Skip to content

[Orbit]Handle iterator exhaustion in Controller.py #13595

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

Open
LINYV0719 wants to merge 1 commit into
base: master
Choose a base branch
from
Open

[Orbit]Handle iterator exhaustion in Controller.py #13595

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
34 changes: 27 additions & 7 deletions orbit/controller.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -273,15 +273,35 @@ def train(self, steps: int, checkpoint_at_completion: bool = True):
"""
self._require("trainer", for_method="train")

# TODO(momernick): Support steps=None or -1 (training to exhaustion).
current_step = self.global_step.numpy() # Cache, since this is expensive.
_log(f"train | step: {current_step: 6d} | training until step {steps}...")
while current_step < steps:
# Calculates steps to run for the next train loop.
num_steps = min(steps - current_step, self.steps_per_loop)
self._train_n_steps(num_steps)

if steps == -1:
_log(f"train | step: {current_step: 6d} | training until exhaustion...")
else:
_log(f"train | step: {current_step: 6d} | training until step {steps}...")

while steps == -1 or current_step < steps:
# Calculates steps to run for the next train loop.
if steps == -1:
num_steps = self.steps_per_loop
else:
num_steps = min(steps - current_step, self.steps_per_loop)

try:
self._train_n_steps(num_steps)
except (tf.errors.OutOfRangeError, StopIteration):
_log("Training stopped because the underlying iterator is exhausted.")
break

self._maybe_save_checkpoint()
current_step = self.global_step.numpy()

new_step = self.global_step.numpy()
# Stop if the iterator is exhausted (step count didn't increase as expected).
if new_step < current_step + num_steps:
_log("Training stopped because the underlying iterator is exhausted.")
break

current_step = new_step

if checkpoint_at_completion:
self._maybe_save_checkpoint(check_interval=False)
Expand Down
24 changes: 24 additions & 0 deletions orbit/controller_test.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -827,6 +827,30 @@ def __call__(self, output):
self.assertIn("eval_loss", output)
self.assertGreaterEqual(output["eval_loss"], 0)

def test_train_until_exhaustion(self):
test_runner = TestRunner()

# Create a finite dataset with only 10 elements (batch size 10) -> 1 step.
def finite_dataset_fn(ctx):
del ctx
inputs = np.zeros((10, 3), dtype=np.float32)
targets = np.ones((10, 4), dtype=np.float32)
dataset = tf.data.Dataset.from_tensor_slices((inputs, targets))
dataset = dataset.batch(1)
return dataset

test_runner.train_dataset = (
test_runner.strategy.distribute_datasets_from_function(finite_dataset_fn))

test_controller = controller.Controller(
trainer=test_runner,
global_step=test_runner.global_step,
steps_per_loop=1)

# Passing -1 should run until the dataset is exhausted (1 step).
test_controller.train(steps=-1)
self.assertEqual(test_runner.global_step.numpy(), 11)

def test_step_per_loop_callable(self):
test_runner = TestRunner()

Expand Down