def create_job(config,
cluster_id,
workspace,
experiment,
job_name,
image_name,
command,
number_of_vms=1):
''' Creates job
'''
input_directories = [
models.InputDirectory(id='SCRIPT',
path='$AZ_BATCHAI_MOUNT_ROOT/{0}/{1}'.format(
config.fileshare_mount_point, job_name)),
models.InputDirectory(id='DATASET',
path='$AZ_BATCHAI_MOUNT_ROOT/{0}/{1}'.format(
config.fileshare_mount_point, 'data'))
]
std_output_path_prefix = "$AZ_BATCHAI_MOUNT_ROOT/{0}".format(
config.fileshare_mount_point)
output_directories = [
models.OutputDirectory(id='MODEL',
path_prefix='$AZ_BATCHAI_MOUNT_ROOT/{0}'.format(
config.fileshare_mount_point),
path_suffix="models"),
models.OutputDirectory(id='NOTEBOOKS',
path_prefix='$AZ_BATCHAI_MOUNT_ROOT/{0}'.format(
config.fileshare_mount_point),
path_suffix="notebooks")
]
parameters = models.JobCreateParameters(
location=config.location,
cluster=models.ResourceId(id=cluster_id),
node_count=number_of_vms,
input_directories=input_directories,
std_out_err_path_prefix=std_output_path_prefix,
output_directories=output_directories,
container_settings=models.ContainerSettings(
image_source_registry=models.ImageSourceRegistry(
image=image_name)),
custom_toolkit_settings=models.CustomToolkitSettings(
command_line=command))
client = client_from(config)
_ = client.jobs.create(config.group_name, workspace, experiment, job_name,
parameters)
def create_custom_job(client,
resource_group,
cluster_id,
job_name,
nodes,
cmd,
job_preparation_cmd=None,
container=None):
"""Creates custom toolkit job
:param BatchAIManagementClient client: client instance.
:param str resource_group: resource group name.
:param str cluster_id: resource Id of the cluster.
:param str job_name: job name.
:param int nodes: number of nodes to execute the job.
:param str cmd: command line to run.
:param str or None job_preparation_cmd: Job preparation command line.
:param models.ContainerSettings or None container: container settings to run the job.
:return models.Job: the created job.
"""
job_preparation = None
if job_preparation_cmd:
job_preparation = models.JobPreparation(
command_line=job_preparation_cmd)
client.experiments.create(resource_group,
Helpers.DEFAULT_WORKSPACE_NAME,
Helpers.DEFAULT_EXPERIMENT_NAME).result()
return client.jobs.create(
resource_group,
Helpers.DEFAULT_WORKSPACE_NAME,
Helpers.DEFAULT_EXPERIMENT_NAME,
job_name,
parameters=models.JobCreateParameters(
cluster=models.ResourceId(id=cluster_id),
node_count=nodes,
std_out_err_path_prefix='$AZ_BATCHAI_MOUNT_ROOT/{0}'.format(
Helpers.AZURE_FILES_MOUNTING_PATH),
output_directories=[
models.OutputDirectory(
id=Helpers.JOB_OUTPUT_DIRECTORY_ID,
path_prefix=Helpers.JOB_OUTPUT_DIRECTORY_PATH,
path_suffix="files")
],
input_directories=[
models.InputDirectory(
id='INPUT',
path='$AZ_BATCHAI_MOUNT_ROOT/{0}/input'.format(
Helpers.AZURE_FILES_MOUNTING_PATH))
],
container_settings=container,
job_preparation=job_preparation,
custom_toolkit_settings=models.CustomToolkitSettings(
command_line=cmd))).result()
def submit_job(config, pretrained_model_type, retraining_type,
output_model_name, num_epochs):
''' Defines and submits a job. Does not check for completion. '''
client = get_client(config)
job_name = 'job{}'.format(
datetime.datetime.utcnow().strftime('%m_%d_%H_%M_%S'))
cluster = client.clusters.get(config.bait_resource_group_name,
config.bait_cluster_name)
# Define the command line arguments to the retraining script
command_line_args = '--input_dir $AZ_BATCHAI_INPUT_TRAININGDATA ' + \
'--validation_dir $AZ_BATCHAI_INPUT_VALIDATIONDATA ' + \
'--output_dir $AZ_BATCHAI_OUTPUT_MODEL ' + \
'--num_epochs {} '.format(num_epochs) + \
'--retraining_type {} '.format(retraining_type) + \
'--model_type {} '.format(pretrained_model_type) + \
'--model_filename $AZ_BATCHAI_INPUT_PRETRAINEDMODELS/'
if pretrained_model_type == 'alexnet':
command_line_args += 'AlexNet.model'
elif pretrained_model_type == 'resnet18':
command_line_args += 'ResNet_18.model'
# Define the job
cntk_settings = tm.CNTKsettings(
language_type='python',
python_script_file_path='$AZ_BATCHAI_INPUT_SCRIPT/' +
'retrain_model_distributed.py',
command_line_args=command_line_args,
process_count=config.bait_vms_per_job) # NC6s -- one GPU per VM
job_create_params = tm.job_create_parameters.JobCreateParameters(
location=config.bait_region,
cluster=tm.ResourceId(cluster.id),
node_count=config.bait_vms_per_job,
std_out_err_path_prefix='$AZ_BATCHAI_MOUNT_ROOT/afs',
output_directories=[
tm.OutputDirectory(id='MODEL',
path_prefix='$AZ_BATCHAI_MOUNT_ROOT/afs')
],
input_directories=[
tm.InputDirectory(id='SCRIPT',
path='$AZ_BATCHAI_MOUNT_ROOT/afs/scripts'),
tm.InputDirectory(
id='PRETRAINEDMODELS',
path='$AZ_BATCHAI_MOUNT_ROOT/afs/pretrainedmodels'),
tm.InputDirectory(
id='TRAININGDATA',
path='$AZ_BATCHAI_MOUNT_ROOT/nfs/training_images'),
tm.InputDirectory(
id='VALIDATIONDATA',
path='$AZ_BATCHAI_MOUNT_ROOT/nfs/validation_images')
],
cntk_settings=cntk_settings)
# Submit the job
job = client.jobs.create(
resource_group_name=config.bait_resource_group_name,
job_name=job_name,
parameters=job_create_params)
return (job_name)
DiscreteParameter(parameter_name="BATCH_SIZE", values=[8, 16, 32]),
DiscreteParameter(parameter_name="T", values=[72, 168, 336]),
DiscreteParameter(parameter_name="LEARNING_RATE",
values=[0.01, 0.001, 0.0001]),
DiscreteParameter(parameter_name="ALPHA", values=[0.1, 0.001, 0])
]
parameters = ParameterSweep(param_specs)
# create a template for Batch AI job
jcp = models.JobCreateParameters(
cluster=models.ResourceId(id=cluster.id),
node_count=1,
std_out_err_path_prefix='$AZ_BATCHAI_JOB_MOUNT_ROOT/logs',
output_directories=[
models.OutputDirectory(id='ALL',
path_prefix='$AZ_BATCHAI_JOB_MOUNT_ROOT/output')
],
custom_toolkit_settings=models.CustomToolkitSettings(
command_line=
'python $AZ_BATCHAI_JOB_MOUNT_ROOT/resources/scripts/FF_multi_step_multivariate.py \
--scriptdir $AZ_BATCHAI_JOB_MOUNT_ROOT/resources/scripts \
--datadir $AZ_BATCHAI_JOB_MOUNT_ROOT/resources/data \
--outdir $AZ_BATCHAI_OUTPUT_ALL \
-l {0} -n {1} -b {2} -T {3} -r {4} -a {5}'.format(
parameters['LATENT_DIM'], parameters['HIDDEN_LAYERS'],
parameters['BATCH_SIZE'], parameters['T'],
parameters['LEARNING_RATE'], parameters['ALPHA'])),
container_settings=models.ContainerSettings(
image_source_registry=models.ImageSourceRegistry(
image=cfg['docker_image'])),
mount_volumes=models.MountVolumes(azure_file_shares=[
location='northeurope',
cluster=models.ResourceId(id=cluster.id),
# The number of VMs in the cluster to use
node_count=1,
# Override the path where the std out and std err files will be written to.
# In this case we will write these out to an Azure Files share
std_out_err_path_prefix='$AZ_BATCHAI_MOUNT_ROOT/{0}'.format(relative_mount_point),
input_directories=[models.InputDirectory(
id='SAMPLE',
path='$AZ_BATCHAI_MOUNT_ROOT/{0}/data'.format(relative_mount_point))],
# Specify directories where files will get written to
output_directories=[models.OutputDirectory(
id='MODEL',
path_prefix='$AZ_BATCHAI_MOUNT_ROOT/{0}'.format(relative_mount_point),
path_suffix="Models")],
# Container configuration
container_settings=models.ContainerSettings(
image_source_registry=models.ImageSourceRegistry(image='microsoft/cntk:2.1-gpu-python3.5-cuda8.0-cudnn6.0')),
# Toolkit specific settings
cntk_settings = models.CNTKsettings(
python_script_file_path='$AZ_BATCHAI_INPUT_SAMPLE/ConvNet_MNIST.py',
command_line_args='$AZ_BATCHAI_INPUT_SAMPLE $AZ_BATCHAI_OUTPUT_MODEL')
)
# Create the job
client.jobs.create(resource_group_name, job_name, parameters).result()
def test_job_level_mounting(self, resource_group, location, cluster,
storage_account, storage_account_key):
"""Tests if it's possible to mount external file systems for a job."""
job_name = 'job'
# Create file share and container to mount on the job level
if storage_account.name != FAKE_STORAGE.name:
files = FileService(storage_account.name, storage_account_key)
files.create_share('jobshare', fail_on_exist=False)
blobs = BlockBlobService(storage_account.name, storage_account_key)
blobs.create_container('jobcontainer', fail_on_exist=False)
job = self.client.jobs.create(
resource_group.name,
job_name,
parameters=models.JobCreateParameters(
location=location,
cluster=models.ResourceId(id=cluster.id),
node_count=1,
mount_volumes=models.
MountVolumes(azure_file_shares=[
models.AzureFileShareReference(
account_name=storage_account.name,
azure_file_url='https://{0}.file.core.windows.net/{1}'.
format(storage_account.name, 'jobshare'),
relative_mount_path='job_afs',
credentials=models.AzureStorageCredentialsInfo(
account_key=storage_account_key),
)
],
azure_blob_file_systems=[
models.AzureBlobFileSystemReference(
account_name=storage_account.name,
container_name='jobcontainer',
relative_mount_path='job_bfs',
credentials=models.
AzureStorageCredentialsInfo(
account_key=storage_account_key),
)
]),
# Put standard output on cluster level AFS to check that the job has access to it.
std_out_err_path_prefix='$AZ_BATCHAI_MOUNT_ROOT/{0}'.format(
AZURE_FILES_MOUNTING_PATH),
# Create two output directories on job level AFS and blobfuse.
output_directories=[
models.OutputDirectory(
id='OUTPUT1',
path_prefix='$AZ_BATCHAI_JOB_MOUNT_ROOT/job_afs'),
models.OutputDirectory(
id='OUTPUT2',
path_prefix='$AZ_BATCHAI_JOB_MOUNT_ROOT/job_bfs')
],
# Check that the job preparation has access to job level file systems.
job_preparation=models.JobPreparation(
command_line=
'echo afs > $AZ_BATCHAI_OUTPUT_OUTPUT1/prep_afs.txt; '
'echo bfs > $AZ_BATCHAI_OUTPUT_OUTPUT2/prep_bfs.txt; '
'echo done'),
# Check that the job has access to job
custom_toolkit_settings=models.CustomToolkitSettings(
command_line=
'echo afs > $AZ_BATCHAI_OUTPUT_OUTPUT1/job_afs.txt; '
'echo bfs > $AZ_BATCHAI_OUTPUT_OUTPUT2/job_bfs.txt; '
'mkdir $AZ_BATCHAI_OUTPUT_OUTPUT1/afs; '
'echo afs > $AZ_BATCHAI_OUTPUT_OUTPUT1/afs/job_afs.txt; '
'mkdir $AZ_BATCHAI_OUTPUT_OUTPUT2/bfs; '
'echo bfs > $AZ_BATCHAI_OUTPUT_OUTPUT2/bfs/job_bfs.txt; '
'echo done'))).result()
self.assertEqual(
wait_for_job_completion(self.is_live, self.client,
resource_group.name, job.name, MINUTE),
models.ExecutionState.succeeded)
job = self.client.jobs.get(resource_group.name, job.name)
# Assert job and job prep standard output is populated on cluster level filesystem
assert_job_files_are(
self, self.client, resource_group.name, job.name,
STANDARD_OUTPUT_DIRECTORY_ID, {
u'stdout.txt': u'done\n',
u'stderr.txt': u'',
u'stdout-job_prep.txt': u'done\n',
u'stderr-job_prep.txt': u''
})
# Assert files are generated on job level AFS
assert_job_files_are(self, self.client, resource_group.name, job.name,
'OUTPUT1', {
u'job_afs.txt': u'afs\n',
u'prep_afs.txt': u'afs\n',
u'afs': None
})
# Assert files are generated on job level blobfuse
assert_job_files_are(self, self.client, resource_group.name, job.name,
'OUTPUT2', {
u'job_bfs.txt': u'bfs\n',
u'prep_bfs.txt': u'bfs\n',
u'bfs': None
})
# Assert subfolders are available via API
assert_job_files_in_path_are(self, self.client, resource_group.name,
job.name, 'OUTPUT1', 'afs',
{u'job_afs.txt': u'afs\n'})
assert_job_files_in_path_are(self, self.client, resource_group.name,
job.name, 'OUTPUT2', 'bfs',
{u'job_bfs.txt': u'bfs\n'})
# Assert that we can access the output files created on job level mount volumes directly in storage using path
# segment returned by the server.
if storage_account.name != FAKE_STORAGE.name:
files = FileService(storage_account.name, storage_account_key)
self.assertTrue(
files.exists(
'jobshare', job.job_output_directory_path_segment + '/' +
OUTPUT_DIRECTORIES_FOLDER_NAME, 'job_afs.txt'))
blobs = BlockBlobService(storage_account.name, storage_account_key)
self.assertTrue(
blobs.exists(
'jobcontainer', job.job_output_directory_path_segment +
'/' + OUTPUT_DIRECTORIES_FOLDER_NAME + '/job_bfs.txt'))
# After the job is done the filesystems should be unmounted automatically, check this by submitting a new job.
checker = self.client.jobs.create(
resource_group.name,
'checker',
parameters=models.JobCreateParameters(
location=location,
cluster=models.ResourceId(id=cluster.id),
node_count=1,
std_out_err_path_prefix='$AZ_BATCHAI_MOUNT_ROOT/{0}'.format(
AZURE_FILES_MOUNTING_PATH),
custom_toolkit_settings=models.CustomToolkitSettings(
command_line='echo job; df | grep -E "job_bfs|job_afs"'))
).result()
# Check the job failed because there are not job level mount volumes anymore
self.assertEqual(
wait_for_job_completion(self.is_live, self.client,
resource_group.name, checker.name, MINUTE),
models.ExecutionState.failed)
# Check that the cluster level AFS was still mounted
assert_job_files_are(self, self.client, resource_group.name,
checker.name, STANDARD_OUTPUT_DIRECTORY_ID, {
u'stdout.txt': u'job\n',
u'stderr.txt': u''
})