def make_workloads() -> workload.Stream:
trainer = utils.TrainAndValidate()
yield from trainer.send(steps=3,
validation_freq=3,
scheduling_unit=10)
training_metrics = trainer.get_avg_training_metrics()
_, validation_metrics = trainer.result()
batch_size = self.hparams["global_batch_size"]
for i, metrics in enumerate(training_metrics):
expect = pytorch_onevar_model.TriangleLabelSum.expect(
batch_size, 10 * i, 10 * (i + 1))
assert "cls_reducer" in metrics
assert metrics["cls_reducer"] == expect
assert "fn_reducer" in metrics
assert metrics["fn_reducer"] == expect
for metrics in validation_metrics:
num_batches = len(
pytorch_onevar_model.OnesDataset()) // batch_size
expect = pytorch_onevar_model.TriangleLabelSum.expect(
batch_size, 0, num_batches)
assert "cls_reducer" in metrics
assert metrics["cls_reducer"] == expect
assert "fn_reducer" in metrics
assert metrics["fn_reducer"] == expect
yield workload.terminate_workload(
), [], workload.ignore_workload_response
def make_workloads() -> workload.Stream:
trainer = utils.TrainAndValidate()
yield from trainer.send(steps=2,
validation_freq=1,
scheduling_unit=1)
yield workload.terminate_workload(
), [], workload.ignore_workload_response
def make_workloads() -> workload.Stream:
trainer = utils.TrainAndValidate()
yield from trainer.send(steps=10,
validation_freq=1,
batches_per_step=100)
training_metrics, validation_metrics = trainer.result()
# We expect the validation error and training loss to be
# monotonically decreasing.
# TODO(DET-1597): actually use a model and optimizer where the losses
# are monotonically decreasing.
for older, newer in zip(training_metrics[::100],
training_metrics[::100][1:]):
assert newer["loss"] <= older["loss"]
for older, newer in zip(validation_metrics,
validation_metrics[1:]):
assert newer["val_categorical_error"] <= older[
"val_categorical_error"]
assert validation_metrics[-1][
"val_categorical_error"] == pytest.approx(0.0)
yield workload.terminate_workload(
), [], workload.ignore_workload_response
def make_test_workloads(
checkpoint_dir: pathlib.Path, config: det.ExperimentConfig
) -> workload.Stream:
print("Start training a test experiment.")
interceptor = workload.WorkloadResponseInterceptor()
print("Training 1 step.")
yield from interceptor.send(workload.train_workload(1), [config.batches_per_step()])
metrics = interceptor.metrics_result()
batch_metrics = metrics["batch_metrics"]
check.eq(len(batch_metrics), config.batches_per_step())
print(f"Finished training. Metrics: {batch_metrics}")
print("Validating.")
yield from interceptor.send(workload.validation_workload(1), [])
validation = interceptor.metrics_result()
v_metrics = validation["validation_metrics"]
print(f"Finished validating. Validation metrics: {v_metrics}")
print(f"Saving a checkpoint to {checkpoint_dir}")
yield workload.checkpoint_workload(), [checkpoint_dir], workload.ignore_workload_response
print(f"Finished saving a checkpoint to {checkpoint_dir}.")
yield workload.terminate_workload(), [], workload.ignore_workload_response
print("The test experiment passed.")
def make_workloads() -> workload.Stream:
nonlocal w
interceptor = workload.WorkloadResponseInterceptor()
for idx, batch in batches:
yield from interceptor.send(workload.train_workload(1), [])
metrics = interceptor.metrics_result()
# Calculate what the loss should be.
loss = trial_class.calc_loss(w, batch)
assert metrics["metrics"]["avg_metrics"][
"loss"] == pytest.approx(loss)
# Update what the weight should be.
w = w - hparams["learning_rate"] * trial_class.calc_gradient(
w, batch)
if test_checkpointing and idx == 3:
# Checkpoint and let the next TrialController finish the work.l
yield workload.checkpoint_workload(), [
checkpoint_dir
], workload.ignore_workload_response
break
yield workload.terminate_workload(
), [], workload.ignore_workload_response
def make_workloads() -> workload.Stream:
trainer = utils.TrainAndValidate()
yield from trainer.send(steps=1,
validation_freq=1,
batches_per_step=1)
yield workload.terminate_workload(
), [], workload.ignore_workload_response
def make_workloads1() -> workload.Stream:
nonlocal controller
yield workload.train_workload(1, 1, 0), [], workload.ignore_workload_response
assert controller is not None, "controller was never set!"
assert controller.trial.counter.__dict__ == {
"validation_steps_started": 0,
"validation_steps_ended": 0,
"checkpoints_ended": 0,
}
yield workload.validation_workload(), [], workload.ignore_workload_response
assert controller.trial.counter.__dict__ == {
"validation_steps_started": 1,
"validation_steps_ended": 1,
"checkpoints_ended": 0,
}
yield workload.checkpoint_workload(), [
checkpoint_dir
], workload.ignore_workload_response
assert controller.trial.counter.__dict__ == {
"validation_steps_started": 1,
"validation_steps_ended": 1,
"checkpoints_ended": 1,
}
yield workload.terminate_workload(), [], workload.ignore_workload_response
def make_workloads() -> workload.Stream:
training_metrics = []
interceptor = workload.WorkloadResponseInterceptor()
total_steps, total_batches_processed = 10, 0
for step_id in range(1, total_steps):
num_batches = step_id
yield from interceptor.send(
workload.train_workload(
step_id,
num_batches=num_batches,
total_batches_processed=total_batches_processed,
),
[],
)
metrics = interceptor.metrics_result()
batch_metrics = metrics["metrics"]["batch_metrics"]
assert len(
batch_metrics
) == num_batches, "did not run for expected num_batches"
training_metrics.extend(batch_metrics)
total_batches_processed += num_batches
yield workload.terminate_workload(
), [], workload.ignore_workload_response
def make_workloads_2() -> workload.Stream:
trainer = utils.TrainAndValidate()
yield from trainer.send(steps=1, validation_freq=1)
yield workload.checkpoint_workload(), [
checkpoint_dir
], workload.ignore_workload_response
yield workload.terminate_workload(), [], workload.ignore_workload_response
def make_workloads(checkpoint_dir: pathlib.Path) -> workload.Stream:
trainer = utils.TrainAndValidate()
yield from trainer.send(steps=10, validation_freq=5, batches_per_step=5)
yield workload.checkpoint_workload(), [
checkpoint_dir
], workload.ignore_workload_response
yield workload.terminate_workload(), [], workload.ignore_workload_response
def make_workloads() -> workload.Stream:
trainer = utils.TrainAndValidate(request_stop_step_id=1)
yield from trainer.send(steps=100,
validation_freq=2,
scheduling_unit=5)
tm, vm = trainer.result()
yield workload.terminate_workload(
), [], workload.ignore_workload_response
def make_workloads() -> workload.Stream:
trainer = utils.TrainAndValidate(
request_stop_step_id=request_stop_step_id)
yield from trainer.send(steps=2,
validation_freq=2,
batches_per_step=5)
tm, vm = trainer.result()
yield workload.terminate_workload(
), [], workload.ignore_workload_response
def make_workloads(tag: str) -> workload.Stream:
trainer = utils.TrainAndValidate()
yield from trainer.send(steps=1000, validation_freq=100)
tm, vm = trainer.result()
training_metrics[tag] = tm
validation_metrics[tag] = vm
yield workload.terminate_workload(), [], workload.ignore_workload_response
def make_workloads() -> workload.Stream:
trainer = utils.TrainAndValidate()
yield from trainer.send(steps=15,
validation_freq=4,
scheduling_unit=5)
training_metrics, validation_metrics = trainer.result()
yield workload.terminate_workload(
), [], workload.ignore_workload_response
def make_workloads() -> workload.Stream:
trainer = utils.TrainAndValidate()
yield from trainer.send(steps=10, validation_freq=10)
training_metrics, validation_metrics = trainer.result()
for metrics in training_metrics:
assert "accuracy" in metrics
yield workload.terminate_workload(), [], workload.ignore_workload_response
def make_workloads(steps: int) -> workload.Stream:
trainer = TrainAndValidate()
yield from trainer.send(steps, validation_freq=1, scheduling_unit=10)
tm, vm = trainer.result()
metrics["training"] += tm
metrics["validation"] += vm
yield workload.terminate_workload(
), [], workload.ignore_workload_response
def make_workloads_2() -> workload.Stream:
interceptor = workload.WorkloadResponseInterceptor()
yield from interceptor.send(workload.validation_workload(), [])
metrics = interceptor.metrics_result()
new_loss = metrics["metrics"]["validation_metrics"]["val_loss"]
assert new_loss == pytest.approx(old_loss)
yield workload.terminate_workload(), [], workload.ignore_workload_response
def make_workloads() -> workload.Stream:
trainer = utils.TrainAndValidate()
yield from trainer.send(steps=1000, validation_freq=100)
training_metrics, validation_metrics = trainer.result()
# We expect the validation error and training loss to be
# monotonically decreasing.
for older, newer in zip(training_metrics, training_metrics[1:]):
assert newer["loss"] <= older["loss"]
yield workload.terminate_workload(), [], workload.ignore_workload_response
def make_workloads(tag: str) -> workload.Stream:
nonlocal training_metrics
nonlocal validation_metrics
trainer = TrainAndValidate()
yield from trainer.send(steps, validation_freq, batches_per_step=batches_per_step)
tm, vm = trainer.result()
training_metrics[tag] = tm
validation_metrics[tag] = vm
yield workload.terminate_workload(), [], workload.ignore_workload_response
def make_workloads_1() -> workload.Stream:
nonlocal old_loss
trainer = utils.TrainAndValidate()
yield from trainer.send(steps=10, validation_freq=10)
training_metrics, validation_metrics = trainer.result()
old_loss = validation_metrics[-1]["val_loss"]
yield workload.checkpoint_workload(), [
checkpoint_dir
], workload.ignore_workload_response
yield workload.terminate_workload(), [], workload.ignore_workload_response
def make_workloads(checkpoint_dir: str = "") -> workload.Stream:
nonlocal training_metrics
trainer = utils.TrainAndValidate()
yield from trainer.send(steps=10, validation_freq=10, batches_per_step=1)
tm, _ = trainer.result()
training_metrics += tm
if checkpoint_dir:
yield workload.checkpoint_workload(), [
checkpoint_dir
], workload.ignore_workload_response
yield workload.terminate_workload(), [], workload.ignore_workload_response
def make_workloads() -> workload.Stream:
trainer = utils.TrainAndValidate()
yield from trainer.send(steps=100, validation_freq=10)
training_metrics, validation_metrics = trainer.result()
# Check the gradient update at every step.
for idx, batch_metrics in enumerate(training_metrics):
pytorch_onevar_model.OneVarTrial.check_batch_metrics(batch_metrics, idx)
# We expect the validation error and training loss to be
# monotonically decreasing.
for older, newer in zip(training_metrics, training_metrics[1:]):
assert newer["loss"] <= older["loss"]
yield workload.terminate_workload(), [], workload.ignore_workload_response
def make_workloads(
steps: int, tag: str, checkpoint_dir: Optional[pathlib.Path] = None
) -> workload.Stream:
trainer = TrainAndValidate()
yield from trainer.send(steps, validation_freq=1, batches_per_step=100)
tm, vm = trainer.result()
training_metrics[tag] += tm
validation_metrics[tag] += vm
if checkpoint_dir is not None:
yield workload.checkpoint_workload(), [
checkpoint_dir
], workload.ignore_workload_response
yield workload.terminate_workload(), [], workload.ignore_workload_response
def make_workloads() -> workload.Stream:
trainer = utils.TrainAndValidate()
yield from trainer.send(steps=10, validation_freq=5, batches_per_step=1000)
training_metrics, validation_metrics = trainer.result()
# We expect the training loss to be monotonically decreasing and the
# accuracy to be monotonically increasing.
for older, newer in zip(training_metrics, training_metrics[1:]):
assert newer["loss"] < older["loss"]
for older, newer in zip(validation_metrics, validation_metrics[1:]):
assert newer["accuracy"] >= older["accuracy"]
# The final accuracy should be 100%.
assert validation_metrics[-1]["accuracy"] == pytest.approx(1.0)
yield workload.terminate_workload(), [], workload.ignore_workload_response
def make_workloads() -> workload.Stream:
trainer = utils.TrainAndValidate()
# Test >1 validation to ensure that resetting the allgather_op list is working.
yield from trainer.send(steps=2,
validation_freq=1,
batches_per_step=1)
training_metrics, validation_metrics = trainer.result()
for metrics in validation_metrics:
assert metrics[
"label_sum_fn"] == estimator_linear_model.validation_label_sum(
)
assert metrics[
"label_sum_cls"] == estimator_linear_model.validation_label_sum(
)
yield workload.terminate_workload(
), [], workload.ignore_workload_response
def _make_test_workloads(checkpoint_dir: pathlib.Path,
config: det.ExperimentConfig) -> workload.Stream:
interceptor = workload.WorkloadResponseInterceptor()
logging.info("Training one batch")
yield from interceptor.send(workload.train_workload(1), [])
metrics = interceptor.metrics_result()
batch_metrics = metrics["metrics"]["batch_metrics"]
check.eq(len(batch_metrics), config.scheduling_unit())
logging.info(f"Finished training, metrics: {batch_metrics}")
logging.info("Validating one batch")
yield from interceptor.send(workload.validation_workload(1), [])
validation = interceptor.metrics_result()
v_metrics = validation["metrics"]["validation_metrics"]
logging.info(f"Finished validating, validation metrics: {v_metrics}")
logging.info(f"Saving a checkpoint to {checkpoint_dir}.")
yield workload.checkpoint_workload(), [checkpoint_dir
], workload.ignore_workload_response
logging.info(f"Finished saving a checkpoint to {checkpoint_dir}.")
yield workload.terminate_workload(), [], workload.ignore_workload_response
logging.info("The test experiment passed.")
def make_workloads2() -> workload.Stream:
yield workload.terminate_workload(
), [], workload.ignore_workload_response
