Skip to content
GitLab
Explore
Sign in
Primary navigation
Search or go to…
Project
bob.learn.tensorflow
Manage
Activity
Members
Labels
Plan
Issues
Issue boards
Milestones
Code
Merge requests
Repository
Branches
Commits
Tags
Repository graph
Compare revisions
Build
Pipelines
Jobs
Pipeline schedules
Artifacts
Deploy
Releases
Model registry
Operate
Environments
Monitor
Incidents
Analyze
Value stream analytics
Contributor analytics
CI/CD analytics
Repository analytics
Model experiments
Help
Help
Support
GitLab documentation
Compare GitLab plans
Community forum
Contribute to GitLab
Provide feedback
Keyboard shortcuts
?
Snippets
Groups
Projects
This is an archived project. Repository and other project resources are read-only.
Show more breadcrumbs
bob
bob.learn.tensorflow
Commits
9d279cac
Commit
9d279cac
authored
6 years ago
by
Amir MOHAMMADI
Browse files
Options
Downloads
Patches
Plain Diff
Add a pixel wise loss
parent
1a2392d8
Branches
Branches containing commit
Tags
Tags containing commit
1 merge request
!79
Add keras-based models, add pixel-wise loss, other improvements
Changes
1
Hide whitespace changes
Inline
Side-by-side
Showing
1 changed file
bob/learn/tensorflow/loss/pixel_wise.py
+63
-0
63 additions, 0 deletions
bob/learn/tensorflow/loss/pixel_wise.py
with
63 additions
and
0 deletions
bob/learn/tensorflow/loss/pixel_wise.py
0 → 100644
+
63
−
0
View file @
9d279cac
from
..dataset
import
tf_repeat
from
.utils
import
(
balanced_softmax_cross_entropy_loss_weights
,
balanced_sigmoid_cross_entropy_loss_weights
,
)
import
tensorflow
as
tf
class
PixelWise
:
"""
A pixel wise loss which is just a cross entropy loss but applied to all pixels
"""
def
__init__
(
self
,
balance_weights
=
True
,
n_one_hot_labels
=
None
,
label_smoothing
=
0.5
,
**
kwargs
):
super
(
PixelWise
,
self
).
__init__
(
**
kwargs
)
self
.
balance_weights
=
balance_weights
self
.
n_one_hot_labels
=
n_one_hot_labels
self
.
label_smoothing
=
label_smoothing
def
__call__
(
self
,
labels
,
logits
):
with
tf
.
name_scope
(
"
PixelWiseLoss
"
):
flatten
=
tf
.
keras
.
layers
.
Flatten
()
logits
=
flatten
(
logits
)
n_pixels
=
logits
.
get_shape
()[
-
1
]
weights
=
1.0
if
self
.
balance_weights
and
self
.
n_one_hot_labels
:
# use labels to figure out the required loss
weights
=
balanced_softmax_cross_entropy_loss_weights
(
labels
,
dtype
=
logits
.
dtype
)
# repeat weights for all pixels
weights
=
tf_repeat
(
weights
[:,
None
],
[
1
,
n_pixels
])
weights
=
tf
.
reshape
(
weights
,
(
-
1
,))
elif
self
.
balance_weights
and
not
self
.
n_one_hot_labels
:
# use labels to figure out the required loss
weights
=
balanced_sigmoid_cross_entropy_loss_weights
(
labels
,
dtype
=
logits
.
dtype
)
# repeat weights for all pixels
weights
=
tf_repeat
(
weights
[:,
None
],
[
1
,
n_pixels
])
if
self
.
n_one_hot_labels
:
labels
=
tf_repeat
(
labels
,
[
n_pixels
,
1
])
labels
=
tf
.
reshape
(
labels
,
(
-
1
,
self
.
n_one_hot_labels
))
# reshape logits too as softmax_cross_entropy is buggy and cannot really
# handle higher dimensions
logits
=
tf
.
reshape
(
logits
,
(
-
1
,
self
.
n_one_hot_labels
))
loss_fn
=
tf
.
losses
.
softmax_cross_entropy
else
:
labels
=
tf
.
reshape
(
labels
,
(
-
1
,
1
))
labels
=
tf_repeat
(
labels
,
[
n_pixels
,
1
])
labels
=
tf
.
reshape
(
labels
,
(
-
1
,
n_pixels
))
loss_fn
=
tf
.
losses
.
sigmoid_cross_entropy
loss_pixel_wise
=
loss_fn
(
labels
,
logits
=
logits
,
weights
=
weights
,
label_smoothing
=
self
.
label_smoothing
,
reduction
=
tf
.
losses
.
Reduction
.
MEAN
,
)
tf
.
summary
.
scalar
(
"
loss_pixel_wise
"
,
loss_pixel_wise
)
return
loss_pixel_wise
This diff is collapsed.
Click to expand it.
Preview
0%
Loading
Try again
or
attach a new file
.
Cancel
You are about to add
0
people
to the discussion. Proceed with caution.
Finish editing this message first!
Save comment
Cancel
Please
register
or
sign in
to comment