due to NCCL requirement, will not work on Windows about idinvert HOT 6 CLOSED

The NCCL is the basis for running this code on multi-GPUs, you could run the code on a single GPU by removing it, but I am not sure whether the performance on a single GPU could reach the performance on multi-GPUs.

from idinvert.

Thank you, and understood. In a practical sense, I assume I could just remove this then, and not replace it with anything? (if using a single GPU):

from idinvert.

Actually, if you use a single GPU, the code can not reach here. By the way, if you use a single GPU to training the encoder, you need to replace the sync_batch_norm in networks_encoder.py with conventional BatchNorm provided by TensorFlow.

from idinvert.

Thanks for the advice so far, looks like the import in optimizer.py wasn't the issue

In fact, I've removed all references to nccl as found here:
https://github.com/genforce/idinvert/search?q=nccl

yet still continue to get nccl related errors

It seems to be from here, but I'm not sure how to trace the issue further
https://github.com/genforce/idinvert/blob/master/dnnlib/tflib/network.py#L291

from idinvert.

The .pkl files contain the network's code I used in training and when it was loaded, the code in it will be directly used.

from idinvert.

Thank you for the explanation, that makes sense. Closing this as an issue.

from idinvert.