def main():
logging.warning("Loading environment variables...")
e = Env()
e.load_environment_variables(env_file_path='local.env',
fallback_to_os=True)
# Get Azure machine learning workspace
logging.warning(
"Getting reference to existing Azure Machine Learning workspace...")
auth = InteractiveLoginAuthentication(tenant_id=e.tenant_id)
ws = get_workspace(e.workspace_name, auth, e.subscription_id,
e.resource_group)
# Get compute target. It has to be a GPU compute as such unit is requested by the 'Feature Extraction Inference' step
compute_target = get_compute_target(ws,
compute_name=e.gpu_compute_name,
vm_size=e.gpu_vm_size)
# Create run configuration
run_config = create_run_configuration(ws)
# -------
# Step 1
# -------
# Define input 'prepared datasets'
input_prepared_datasets = []
experiment_configuration = ExperimentConfigurationWrapper()
experiment_configuration.load(
os.path.join(cfg.StepsStructure.SNAPSHOT_ROOT_DIR,
cfg.StepsStructure.get_experiments_config_filepath())
) # noqa: E501
for data_config in experiment_configuration.json['OBJECT_DETECTION'][
'inference']['data']:
dataset_name = data_config['input']['dataset_name']
dataset = ws.datasets.get(dataset_name)
input_prepared_datasets.extend([dataset.as_named_input(dataset_name)])
# Create pipeline datastore objects to create links between steps, so they are executed in a sequence, not in parallel
pipeline_datastore = ws.get_default_datastore()
object_detection_inference_output = PipelineData(
name="centers", datastore=pipeline_datastore, is_directory=True)
step_object_detection_inference = PythonScriptStep(
name="Object Detection - Inference",
source_directory=cfg.StepsStructure.SNAPSHOT_ROOT_DIR,
script_name=cfg.StepsStructure.ObjectDetection.
INFERENCE_STEP_SCRIPT_PATH,
arguments=[
'--subscription_id', e.subscription_id, '--resource_group',
e.resource_group, '--workspace_name', e.workspace_name,
'--experiments_config_filepath',
cfg.StepsStructure.get_experiments_config_filepath(),
'--model_name', cfg.MLModelNames.OBJECT_DETECTION_MODEL,
'--model_version',
cfg.MLModelNames.OBJECT_DETECTION_MODEL_BEST_VERSION,
'--output_folder', object_detection_inference_output,
'--should_register_dataset', True
],
inputs=input_prepared_datasets,
outputs=[object_detection_inference_output],
compute_target=compute_target,
runconfig=run_config,
allow_reuse=True)
# -------
# Step 2
# -------
# input should contain 'prepared datasets' and centers
object_extraction_input = object_detection_inference_output.as_input(
'centers')
object_extraction_inputs = [object_extraction_input]
object_extraction_output = PipelineData(name="cropped_objects",
datastore=pipeline_datastore,
is_directory=True)
step_object_extraction = PythonScriptStep(
name="Object Extraction",
source_directory=cfg.StepsStructure.SNAPSHOT_ROOT_DIR,
script_name=cfg.StepsStructure.ObjectExtraction.STEP_SCRIPT_PATH,
arguments=[
"--subscription_id",
e.subscription_id,
"--resource_group",
e.resource_group,
"--workspace_name",
e.workspace_name,
"--experiments_config_filepath",
cfg.StepsStructure.get_experiments_config_filepath(),
"--output_folder",
object_extraction_output,
"--should_register_dataset",
True,
# This flag might be handy when we really want to recreate a cropped objects dataset (e.g. changed implementation
# of the NucleiExtractor, although there are no changes in the input datasets).
"--force_dataset_recreation",
True
],
inputs=object_extraction_inputs,
outputs=[object_extraction_output],
compute_target=compute_target,
runconfig=run_config,
allow_reuse=True,
)
# -------
# Step 3a
# -------
step_object_images_upload = PythonScriptStep(
name="Cropped Object Images Upload to Blob Storage",
source_directory=cfg.StepsStructure.SNAPSHOT_ROOT_DIR,
script_name=cfg.StepsStructure.ObjectImagesUpload.STEP_SCRIPT_PATH,
arguments=[
# '--subscription_id', e.subscription_id,
# '--resource_group', e.resource_group,
# '--workspace_name', e.workspace_name,
'--experiments_config_filepath',
cfg.StepsStructure.get_experiments_config_filepath(),
# '--model_name', cfg.MLModelNames.FEATURE_EXTRACTION_MODEL,
# '--model_version', cfg.MLModelNames.FEATURE_EXTRACTION_MODEL_BEST_VERSION,
# '--output_folder', feature_extraction_inference_output,
# '--should_register_dataset', True
],
inputs=[object_extraction_output.as_input('cropped_objects')],
outputs=[],
compute_target=compute_target,
runconfig=run_config,
allow_reuse=True)
# -------
# Step 3b
# -------
feature_extraction_inference_input = object_extraction_output.as_input(
'cropped_objects')
feature_extraction_inference_inputs = [feature_extraction_inference_input]
feature_extraction_inference_output = PipelineData(
name="latent_dims", datastore=pipeline_datastore, is_directory=True)
step_feature_extraction_inference = PythonScriptStep(
name="Feature Extraction - Inference",
source_directory=cfg.StepsStructure.SNAPSHOT_ROOT_DIR,
script_name=cfg.StepsStructure.FeatureExtraction.
INFERENCE_STEP_SCRIPT_PATH,
arguments=[
'--subscription_id', e.subscription_id, '--resource_group',
e.resource_group, '--workspace_name', e.workspace_name,
'--experiments_config_filepath',
cfg.StepsStructure.get_experiments_config_filepath(),
'--model_name', cfg.MLModelNames.FEATURE_EXTRACTION_MODEL,
'--model_version',
cfg.MLModelNames.FEATURE_EXTRACTION_MODEL_BEST_VERSION,
'--output_folder', feature_extraction_inference_output,
'--should_register_dataset', True
],
inputs=feature_extraction_inference_inputs,
outputs=[feature_extraction_inference_output],
compute_target=compute_target,
runconfig=run_config,
allow_reuse=True)
# -------
# Pipeline composition
# -------
pipeline_steps = [
step_object_detection_inference, step_object_extraction,
step_object_images_upload, step_feature_extraction_inference
]
pipeline = Pipeline(workspace=ws, steps=pipeline_steps)
# Create and submit an experiment
logging.warning("Submitting experiment...")
experiment = Experiment(ws, cfg.ExperimentNames.INFERENCE_REMOTE)
experiment.submit(
pipeline, regenerate_outputs=False) # Allow data reuse for this run
logging.warning('Experiment submitted!')
"--input",
ds_input.as_download(), "--output",
prepared_data
],
inputs=[ds_input],
outputs=[prepared_data],
compute_target=compute_target,
runconfig=aml_run_config,
allow_reuse=True)
# Step 2: Train our model
training_results = PipelineData(name="training_results",
datastore=blob_store)
train_step = PythonScriptStep(script_name=train_entry_point,
source_directory=train_source_dir,
arguments=[
"--prepped_data",
prepared_data.as_input(),
"--training_results", training_results
],
compute_target=compute_target,
runconfig=aml_run_config,
allow_reuse=True)
# Step 3: Build the pipeline
pipeline_steps = [data_prep_step, train_step]
pipeline = Pipeline(workspace=workspace, steps=[pipeline_steps])
pipeline_run1 = Experiment(workspace, "mckee_bot").submit(pipeline)
pipeline_run1.wait_for_completion()