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

Jump to bottom

Fix random sampling pytorch non-determinism. #145

Merged
merged 1 commit into from
Jan 29, 2024
Merged

Fix random sampling pytorch non-determinism. #145

Show file tree
Hide file tree
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
2 changes: 1 addition & 1 deletion .github/workflows/python-package-conda.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -185,7 +185,7 @@ jobs:
run: |
source ${VENV}/bin/activate
mpirun -np 2 dlio_benchmark workload=unet3d ++workload.train.computation_time=0.05 ++workload.evaluation.eval_time=0.01 ++workload.workflow.train=False ++workload.workflow.generate_data=True ++workload.dataset.num_files_train=8 ++workload.dataset.num_files_eval=8 ++workload.reader.read_threads=2 ++workload.dataset.record_length=4096 ++workload.dataset.record_length_stdev=0
mpirun -np 2 dlio_benchmark workload=unet3d ++workload.train.computation_time=0.05 ++workload.evaluation.eval_time=0.01 ++workload.train.epochs=1 ++workload.workflow.train=True ++workload.workflow.generate_data=False ++workload.dataset.num_files_train=8 ++workload.dataset.num_files_eval=8 ++workload.reader.read_threads=2 ++workload.dataset.record_length=4096 ++workload.dataset.record_length_stdev=0
mpirun -np 2 dlio_benchmark workload=unet3d ++workload.train.computation_time=0.05 ++workload.evaluation.eval_time=0.01 ++workload.train.epochs=1 ++workload.workflow.train=True ++workload.workflow.generate_data=False ++workload.dataset.num_files_train=8 ++workload.dataset.num_files_eval=8 ++workload.reader.read_threads=0 ++workload.dataset.record_length=4096 ++workload.dataset.record_length_stdev=0
- name: test-tf-loader-npz
run: |
source ${VENV}/bin/activate
Expand Down
14 changes: 10 additions & 4 deletions dlio_benchmark/data_loader/torch_data_loader.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -75,7 +75,7 @@ def __len__(self):
def __getitem__(self, image_idx):
self.num_images_read += 1
step = int(math.ceil(self.num_images_read / self.batch_size))
logging.debug(f"{utcnow()} Rank {DLIOMPI.get_instance().rank()} reading {image_idx} sample")
logging.info(f"{utcnow()} Rank {DLIOMPI.get_instance().rank()} reading {image_idx} sample")
return self.reader.read_index(image_idx, step)

class TorchDataLoader(BaseDataLoader):
Expand All @@ -85,10 +85,16 @@ def __init__(self, format_type, dataset_type, epoch_number):

@dlp.log
def read(self):
do_shuffle = True if self._args.sample_shuffle != Shuffle.OFF else False
dataset = TorchDataset(self.format_type, self.dataset_type, self.epoch_number, self.num_samples, self._args.read_threads, self.batch_size)
if do_shuffle:
sampler = RandomSampler(dataset)
if self._args.sample_shuffle != Shuffle.OFF:
# torch seed is used for all functions within.
torch.manual_seed(self._args.seed)
seed = int(torch.empty((), dtype=torch.int64).random_().item())
# generator needs to load up torch seed.
torch_generator = torch.Generator()
torch_generator.manual_seed(seed)
# Pass generator to sampler
sampler = RandomSampler(dataset, generator=torch_generator)
else:
sampler = SequentialSampler(dataset)
if self._args.read_threads >= 1:
Expand Down
15 changes: 8 additions & 7 deletions dlio_benchmark/utils/config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -312,13 +312,14 @@ def get_global_map_index(self, file_list, total_samples):

@dlp.log
def reconfigure(self, epoch_number, dataset_type):
if self.file_shuffle is not Shuffle.OFF:
if self.seed_change_epoch:
np.random.seed(self.seed + epoch_number)
else:
np.random.seed(self.seed)
np.random.shuffle(self.file_list_train) if dataset_type is DatasetType.TRAIN else np.random.shuffle(
self.file_list_eval)
if self.data_loader_sampler == DataLoaderSampler.ITERATIVE:
if self.file_shuffle is not Shuffle.OFF:
if self.seed_change_epoch:
np.random.seed(self.seed + epoch_number)
else:
np.random.seed(self.seed)
np.random.shuffle(self.file_list_train) if dataset_type is DatasetType.TRAIN else np.random.shuffle(
self.file_list_eval)

if self.data_loader_sampler == DataLoaderSampler.ITERATIVE:
if dataset_type is DatasetType.TRAIN:
Expand Down
Loading