Skip to content

limit crop_to_factor #5

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
abred wants to merge 3 commits into
base: master
Choose a base branch
from
Draft

limit crop_to_factor #5

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
23d00ab
crop_to_factor only on last level
abred Dec 16, 2020
91f287f
only crop_to_factor if not training
abred Dec 16, 2020
a0e30b0
ensure output size > crop_factor
abred Dec 16, 2020
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
18 changes: 12 additions & 6 deletions funlib/learn/torch/models/unet.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -210,7 +210,7 @@ def forward(self, f_left, g_out):

g_up = self.up(g_out)

if self.next_conv_kernel_sizes is not None:
if not self.training and self.next_conv_kernel_sizes is not None:
g_cropped = self.crop_to_factor(
g_up,
self.crop_factor,
Expand Down Expand Up @@ -350,7 +350,6 @@ def __init__(
kernel_size_up = [[(3, 3, 3), (3, 3, 3)]]*(self.num_levels - 1)

# compute crop factors for translation equivariance
crop_factors = []
factor_product = None
for factor in downsample_factors[::-1]:
if factor_product is None:
Expand All @@ -359,8 +358,7 @@ def __init__(
factor_product = list(
f*ff
for f, ff in zip(factor, factor_product))
crop_factors.append(factor_product)
crop_factors = crop_factors[::-1]
self.crop_factor = factor_product

# modules

Expand Down Expand Up @@ -392,8 +390,11 @@ def __init__(
mode='nearest' if constant_upsample else 'transposed_conv',
in_channels=num_fmaps*fmap_inc_factor**(level + 1),
out_channels=num_fmaps*fmap_inc_factor**(level + 1),
crop_factor=crop_factors[level],
next_conv_kernel_sizes=kernel_size_up[level])
crop_factor=(self.crop_factor if level == self.num_levels - 2
else None),
next_conv_kernel_sizes=(kernel_size_up[level]
if level == self.num_levels - 2
else None))
for level in range(self.num_levels - 1)
])
for _ in range(num_heads)
Expand Down Expand Up @@ -455,6 +456,11 @@ def forward(self, x):

y = self.rec_forward(self.num_levels - 1, x)

if self.training:
out_spatial_shape = y.size()[-self.dims:]
assert out_spatial_shape > self.crop_factor, \
"To avoid tile-and-stitch inconsistencies, the output size during training has to be strictly larger than prod(downsample_factors) (output {}, prod(downsample_factors) {} (= crop_factor))".format(out_spatial_shape, self.crop_factor)

if self.num_heads == 1:
return y[0]

Expand Down