search

BaguaSys / bagua

Bagua Speeds up PyTorch
https://tutorials-8ro.pages.dev/
MIT License
875 stars 83 forks source link

low decentralized algorithm's accuracy is not in accord with its torch's counterpart #165

Closed wangraying closed 3 years ago

wangraying commented 3 years ago

This issue seems caused by cargo profile settings.

When bagua-core is built for development, the accuracies (see test_low_precision_decentralized.py for the measurement of accuracies) of these two are completely the same.

When it is built for release, the accuracies of these two slightly differ, but is still reasonable (by calculating the difference of model weights against the max norm of model weights).

I will close this issue and relax the constraints in test_low_precision_decentralized.