hzhuangdy

那是不是DBNet++比较不适合使用AMP？我普通训练了20epoch之后用这个模型去做重训开启AMP还是会有同样的问题

Thanks for reply I have set the batch_size to 4 and it still take a long time (at least 10m per IoU thresh)

I tried both version and they perform the same. My test data indeed has almost 30+ polygons per image, but the key is that it seems like the test process...

OK, I will try it again later. By the way, I use DBNet++ and I found that on 1.x branch the EastRandomCrop op is replaced by a normal RandomCrop op....

Thanks

Is there any method to add a progressbar(or logger) inside the evaluation? I upgrade to MMOCR1.x and run dist_test.sh. And it stuck at the following line. local loads checkpoint from...

For the normal situations: Elapsed time: 0.21742606163024902 gt_num:42, pred_num:42 Elapsed time: 0.20623540878295898 gt_num:28, pred_num:28 Elapsed time: 0.0869746208190918 gt_num:28, pred_num:28 Elapsed time: 0.11585068702697754 gt_num:28, pred_num:28 Elapsed time: 0.09441208839416504 gt_num:22, pred_num:22 Elapsed...

I'm sorry, my datasets are not publicly accessible. However, My datasets are all document images, so I think some public datasets like PubLayNet might help.

I also tried the code from Wenmuzhou(https://github.com/WenmuZhou/DBNet.pytorch), his evaluation runs fast. But I don't know the difference between his and MMOCR's

No, I didn't. ***@***.*** From: sudhitpanchal Date: 2023-10-16 16:05 To: clovaai/donut CC: hzhuangdy; Author Subject: Re: [clovaai/donut] Switch pretrained Decoder(for example, uses RoBERTa-xlm to replace mBART) (Issue #112) You got...