Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Using torch.bfloat16 to prevent overflow instead of default fp16 in AMP #345

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Using torch.bfloat16 to prevent overflow instead of default fp16 in AMP #345

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 4 additions & 2 deletions main.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -184,7 +184,8 @@ def train_one_epoch(config, model, criterion, data_loader, optimizer, epoch, mix
if mixup_fn is not None:
samples, targets = mixup_fn(samples, targets)

with torch.cuda.amp.autocast(enabled=config.AMP_ENABLE):
# Using torch.bfloat16 to prevent overflow. Float16 has three less integer bits compared to bfloat16 which causes NaN loss and NaN grad norms during AMP training.
with torch.cuda.amp.autocast(enabled=config.AMP_ENABLE, dtype=torch.bfloat16):
outputs = model(samples)
loss = criterion(outputs, targets)
loss = loss / config.TRAIN.ACCUMULATION_STEPS
Expand Down Expand Up @@ -241,7 +242,8 @@ def validate(config, data_loader, model):
target = target.cuda(non_blocking=True)

# compute output
with torch.cuda.amp.autocast(enabled=config.AMP_ENABLE):
# Using torch.bfloat16 to prevent overflow. Float16 has three less integer bits compared to bfloat16 which causes NaN loss and NaN grad norms during AMP training.
with torch.cuda.amp.autocast(enabled=config.AMP_ENABLE, dtype=torch.bfloat16):
output = model(images)

# measure accuracy and record loss
Expand Down