assert numpy.array([img[:-7] == lbl[:-9] for img, lbl in zip(rgb, labels)]).all() == True
numpy.random.seed(1203412412)
indexes = numpy.random.permutation(len(rgb))
perm_rgb = numpy.array(rgb)[indexes].tolist()
perm_labels = numpy.array(labels)[indexes].tolist()
train_data = Reader(perm_rgb[:-10], perm_labels[:-10], 2)
valid_data = Reader(perm_rgb[-10:], perm_labels[-10:], 2)
transforms = [dc.rgba2rgb(),
dc.clip_patch((192, 192)),
dc.flip_horizontally(),
dc.flip_vertically(),
dc.rotate90(),
dc.random_transform(0.1, 90, 0),
dc.blur(),
dc.normalize(0.5, 0.5)]
generator = train_data.create_batch_generator(30, transforms=transforms)
mask_builder = dc.build_halo_mask()
net = dc.EUnet(3, 9, 4, 3, 1, depth=3, padding=1, init_xavier=True, use_bn=False, use_dropout=True).to(device)
model, errors = dc.train(generator=generator,
model=net,
mask_builder=mask_builder,
niter=10000,
caption=join(directory, "model"))
print(evaluate(valid_data, net, 65))
numpy.random.seed(1203412412)
indexes = numpy.random.permutation(len(rgb))
perm_rgb = numpy.array(rgb)[indexes].tolist()
perm_labels = numpy.array(labels)[indexes].tolist()
train_data = dc.Reader(perm_rgb[:-10], perm_labels[:-10], 1)
valid_data = dc.Reader(perm_rgb[-10:], perm_labels[-10:], 1)
transforms = [
dc.clip_patch((256, 256)),
dc.flip_horizontally(),
dc.flip_vertically(),
dc.rotate90(),
dc.rgb2gray(),
dc.normalize(0.5, 0.5)
]
generator = train_data.create_batch_generator(30, transforms=transforms)
mask_builder = clip_mask_builder(
dc.build_halo_mask(fixed_depth=100, margin=35, min_fragment=10))
net = dc.EUnet(1, 6, k=2, s=4, depth=3, l=1, padding=0,
use_dropout=True).to(device)
model, errors = dc.train(generator=generator,
model=net,
mask_builder=mask_builder,
niter=7000,
k_neg=5.,
lr=1e-3,
caption=join(directory, "model"))