mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-08-09 01:52:41 +00:00
Code Issues Packages Projects Releases Wiki Activity

Implement 2o schedule

Browse Source
This commit is contained in:
Daniel Povey 2022-04-09 18:37:03 +08:00
parent db72aee1f0
commit 4d41ee0caa
2 changed files with 8 additions and 20 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
egs/librispeech/ASR/pruned_transducer_stateless2/optim.py
View File

@ -100,9 +100,6 @@ class Eve(Optimizer):
if p.grad is None:
continue
# Perform optimization step
grad = p.grad
if grad.is_sparse:
@ -144,12 +141,3 @@ class Eve(Optimizer):
p.addcdiv_(exp_avg, denom, value=-step_size)
return loss
# Note on avg-change per epoch..
# suppose epoch is 4k iters.
# if avg-change as rms(diff) / rms(params) equals 0.2, and rms(params) = 0.1,
# then rm(diff) 0.1 * 0.2, var(diff) = (0.1 * 0.2)**2, = 0.0004. So var(diff per minibatch)
# = 0.0004 / 4000 = 1e-07, rms(diff per minibatch) = 3.16e-04. So LR would be 3e-04.
# Suggested lr_schedule?
#
# .. 6e-05 is 1/5 of that...

16
egs/librispeech/ASR/pruned_transducer_stateless2/train.py
View File

@ -154,15 +154,15 @@ def get_parser():
parser.add_argument(
"--lr-begin-steps",
type=float,
default=25000,
default=5000,
help="Number of steps that affects how rapidly the learning rate initially decreases"
)
parser.add_argument(
"--lr-end-epochs",
"--lr-epochs",
type=float,
default=-1,
help="""Number of epochs that affects how rapidly the learning rate finally decreases;
help="""Number of epochs for purposes of the learning-rate schedule;
if -1, will be set the same as --num-epochs
"""
)
@ -784,15 +784,15 @@ def run(rank, world_size, args):
model.parameters(),
lr=params.initial_lr)
# The `epoch` variable in the lambda expression binds to the value below
# in `for epoch in range(params.start_epoch, params.num_epochs):`. But set it to 0
# The `epoch` variable in the lambda expression picks up to the value below
# in `for epoch in range(params.start_epoch, params.num_epochs):`. Set it to 0
# here to avoid crash in constructor.
epoch = 0
lr_end_epochs = params.lr_end_epochs if params.lr_end_epochs > 0 else params.num_epochs
lr_epochs = params.lr_epochs if params.lr_epochs > 0 else params.num_epochs
scheduler = torch.optim.lr_scheduler.LambdaLR(
optimizer,
lambda step: (((step + params.lr_begin_steps) / params.lr_begin_steps) ** -0.5 *
((epoch + lr_end_epochs) / lr_end_epochs) ** -2.0))
lambda step: (((step**2 + params.lr_begin_steps**2) / params.lr_begin_steps**2) ** -0.25 *
((epoch + lr_epochs) / lr_epochs) ** -0.5))
if checkpoints and "optimizer" in checkpoints: