pytorch-lamb 0.0.0

No project description provided

Implementation of https://arxiv.org/abs/1904.00962 for large batch, large learning rate training.

Bonus: TensorboardX logging (example below).

Try the sample

git clone [e-mail]:cybertronai/pytorch-lamb.git
cd pytorch-lamb
pip install -e .
python test_lamb.py
tensorboard --logdir=runs

Sample results

At --lr=.1, the Adam optimizer is unable to train. With a little weight decay, LAMB avoids diverging!

Green: python test_lamb.py --batch-size=512 --lr=.1 --wd=0 --log-interval=30 --optimizer=lamb

Blue: python test_lamb.py --batch-size=512 --lr=.1 --wd=.01 --log-interval=30 --optimizer=lamb

r1 is the L2 norm of the weights. You can see in the green plot that some of the weights start to run away, which leads to divergence. This is why weight decay helps.