Active LR optimizer for AdamW, RAdam, SGD with momentum, and Adabelief (ActiveAdamW, ActiveRAdam, ActiveSGD, and ActiveBelief, respectively).
This is the PyTorch implementation of the Active LR optimizer on top of AdamW, RAdam, SGD with momentum, and Adabelief as we proposed and benchmarked in "Read the Signs: Towards Invariance to Gradient Descent’s Hyperparameter Values" under review in ICML 2022.
Although following Algorithm 1 in the paper you can modify existing gradient descent optimizers to their ActiveLR version, we have provided the ActiveLR version of AdamW, RAdam, SGD with momentum, and Adabelief in the optimizers subdirectory.
Please use conda to create an environment based on the environment.yml file and copy the optimizer directory into the same directory as your python script.
Vanilla Adam LR = 1e-3
Vanilla Adam LR = 1e-5
Active Adam LR = 1e-5