pass
PATCH_SIZE = np.array([64, 64, 32])
# dataset
raw_dataset = BraTS2013(BRATS_PATH)
dataset = apply(CropToBrain(raw_dataset), load_image=partial(min_max_scale, axes=0))
train_dataset = cache_methods(ChangeSliceSpacing(dataset, new_slice_spacing=CONFIG['source_slice_spacing']))
# cross validation
split = load_json(SPLIT_PATH)
train_ids, val_ids, test_ids = split[int(FOLD)]
# batch iterator
batch_iter = Infinite(
load_by_random_id(train_dataset.load_image, train_dataset.load_gt, ids=train_ids),
unpack_args(tumor_sampling, patch_size=PATCH_SIZE, tumor_p=.5),
random_apply(.5, unpack_args(lambda image, gt: (np.flip(image, 1), np.flip(gt, 0)))),
batch_size=CONFIG['batch_size'], batches_per_epoch=CONFIG['batches_per_epoch']
)
# model
model = nn.Sequential(
nn.Conv3d(dataset.n_modalities, 8, kernel_size=3, padding=1),
layers.FPN(
layers.ResBlock3d, downsample=nn.MaxPool3d(2, ceil_mode=True), upsample=nn.Identity,
merge=lambda left, down: torch.cat(layers.interpolate_to_left(left, down, 'trilinear'), dim=1),
structure=[
[[8, 8, 8], [16, 8, 8]],
[[8, 16, 16], [32, 16, 8]],
[[16, 32, 32], [64, 32, 16]],
# dataset
raw_dataset = BinaryGT(BraTS2013(BRATS_PATH),
positive_classes=CONFIG['positive_classes'])
dataset = cache_methods(
ZooOfSpacings(apply(CropToBrain(raw_dataset),
load_image=partial(min_max_scale, axes=0)),
slice_spacings=CONFIG['slice_spacings']))
# cross validation
split = load_json(SPLIT_PATH)
train_ids, val_ids, test_ids = split[int(FOLD)]
# batch iterator
batch_iter = Infinite(
load_by_random_id(dataset.load_image,
dataset.load_spacing,
dataset.load_gt,
ids=train_ids),
unpack_args(get_random_slices, n_slices=CONFIG['n_slices']),
random_apply(
.5,
unpack_args(lambda image, spacing, gt:
(np.flip(image, 1), spacing, np.flip(gt, 0)))),
apply_at(2, prepend_dims),
apply_at([1, 2], np.float32),
batch_size=CONFIG['batch_size'],
batches_per_epoch=CONFIG['batches_per_epoch'],
combiner=combine_pad)
# model
unet_2d = nn.Sequential(
nn.Conv2d(dataset.n_modalities, 8, kernel_size=3, padding=1),