get_steps_per_epoch fails if block_shape is None

[hvgazula]

The get_steps_per_epoch method of the nobrainer.dataset.Dataset class assumes blocks are created from the input images. This method needs refactoring to handle the block_shape = None case.

nobrainer/nobrainer/dataset.py

Lines 249 to 266 in ed0d609

	def get_steps_per_epoch(self):
	def get_n(a, k):
	return (a - k) / k + 1
	n_blocks = tuple(
	get_n(aa, kk) for aa, kk in zip(self.volume_shape, self.block_shape)
	)
	for n in n_blocks:
	if not n.is_integer() or n < 1:
	raise ValueError(
	"cannot create non-overlapping blocks with the given parameters."
	)
	n_blocks_per_volume = np.prod(n_blocks).astype(int)
	steps = n_blocks_per_volume * self.n_volumes / self.batch_size
	steps = math.ceil(steps)
	return steps

[hvgazula]

won't fail. all good.

nobrainer/nobrainer/dataset.py

Lines 241 to 243 in ed0d609

	@property
	def block_shape(self):
	return tuple(self.dataset.element_spec[0].shape[1:4].as_list())

[hvgazula]

fails if n_volumes is not set which is the case when the from_tfrecords function is called.

[hvgazula]

this line

nobrainer/nobrainer/dataset.py

Line 122 in ed0d609

block_length = len([0 for _ in first_shard])

is a huge bottleneck in the code. For context, please refer to https://stackoverflow.com/questions/70992022/how-to-get-the-correct-cardinality-of-a-tensorflow-dataset-after-filtering

[hvgazula]

this is resolved by calculating the number of files within each shard. This case doesn't account for the situation where the last shard may have fewer files but that shouldn't impact the training process in any way.