SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Properly calculate the total train iterations and recalculate num epochs in no_trainer scripts (#17856)

Browse Source
This commit is contained in:
Zachary Mueller
2022-06-23 15:46:01 -04:00
committed by GitHub
parent 7c1b91281f
commit 75259b44bf
12 changed files with 70 additions and 37 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
examples/pytorch/language-modeling/run_clm_no_trainer.py
View File

@@ -474,11 +474,11 @@ def main():
model.tie_weights()
# Scheduler and math around the number of training steps.
overrode_max_train_steps = False
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
if args.max_train_steps is None:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
else:
args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
overrode_max_train_steps = True
lr_scheduler = get_scheduler(
name=args.lr_scheduler_type,
@@ -494,7 +494,10 @@ def main():
# We need to recalculate our total training steps as the size of the training dataloader may have changed.
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
if overrode_max_train_steps:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
# Afterwards we recalculate our number of training epochs
args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
# Figure out how many steps we should save the Accelerator states
if hasattr(args.checkpointing_steps, "isdigit"):

9
examples/pytorch/language-modeling/run_mlm_no_trainer.py
View File

@@ -518,11 +518,11 @@ def main():
# shorter in multiprocess)
# Scheduler and math around the number of training steps.
overrode_max_train_steps = False
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
if args.max_train_steps is None:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
else:
args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
overrode_max_train_steps = True
lr_scheduler = get_scheduler(
name=args.lr_scheduler_type,
@@ -538,7 +538,10 @@ def main():
# We need to recalculate our total training steps as the size of the training dataloader may have changed.
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
if overrode_max_train_steps:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
# Afterwards we recalculate our number of training epochs
args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
# Figure out how many steps we should save the Accelerator states
if hasattr(args.checkpointing_steps, "isdigit"):