative parts of loss please seen paper
5. lr - learining rate
6. caption - name of errors file and model