-
Notifications
You must be signed in to change notification settings - Fork 1
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.
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
Gpu enable roar #20
Gpu enable roar #20
Conversation
94838dc
to
3f45565
Compare
3f45565
to
95ccbe2
Compare
@@ -133,7 +146,7 @@ def _mask_batch(self, batch): | |||
|
|||
def _mask_dataset(self, dataloader, name): | |||
outputs = [] | |||
for batch in tqdm(dataloader(batch_size=self.batch_size, num_workers=0, shuffle=False), | |||
for batch in tqdm(dataloader(batch_size=8, num_workers=0, shuffle=False), |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Is the batch size fixed here due to memory issues?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yes. I haven't found another way. It doesn't appear to affect performance much, but properly something I should look more into.
@@ -21,4 +21,4 @@ pip3 install --no-index --find-links $HOME/python_wheels \ | |||
# Install comp550 | |||
cd $HOME/workspace/comp550 | |||
pip3 install --no-index --no-deps -e . | |||
python3 -u -X faulthandler "$@" --use-gpu True --num-workers 4 --persistent-dir $SCRATCH/comp550 | |||
python3 -u -X faulthandler "$@" --use-gpu True --num-workers 3 --persistent-dir $SCRATCH/comp550 |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Why did we enable the fault handler again? I think you mentioned it in a comment before, but I couldn't find it
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
It essentially has no performance impact and in case of a Segmentation Fault, it provides a stack trace based on the coredump. I was getting Segmentation Faults when looking into memory issues.
This looks good to merge to me. |
Depends on #19 landing. I will rebase this once #19 landsI enabled CUDA for ROAR and recallibrated all the walltime estimates. Everything should be working and running in a resonable amount of time now.
I did look into using
Trainer().predict()
which was added in 1.2.0. However, it doesn't appear to work and while in the CHANGELOG it is still undocumented.