search

juntang-zhuang / ShelfNet

implementation for paper "ShelfNet for fast semantic segmentation"
MIT License
252 stars 41 forks source link

编译so的环境 #5

Closed luuuyi closed 5 years ago

luuuyi commented 5 years ago

你好，感谢你的工作。能问一下你调用的enclib_cpu.so和enclib_gpu.lib的编译环境么，使用过程中出现importerror问题

juntang-zhuang commented 5 years ago

Thanks for your interest. I used Ubuntu 16.04, CUDA 9.0, PyTorch 0.4.1, gcc 5.4.0. Please check if running "python setup.py install" will solve the problem. Or refer to Pytorch-encoding for installation of enclib. But in this implementation we did not use their encoding function, we just used their over-all framework for image processing, data loader and synchronized-bn et al.

luuuyi commented 5 years ago

谢谢，现在问题解决了。多问一句，同步BN如果不使用的话还能有那么高的性能么。

juntang-zhuang commented 5 years ago

I'm not sure. From my experience, the accuracy is very close, but it may converge faster with syncbn. I have not done extensive experiments on this, and this is just my impression.