How can i train encnet in one GPU?

[neverstoplearn]

[zhanghang1989]

Similar issues: https://github.com/zhanghang1989/PyTorch-Encoding/issues/205 https://github.com/zhanghang1989/PyTorch-Encoding/issues/210#issuecomment-578752998

The code was written for multiple GPUs. If you would like run on single GPU, the following part need to change: 1). change SyncBatchNorm to regular nn.BatchNorm 2). The network output shape may need some work, which is a tuple of list of tensors (ngpus x noutputs).

[neverstoplearn]

thanks for your answer. i will try it.

[SeanCho1996]

Hi there, did you solve the problem with the output shape? Cuz I've encounterer the same error. Thank U!

[StacyYang]

checked with the author @zhanghang1989 , the next version should be released in 2~3 weeks, which will support single gpu training.

[zhanghang1989]

Using https://github.com/zhanghang1989/PyTorch-Encoding/blob/master/experiments/segmentation/train_dist.py should solve this problem. I am close it. Feel free to reopen if error exists.

[wanghao9610]

Using /experiments/segmentation/train_dist.py@master should solve this problem. I am close it. Feel free to reopen if error exists.

Can you give an example of train_dist.py?

[zhanghang1989]

The cmd is actually the same:

python train_dist.py --dataset ADE20K --model fcn  --aux --backbone resnest50

[wanghao9610]

The cmd is actually the same:

python train_dist.py --dataset ADE20K --model fcn  --aux --backbone resnest50

I have tried, but here came an error, how to fix this? _pickle.PicklingError: Can't pickle <function main_worker at 0x7f104ff1b700>: attribute lookup main_worker on main failed

[zhanghang1989]

That's wired. Are you using PyTorch 1.4.0?

[wanghao9610]

That's wired. Are you using PyTorch 1.4.0? Yes, pytorch version is 1.4.0, encoding version is 1.2.1b20200527. Is it not compatible with pytorch1.4.0?

[zhanghang1989]

Yes, it should be compatible with pytorch 1.4.0. I am using that version.