Jianquan Ye

> It is impossible to perform `gather` operation in cpu because the operation is based on NCCL backend. But there is a way to avoid gathering in GPU on-the-fly, that...

I have met this situation, too

> The test protocol of lfw is 10-fold cross-validation. It has 6000 pairs of face images and has been split into 10 parts. OK,Thank you for your answering.

May I ask you one more question? Is it unnecessary to use margin.ckpt in evaluation? Margin only help feature learing in training,is it right?

I change my code use tta.transforms instead of TTAWrapper. Then I can do normalization to the tensors(make it to [0,1]), and that could avoid the negative value in tensors. So...

> Hi authors, I just came into this project and found it written well! I wonder if you would like to raise a PR to MMOCR so that more researchers...

> Hi, are you still interested in raising a PR? If you need any assistance from us (and the community), just mail to [email protected] and I’ll get you in our...

> I want use the same datasets ,but I don't know where to load it. Thanks. Due to the company’s privacy policy, the PSENet training data labeled in the intranet...

> i want to only test the acc of Structure recognition accuracy and use mmocr_teds_acc_mp.py, then i use pretrain model name as you supply epoch_16_0.7767.pth, to get only table structure...

> Yes, you're right. It's doesn't go wrong because of 'self.cls_layer' length is 1. Your solution is more reasonable.