search

HRNet / HRNet-Semantic-Segmentation

The OCR approach is rephrased as Segmentation Transformer: https://arxiv.org/abs/1909.11065. This is an official implementation of semantic segmentation for HRNet. https://arxiv.org/abs/1908.07919
Other
3.09k stars 683 forks source link

InPlaceABNSync error #3

Closed nemonameless closed 4 years ago

nemonameless commented 5 years ago

hi ,the InPlaceABNSync seems not working, the training process was stuck by https://github.com/HRNet/HRNet-Semantic-Segmentation/blob/master/lib/models/seg_hrnet.py#L269

only if I change InPlaceABNSync into nn.BatchNorm2d, then I can train without Obstruction.

sunke123 commented 5 years ago

Can you provide the error info and related environment info? I can run this code successfully on pytorch 0.4.1. If you use the pytorch 1.0, you need to reinstall the sync-bn (https://github.com/mapillary/inplace_abn).

nemonameless commented 5 years ago

@sunke123 I also run on pytorch 0.4.1, it's just stuck at the InPlaceABNSync without any error log. However I can run test code using you InPlaceABNSync but got 56.04 in LIP, a little lower than yours

sunke123 commented 5 years ago

The model achieves 56.04 on LIP as reported. What‘ s the meaning of "a little lower than yours"? I will double check the InPlaceABNSync. I can run this code successfully. Could you try my docker? You can pull my docker from dockerhub, kesun/pose:pytorch-0.4.1-bn-cudnn-off.

nemonameless commented 5 years ago

@sunke123 Sorry for misreading 56.04. Thanks for your docker, I will try later and also check the InPlaceABNSync

xiaohuihui52309 commented 5 years ago

Install CUDA 9.2, I solve the problem.

ngunauj commented 5 years ago

Install CUDA 9.2, I solve the problem.

hi ,the InPlaceABNSync seems not working, the training process was stuck by https://github.com/HRNet/HRNet-Semantic-Segmentation/blob/master/lib/models/seg_hrnet.py#L269

only if I change InPlaceABNSync into nn.BatchNorm2d, then I can train without Obstruction.

i have the same problem. how to solve?

superxudou commented 4 years ago

i use the pytorch 1.0 and reinstall the sync-bn (https://github.com/mapillary/inplace_abn).,i get segment default .my god