Skip to content

support random seed for RA sampler #5053

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

Merged
merged 3 commits into from
Dec 8, 2021
Merged

support random seed for RA sampler #5053

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
c7f46e9
support random seed used to shuffle the sampler
xiaohu2015 Dec 8, 2021
d23d118
fix bug for RA samper
xiaohu2015 Dec 8, 2021
17ac681
Merge branch 'main' into patch-1
datumbox Dec 8, 2021
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
5 changes: 3 additions & 2 deletions references/classification/sampler.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,7 +15,7 @@ class RASampler(torch.utils.data.Sampler):
https://github.com/facebookresearch/deit/blob/main/samplers.py
"""

def __init__(self, dataset, num_replicas=None, rank=None, shuffle=True):
def __init__(self, dataset, num_replicas=None, rank=None, shuffle=True, seed=0):
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I agree this might be useful but it's also a paradigm we don't do on TorchVision. Don't get me wrong, this approach is popular with other libraries it's just that setting seeds on constructors is something we don't typically do.

@fmassa thoughts?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@datumbox The RA sampler is modified from torch.utils.data.DistributedSampler which also has a parameter seed. The default value of seed is 0, so it does not affect the behavior of RA samper if users don't explicitly set a seed.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Fair enough. Since DistributedSampler supports it we can also have it.

if num_replicas is None:
if not dist.is_available():
raise RuntimeError("Requires distributed package to be available!")
Expand All @@ -32,11 +32,12 @@ def __init__(self, dataset, num_replicas=None, rank=None, shuffle=True):
self.total_size = self.num_samples * self.num_replicas
self.num_selected_samples = int(math.floor(len(self.dataset) // 256 * 256 / self.num_replicas))
self.shuffle = shuffle
self.seed = seed

def __iter__(self):
# Deterministically shuffle based on epoch
g = torch.Generator()
g.manual_seed(self.epoch)
g.manual_seed(self.seed + self.epoch)
if self.shuffle:
indices = torch.randperm(len(self.dataset), generator=g).tolist()
else:
Expand Down
2 changes: 1 addition & 1 deletion references/classification/train.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,7 +9,7 @@
import torchvision
import transforms
import utils
from references.classification.sampler import RASampler
from sampler import RASampler
Copy link
Contributor

@datumbox datumbox Dec 8, 2021 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Good catch, if this PR is not merged could you please bring this separately?

from torch import nn
from torch.utils.data.dataloader import default_collate
from torchvision.transforms.functional import InterpolationMode
Expand Down