nachiket273 / lookahead_pytorch

Pytorch implementation of lookahead optimizer(https://arxiv.org/pdf/1907.08610.pdf) and RAdam(https://arxiv.org/pdf/1908.03265.pdf)

Implementation of Lookahead optimizer and RAdam.

Lookahead Optimizer

Pytorch implementation of Lookahead Optimizer

RAdam Optimizer

RAdam optimizer

Dependencies

• PyTorch

Usage

Lookahead

from Lookahead import Lookahead
optim = torch.optim.Adam(model.parameters(), lr=0.001 )
optimizer = Lookahead( optim, alpha= 0.6 , k = 10)

RAdam

from RAdam import RAdam
optim = RAdam(model.parameters(), lr=1e-3, weight_decay=1e-4)

Results

• Reported results of run on CIFAR-10 with 3 different seed.
• Used architecture ResNet-18 and trained for 90 epochs.
• Used lr schedule to divide learning rate by 5 after every 30 epochs.

Seed / Optimizer	SGD	Adamw	Lookahead with SGD	RAdam	Lookahead with Radam
42	93.31	92.78	93.34	93.01	93.21
17	93.30	92.77	93.36	93.02	93.20
11	93.33	92.78	93.40	93.03	93.14