def test_mxnet_with_rules_and_debugger_hook_config(
sagemaker_session,
mxnet_training_latest_version,
mxnet_training_latest_py_version,
cpu_instance_type,
):
with timeout(minutes=TRAINING_DEFAULT_TIMEOUT_MINUTES):
rules = [
Rule.sagemaker(rule_configs.vanishing_gradient()),
Rule.sagemaker(base_config=rule_configs.all_zero(),
rule_parameters={"tensor_regex": ".*"}),
Rule.sagemaker(rule_configs.loss_not_decreasing()),
]
debugger_hook_config = DebuggerHookConfig(s3_output_path=os.path.join(
"s3://", sagemaker_session.default_bucket(), str(uuid.uuid4()),
"tensors"))
script_path = os.path.join(DATA_DIR, "mxnet_mnist", "mnist_gluon.py")
data_path = os.path.join(DATA_DIR, "mxnet_mnist")
mx = MXNet(
entry_point=script_path,
role="SageMakerRole",
framework_version=mxnet_training_latest_version,
py_version=mxnet_training_latest_py_version,
instance_count=1,
instance_type=cpu_instance_type,
sagemaker_session=sagemaker_session,
rules=rules,
debugger_hook_config=debugger_hook_config,
)
train_input = mx.sagemaker_session.upload_data(
path=os.path.join(data_path, "train"),
key_prefix="integ-test-data/mxnet_mnist/train")
test_input = mx.sagemaker_session.upload_data(
path=os.path.join(data_path, "test"),
key_prefix="integ-test-data/mxnet_mnist/test")
mx.fit({"train": train_input, "test": test_input})
job_description = mx.latest_training_job.describe()
for index, rule in enumerate(rules):
assert (job_description["DebugRuleConfigurations"][index]
["RuleConfigurationName"] == rule.name)
assert (job_description["DebugRuleConfigurations"][index]
["RuleEvaluatorImage"] == rule.image_uri)
assert job_description["DebugRuleConfigurations"][index][
"VolumeSizeInGB"] == 0
assert (job_description["DebugRuleConfigurations"][index]
["RuleParameters"]["rule_to_invoke"] ==
rule.rule_parameters["rule_to_invoke"])
assert job_description[
"DebugHookConfig"] == debugger_hook_config._to_request_dict()
assert (job_description["DebugRuleEvaluationStatuses"] ==
mx.latest_training_job.rule_job_summary())
_wait_and_assert_that_no_rule_jobs_errored(
training_job=mx.latest_training_job)
def test_training_job_with_debugger_and_profiler(
sagemaker_session,
pipeline_name,
role,
pytorch_training_latest_version,
pytorch_training_latest_py_version,
):
instance_count = ParameterInteger(name="InstanceCount", default_value=1)
instance_type = ParameterString(name="InstanceType",
default_value="ml.m5.xlarge")
rules = [
Rule.sagemaker(rule_configs.vanishing_gradient()),
Rule.sagemaker(base_config=rule_configs.all_zero(),
rule_parameters={"tensor_regex": ".*"}),
Rule.sagemaker(rule_configs.loss_not_decreasing()),
]
debugger_hook_config = DebuggerHookConfig(
s3_output_path=
f"s3://{sagemaker_session.default_bucket()}/{uuid.uuid4()}/tensors")
base_dir = os.path.join(DATA_DIR, "pytorch_mnist")
script_path = os.path.join(base_dir, "mnist.py")
input_path = sagemaker_session.upload_data(
path=os.path.join(base_dir, "training"),
key_prefix="integ-test-data/pytorch_mnist/training",
)
inputs = TrainingInput(s3_data=input_path)
pytorch_estimator = PyTorch(
entry_point=script_path,
role="SageMakerRole",
framework_version=pytorch_training_latest_version,
py_version=pytorch_training_latest_py_version,
instance_count=instance_count,
instance_type=instance_type,
sagemaker_session=sagemaker_session,
rules=rules,
debugger_hook_config=debugger_hook_config,
)
step_train = TrainingStep(
name="pytorch-train",
estimator=pytorch_estimator,
inputs=inputs,
)
pipeline = Pipeline(
name=pipeline_name,
parameters=[instance_count, instance_type],
steps=[step_train],
sagemaker_session=sagemaker_session,
)
try:
response = pipeline.create(role)
create_arn = response["PipelineArn"]
execution = pipeline.start()
response = execution.describe()
assert response["PipelineArn"] == create_arn
try:
execution.wait(delay=10, max_attempts=60)
except WaiterError:
pass
execution_steps = execution.list_steps()
assert len(execution_steps) == 1
assert execution_steps[0].get("FailureReason", "") == ""
assert execution_steps[0]["StepName"] == "pytorch-train"
assert execution_steps[0]["StepStatus"] == "Succeeded"
training_job_arn = execution_steps[0]["Metadata"]["TrainingJob"]["Arn"]
job_description = sagemaker_session.sagemaker_client.describe_training_job(
TrainingJobName=training_job_arn.split("/")[1])
for index, rule in enumerate(rules):
config = job_description["DebugRuleConfigurations"][index]
assert config["RuleConfigurationName"] == rule.name
assert config["RuleEvaluatorImage"] == rule.image_uri
assert config["VolumeSizeInGB"] == 0
assert (config["RuleParameters"]["rule_to_invoke"] ==
rule.rule_parameters["rule_to_invoke"])
assert job_description[
"DebugHookConfig"] == debugger_hook_config._to_request_dict()
assert job_description["ProfilingStatus"] == "Enabled"
assert job_description["ProfilerConfig"][
"ProfilingIntervalInMilliseconds"] == 500
finally:
try:
pipeline.delete()
except Exception:
pass
def test_mxnet_with_profiler_and_debugger_then_disable_framework_metrics(
sagemaker_session,
mxnet_training_latest_version,
mxnet_training_latest_py_version,
cpu_instance_type,
):
with timeout(minutes=TRAINING_DEFAULT_TIMEOUT_MINUTES):
rules = [
Rule.sagemaker(rule_configs.vanishing_gradient()),
Rule.sagemaker(base_config=rule_configs.all_zero(),
rule_parameters={"tensor_regex": ".*"}),
ProfilerRule.sagemaker(rule_configs.ProfilerReport(),
name="CustomProfilerReportRule"),
]
debugger_hook_config = DebuggerHookConfig(
s3_output_path=
f"s3://{sagemaker_session.default_bucket()}/{str(uuid.uuid4())}/tensors",
)
profiler_config = ProfilerConfig(
s3_output_path=
f"s3://{sagemaker_session.default_bucket()}/{str(uuid.uuid4())}/system",
system_monitor_interval_millis=1000,
framework_profile_params=FrameworkProfile(),
)
script_path = os.path.join(DATA_DIR, "mxnet_mnist", "mnist_gluon.py")
data_path = os.path.join(DATA_DIR, "mxnet_mnist")
mx = MXNet(
entry_point=script_path,
role="SageMakerRole",
framework_version=mxnet_training_latest_version,
py_version=mxnet_training_latest_py_version,
instance_count=1,
instance_type=cpu_instance_type,
sagemaker_session=sagemaker_session,
rules=rules,
debugger_hook_config=debugger_hook_config,
profiler_config=profiler_config,
)
train_input = mx.sagemaker_session.upload_data(
path=os.path.join(data_path, "train"),
key_prefix="integ-test-data/mxnet_mnist/train")
test_input = mx.sagemaker_session.upload_data(
path=os.path.join(data_path, "test"),
key_prefix="integ-test-data/mxnet_mnist/test")
training_job_name = unique_name_from_base(
"test-profiler-mxnet-training")
mx.fit(
inputs={
"train": train_input,
"test": test_input
},
job_name=training_job_name,
wait=False,
)
job_description = mx.latest_training_job.describe()
assert job_description[
"ProfilerConfig"] == profiler_config._to_request_dict()
assert job_description[
"DebugHookConfig"] == debugger_hook_config._to_request_dict()
assert job_description.get("ProfilingStatus") == "Enabled"
profiler_rule_configuration = job_description.get(
"ProfilerRuleConfigurations")[0]
assert profiler_rule_configuration[
"RuleConfigurationName"] == "CustomProfilerReportRule"
assert profiler_rule_configuration["RuleEvaluatorImage"] == mx.rules[
0].image_uri
assert profiler_rule_configuration["RuleParameters"] == {
"rule_to_invoke": "ProfilerReport",
}
for index, rule in enumerate(mx.debugger_rules):
assert (job_description["DebugRuleConfigurations"][index]
["RuleConfigurationName"] == rule.name)
assert (job_description["DebugRuleConfigurations"][index]
["RuleEvaluatorImage"] == rule.image_uri)
_wait_until_training_can_be_updated(sagemaker_session.sagemaker_client,
training_job_name)
mx.update_profiler(disable_framework_metrics=True)
job_description = mx.latest_training_job.describe()
assert job_description["ProfilerConfig"]["ProfilingParameters"] == {}
