How to train the model on single GPU?

CoinCheung / DeepLab-v3-plus-cityscapes

mIOU=80.02 on cityscapes. My implementation of deeplabv3+ (also know as 'Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation' based on the dataset of cityscapes).

MIT License

174 stars 27 forks source link

How to train the model on single GPU? #2

Closed zhixuanli closed 5 years ago

zhixuanli commented 5 years ago

This ninja error continuously occurs:

Command '['ninja', '-v']' returned non-zero exit status 1

As the instruction of the inplace_abn author: https://github.com/mapillary/inplace_abn/issues/106

I have to change the GCC version to 7.3.

But I'm not the root of the server, so I could not install any new GCC.

How can I train your code on a single GPU? I have a GPU on my own machine.

Thank you so much for your neat codes!

CoinCheung commented 5 years ago

Hi, I dont think you can train this code on single gpu, since I used batch size of 8 on two 1080ti gpus. If you train on one gpu, you will have to endure a batch size of 4 which will significantly harm the performance of BN based models. Therefore, I do not recommend you to try it.

zhixuanli commented 5 years ago

Thank you!