def run_in_situ(self) -> None:
"""
Actually run the AzureML job; this method will typically run on an Azure VM.
"""
# Only set the logging level now. Usually, when we set logging to DEBUG, we want diagnostics about the model
# build itself, but not the tons of debug information that AzureML submissions create.
logging_to_stdout(self.azure_config.log_level)
suppress_logging_noise()
pytest_failed = False
training_failed = False
pytest_passed = True
# Ensure that both model training and pytest both get executed in all cases, so that we see a full set of
# test results in each PR
outputs_folder = self.model_config.outputs_folder
try:
logging_to_file(self.model_config.logs_folder / LOG_FILE_NAME)
try:
self.create_ml_runner().run()
except Exception as ex:
print_exception(ex, "Model training/testing failed.")
training_failed = True
if self.azure_config.pytest_mark:
try:
pytest_passed, results_file_path = run_pytest(
self.azure_config.pytest_mark, outputs_folder)
if not pytest_passed:
logging.error(
f"Not all PyTest tests passed. See {results_file_path}"
)
except Exception as ex:
print_exception(ex, "Unable to run PyTest.")
pytest_failed = True
finally:
# wait for aggregation if required, and only if the training actually succeeded.
if not training_failed and self.model_config.should_wait_for_other_cross_val_child_runs(
):
self.wait_for_cross_val_runs_to_finish_and_aggregate()
disable_logging_to_file()
message = []
if training_failed:
message.append("Training failed")
if pytest_failed:
message.append("Unable to run Pytest")
if not pytest_passed:
message.append("At least 1 test in Pytest failed")
# Terminate if pytest or model training has failed. This makes the smoke test in
# PR builds fail if pytest fails.
if message:
raise ValueError(
f"One component of the training pipeline failed: {'. '.join(message)}"
)
def test_logging_to_file(test_output_dirs: OutputFolderForTests) -> None:
# Log file should go to a new, non-existent folder, 2 levels deep
file_path = test_output_dirs.root_dir / "subdir1" / "subdir2" / "logfile.txt"
common_util.logging_to_file_handler = None
common_util.logging_to_file(file_path)
assert common_util.logging_to_file_handler is not None
log_line = "foo bar"
logging.getLogger().setLevel(logging.INFO)
logging.info(log_line)
common_util.disable_logging_to_file()
should_not_be_present = "This should not be present in logs"
logging.info(should_not_be_present)
assert common_util.logging_to_file_handler is None
# Wait for a bit, tests sometimes fail with the file not existing yet
time.sleep(2)
assert file_path.exists()
assert log_line in file_path.read_text()
assert should_not_be_present not in file_path.read_text()
def run_in_situ(self) -> None:
"""
Actually run the AzureML job; this method will typically run on an Azure VM.
"""
# Only set the logging level now. Usually, when we set logging to DEBUG, we want diagnostics about the model
# build itself, but not the tons of debug information that AzureML submissions create.
logging_to_stdout(self.azure_config.log_level)
suppress_logging_noise()
error_messages = []
# For the PR build in AzureML, we can either pytest, or the training of the simple PR model. Running both
# only works when using DDP_spawn, but that has as a side-effect that it messes up memory consumption of the
# large models.
if self.azure_config.pytest_mark:
try:
outputs_folder = Path.cwd() / fixed_paths.DEFAULT_AML_UPLOAD_DIR
pytest_passed, results_file_path = run_pytest(self.azure_config.pytest_mark, outputs_folder)
if not pytest_passed:
pytest_failures = f"Not all PyTest tests passed. See {results_file_path}"
logging.error(pytest_failures)
error_messages.append(pytest_failures)
except Exception as ex:
print_exception(ex, "Unable to run PyTest.")
error_messages.append(f"Unable to run PyTest: {ex}")
else:
# Set environment variables for multi-node training if needed.
# In particular, the multi-node environment variables should NOT be set in single node
# training, otherwise this might lead to errors with the c10 distributed backend
# (https://github.com/microsoft/InnerEye-DeepLearning/issues/395)
if self.azure_config.num_nodes > 1:
set_environment_variables_for_multi_node()
try:
logging_to_file(self.model_config.logs_folder / LOG_FILE_NAME)
try:
self.create_ml_runner().run()
except Exception as ex:
print_exception(ex, "Model training/testing failed.")
error_messages.append(f"Training failed: {ex}")
finally:
disable_logging_to_file()
# Terminate if pytest or model training has failed. This makes the smoke test in
# PR builds fail if pytest fails.
if error_messages:
raise ValueError(
f"At least one component of the runner failed: {os.linesep} {os.linesep.join(error_messages)}")
def run_in_situ(self) -> None:
"""
Actually run the AzureML job; this method will typically run on an Azure VM.
"""
# Only set the logging level now. Usually, when we set logging to DEBUG, we want diagnostics about the model
# build itself, but not the tons of debug information that AzureML submissions create.
# Suppress the logging from all processes but the one for GPU 0 on each node, to make log files more readable
logging_to_stdout(self.azure_config.log_level if is_local_rank_zero() else "ERROR")
suppress_logging_noise()
if is_global_rank_zero():
self.print_git_tags()
# For the PR build in AzureML, we can either pytest, or the training of the simple PR model. Running both
# only works when using DDP_spawn, but that has as a side-effect that it messes up memory consumption of the
# large models.
if self.azure_config.pytest_mark:
outputs_folder = Path.cwd() / fixed_paths.DEFAULT_AML_UPLOAD_DIR
pytest_passed, results_file_path = run_pytest(self.azure_config.pytest_mark, outputs_folder)
if not pytest_passed:
# Terminate if pytest has failed. This makes the smoke test in
# PR builds fail if pytest fails.
pytest_failures = f"Not all PyTest tests passed. See {results_file_path}"
raise ValueError(pytest_failures)
else:
# Set environment variables for multi-node training if needed.
# In particular, the multi-node environment variables should NOT be set in single node
# training, otherwise this might lead to errors with the c10 distributed backend
# (https://github.com/microsoft/InnerEye-DeepLearning/issues/395)
if self.azure_config.num_nodes > 1:
set_environment_variables_for_multi_node()
ml_runner = self.create_ml_runner()
ml_runner.setup()
ml_runner.start_logging_to_file()
try:
ml_runner.run()
finally:
disable_logging_to_file()
