class remapping in cityscapes

chenxi116 / DeepLabv3.pytorch

PyTorch implementation of DeepLabv3

Other

254 stars 47 forks source link

class remapping in cityscapes #17

Closed xiaoaoran closed 4 years ago

xiaoaoran commented 4 years ago

Thank you for the repo.

I am using this repo to train cityscapes. But in your cityscapes.py, you defined 19 classes except for ignored 255, which is the num_class for the classifier. But label data do not remap to 19 classes but remain original 33, this will raise an error when calculating loss as

THCudaCheck FAIL file=/pytorch/aten/src/THC/THCReduceAll.cuh line=327 error=700 : an illegal memory access was encountered

chenxi116 commented 4 years ago

I thought if you follow the "Prepare Cityscapes Dataset" section in README, you will be working with the remap-to-19-classes label, so this error won't occur?

xiaoaoran commented 4 years ago

I just found it. Sorry for my ignorance. I'll close this issue.