Add support for multi-channel images in PyTorchObjectDetector #1633
Conversation
Signed-off-by: Beat Buesser <beat.buesser@ie.ibm.com>
Codecov Report
@@ Coverage Diff @@
## dev_1.10.1 #1633 +/- ##
===========================================
Coverage 88.09% 88.10%
===========================================
Files 259 259
Lines 21328 21328
Branches 3779 3779
===========================================
+ Hits 18788 18790 +2
+ Misses 1597 1596 -1
+ Partials 943 942 -1
|
| @@ -214,7 +214,7 @@ def _get_losses( | |||
| x_grad.requires_grad = True | |||
| else: | |||
| x_grad = x[i].to(self.device) | |||
| if x_grad.shape[-1] in [1, 3]: | |||
| if x_grad.shape[2] < x_grad.shape[0] and x_grad.shape[2] < x_grad.shape[1]: | |||
There was a problem hiding this comment.
Could we keep the channels_first variable to allow manual configuration of the channels, and if it is not supplied, infer the channel number via this logic? There could be unusual cases in which the channels are more than one of the height/width dimensions of the input.
There was a problem hiding this comment.
I agree. For this patch release we decided to keep the channels_first unused, but update the documentation accordingly, with the goal not to break the estimator. I agree, there could be images with more channels than pixels in height and width. All the models supported by this estimator are expecting channels first, therefore we should update the channels_First argument` for ART 1.11, I have opened issue #1637.
Signed-off-by: Beat Buesser beat.buesser@ie.ibm.com
Description
This pull request adds support for multi-channel images in
PyTorchObjectDetector.Fixes #1630
Type of change
Please check all relevant options.