Make LCPAN also pass the activator to DepthwiseSeparable #9280
Conversation
Without this, nn.Hardswish() will be selected in ConvBNLayer, which causes the opset version to be bumped to 14, in turn breaking a ESP-DL onnx quantization, which only supports 13. I would also suspect that this is more correct also in other cases as one would not be likely to not want to use a specific activation function everywhere, would one supply it. Also, it aligns with how the act-parameter is handled in other cases.
|
Thanks for your contribution! |
|
|
There was a problem hiding this comment.
These changes effectively allow passing the configured activation function. Without them, setting act=relu6 in LCPAN still results in a hardswish activation function in the exported ONNX.
I successfully exported an ONNX network with opset 11 for ESP-DL (which does not support hardswish activation and opset 14) using these changes.
|
@LokeZhou @jzhang533 |
|
@LokeZhou @jzhang533 isn't this an obvious bug? I have to correct this each time i use a new version of PaddleDetection? |
I agree |
|
@zhangyubo0722 , @SigureMo , can you please review this? @simoberny Sorry, accidentally requested a review from you. |
4 participants
Without this, nn.Hardswish() will be selected in ConvBNLayer, which causes the opset version to be bumped to 14, in turn, and in my case, breaks ESP-DL ONNX quantization, which only supports 13.
I would also suspect that this is more correct also in other cases as one would not be likely to not want to use a specific activation function everywhere, would one supply it.
Also, it aligns with how the act-parameter is handled in other cases.
(accidentally first did the PR against a release branch)