-
Notifications
You must be signed in to change notification settings - Fork 426
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.
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
PERMANENT DRAFT: TF grouped convolutions check #1183
Conversation
ok SeperableConv2D and DepthwiseConv2D does not support different row/col stride values currently which would break the rectangular pooling implementations 😩
|
Maybe we can keep this as Draft and check if the next TF release brings a fix |
Thanks for the suggestion Felix! Yes as you mentioned, I think this is a risky move 😅 |
Yeah lets keep this Draft and i will try to check after each TF release if it is maybe fixed 😅 |
cce9376
to
fc759fd
Compare
fc759fd
to
b06df1a
Compare
b06df1a
to
177d28e
Compare
177d28e
to
68f12d4
Compare
Codecov Report
@@ Coverage Diff @@
## main #1183 +/- ##
==========================================
- Coverage 95.78% 95.76% -0.02%
==========================================
Files 154 154
Lines 6903 6903
==========================================
- Hits 6612 6611 -1
- Misses 291 292 +1
Flags with carried forward coverage won't be shown. Click here to find out more. |
This PR:
tensorflow.python.framework.errors_impl.InvalidArgumentError: {{function_node __wrapped__Conv2DBackpropInput_device_/job:localhost/replica:0/task:0/device:CPU:0}} Gradients for grouped convolutions are not supported on CPU. Please file a feature request if you run into this issue. Computed input depth 576 doesn't match filter input depth 1 [Op:Conv2DBackpropInput]
The big disadvantage there would be, that we would need to retrain all models which uses mobilenet as backbone and the classification models itself
The export for TF mobilenet from #1182 still works with the changes from this PR
@frgfm What do you think makes it sense in that case of retraining all the stuff to fix this issue ?
And the more related question @olivmindee @charlesmindee
Could you retrain the models ? Especially the rotation classification model and the detection/recogition ones depends on your datasets (computation power would not be a problem on my side but the data is 😅 )