search

open-mmlab / mmocr

OpenMMLab Text Detection, Recognition and Understanding Toolbox
https://mmocr.readthedocs.io/en/dev-1.x/
Apache License 2.0
4.3k stars 746 forks source link

robust scanner fusion question #184

Open Goghzki opened 3 years ago

Goghzki commented 3 years ago

hi~, I think the robust_scanner_fusion_layer.py is not the original method in is's parper https://arxiv.org/pdf/2007.07542.pdf In git: output = self.linear_layer(fusion_input) output = self.glu_layer(output) In parper : -> output1 = self.linear_layer(fusion_input) output1 = torch.sigmoid(output1) output2 = self.linear_layer(fusion_input) output = torch.mul(output1 , output2 )

Actually In my sentence level dataset, I found many alignment drift case, which should be fixed by robust_scanner_fusion_layer. I don't know is this right?

innerlee commented 3 years ago

@cuhk-hbsun Pytorch has an implementation of glu, maybe we can use it.

Goghzki commented 3 years ago

image mabey it is different? torch split the input in to two parts a and b, and use sigmoid function only to b(not both) , then mul a and sigmoid b