Link Search Menu Expand Document

Frequently Asked Questions


Low performance compared to the original dataset

Since our data generation code is based on pytorch-lightning, which supports a convenient toolkit for DDP(Distributed Data Parallel), several side-effects occur when you have installed a different version. For instance, in PyTorch-lightning 1.6.0, we found that the global step is fixed during the training phase. This affects on the learning rate since our code uses a learning rate scheduler based on the global step.