Closed pawopawo closed 4 years ago
The number of the last block channel of mobilenetv3 in the paper is reduced by a factor of 2, and atrous conv is used
If you don't want to use Docker you may locally install Python>=3.6, NVIDIA driver >= 418, CUDA >= 10.1, tensorflow-gpu>=2.0.0b1 and other packages from requirements.txt and Docker file.
Regarding the second message, actually it's quite vague in the paper but according to the figure 10 they don't actually use atrous convolutions, they propose Lite-RASPP module which exploits 1x1 convolutions and average pooling with big kernel and stride parameters. What about the number of channels you can easily change it in proposed code to get full consistence with the paper, but highly likely it won't affect accuracy and performance so much.
Thanks for your nice work ! How to allow programs in Ubuntu environment, not docker, cannot allow
make build