This repository contains an implementation of AdamW optimization algorithm and cosine learning rate scheduler described in "Decoupled Weight Decay Regularization". AdamW implementation is straightforward and does not differ much from existing Adam implementation for PyTorch, except that it separates weight decaying from batch gradient calculations. Cosine annealing scheduler with restarts allows model to converge to a (possibly) different local minimum on every restart and normalizes weight decay hyperparameter value according to the length of restart period. Unlike schedulers presented in standard PyTorch scheduler suite this scheduler adjusts optimizer's learning rate not on every epoch, but on every batch update, according to the paper.
Besides "cosine" and "arccosine" policies (arccosine
has steeper profile at the limiting points), there are "triangular", triangular2
and exp_range
, which implement policies proposed in "Cyclical Learning Rates for Training Neural Networks".
The ratio of increasing and decreasing phases for triangular policy could be adjusted with triangular_step
parameter. Minimum allowed lr is adjusted by min_lr
parameter.
triangular
schedule is enabled by passingpolicy="triangular"
parameter.triangular2
schedule reduces maximum lr by half on each restart cycle and is enabled by passingpolicy="triangular2"
parameter, or by combining parameterspolicy="triangular", eta_on_restart_cb=ReduceMaxLROnRestart(ratio=0.5)
. Theratio
parameter regulates the factor by which lr is scaled on each restart.exp_range
schedule is enabled by passingpolicy="exp_range"
parameter. It exponentially scales maximum lr depending on iteration count. The base of exponentiation is set bygamma
parameter.
These schedules could be combined with shrinking/expanding restart periods, weight decay normalization and could be used with AdamW and other PyTorch optimizers.
batch_size = 32
epoch_size = 1024
model = resnet()
optimizer = AdamW(model.parameters(), lr=1e-3, weight_decay=1e-5)
scheduler = CyclicLRWithRestarts(optimizer, batch_size, epoch_size, restart_period=5, t_mult=1.2, policy="cosine")
for epoch in range(100):
scheduler.step()
train_for_every_batch(...)
...
optimizer.zero_grad()
loss.backward()
optimizer.step()
scheduler.batch_step()
validate(...)