Feature: DISTS metric #131
Conversation
Codecov Report
@@ Coverage Diff @@
## master #131 +/- ##
==========================================
+ Coverage 90.49% 91.00% +0.50%
==========================================
Files 21 22 +1
Lines 1505 1556 +51
==========================================
+ Hits 1362 1416 +54
+ Misses 143 140 -3
|
|
|
||
| class DISTS(ContentLoss): | ||
| r"""Deep Image Structure and Texture Similarity metric. | ||
| Expects input to be in range [0, 1] or normalized with ImageNet statistics into range [-1, 1] |
There was a problem hiding this comment.
Do we really support [-1, 1]? Let's add images with [-1, 1] dynamic range to the tests.
There was a problem hiding this comment.
Yes, but forward pass fails because of _validate_input function :\
piq/perceptual.py
Outdated
| def construct_kernel(self, x: torch.Tensor) -> torch.Tensor: | ||
| r"""Returns 2D Hann window kernel with number of channels equal to input channels""" | ||
| C = x.size(1) | ||
|
|
||
| # Take bigger window and drop borders | ||
| window = torch.hann_window(self.kernel_size + 2, periodic=False)[1:-1] | ||
| kernel = window[:, None] * window[None, :] | ||
|
|
||
| # Normalize and reshape kernel | ||
| self.kernel = (kernel / kernel.sum()).repeat((C, 1, 1, 1)).to(x) |
There was a problem hiding this comment.
Since we added functional.py into the project, which contains some filters, I would propose to move to there.
There was a problem hiding this comment.
I propose the following:
- Make a
utilsfolder - Put the
functional.pyhere "as is" - Rename
utils.py->utils/common.py - Create a new file
utils/layers.py - Add
L2Pool2dhere
I also think that this refactoring is the subject of another PR. Current structure will work for the current PR.
There was a problem hiding this comment.
I did this refactoring in a single commit. It's not so big to require a separate PR
piq/perceptual.py
Outdated
| def construct_kernel(self, x: torch.Tensor) -> torch.Tensor: | ||
| r"""Returns 2D Hann window kernel with number of channels equal to input channels""" | ||
| C = x.size(1) | ||
|
|
||
| # Take bigger window and drop borders | ||
| window = torch.hann_window(self.kernel_size + 2, periodic=False)[1:-1] | ||
| kernel = window[:, None] * window[None, :] | ||
|
|
||
| # Normalize and reshape kernel | ||
| self.kernel = (kernel / kernel.sum()).repeat((C, 1, 1, 1)).to(x) |
There was a problem hiding this comment.
I propose the following:
- Make a
utilsfolder - Put the
functional.pyhere "as is" - Rename
utils.py->utils/common.py - Create a new file
utils/layers.py - Add
L2Pool2dhere
I also think that this refactoring is the subject of another PR. Current structure will work for the current PR.
|
Also closed #132 in one of the commits |
zakajd
force-pushed
the
feature/issue#123
branch
from
bab2b6d
to
4022043
Compare
|
Kudos, SonarCloud Quality Gate passed!
|
|
@snk4tr Ready to be merged |
Closes #123
Proposed Changes
use_average_poolingtoreplace_poolingbecause in DISTS pooling is replaced by L2pooling layer, not AveragePool.__init__descriptions to class description, so that it's visible in VS Code and Jupyter IDA