Vertex AI Neural Architecture Search client library

This document describes the Vertex AI Neural Architecture Search client library.

Neural Architecture Search client (in vertex_nas_cli.py) wraps the job management API and facilitates the Neural Architecture Search development. It provides the following subcommands:

  • vertex_nas_cli.py build: builds Neural Architecture Search containers and pushes to Google Cloud container registry.
  • vertex_nas_cli.py run_latency_calculator_local: runs latency calculator locally for Neural Architecture Search stage-1 search job.
  • vertex_nas_cli.py search_in_local: runs Neural Architecture Search job locally on your machine with a randomly sampled architecture.
  • vertex_nas_cli.py search: runs Neural Architecture Search job with stage-1 search and stage-2 training on Google Cloud.
  • vertex_nas_cli.py search_resume: resumes a previous Neural Architecture Search job on Google Cloud.
  • vertex_nas_cli.py list_trials: lists Neural Architecture Search trials for specific job.
  • vertex_nas_cli.py train: trains searched model architecture (trial) in Google Cloud.

Build

Run the following command to see the list of arguments supported by vertex_nas_cli.py build:

python3 vertex_nas_cli.py build -h

If --trainer_docker_id is specified, it builds the trainer docker from the docker file specified by the flag --trainer_docker_file. The docker is built with full URI gcr.io/project_id/trainer_docker_id and pushed to Google Cloud container registry.

If --latency_calculator_docker_id is specified, it builds the latency calculator docker from the docker file specified by the flag --latency_calculator_docker_file. The docker is built with full URI gcr.io/project_id/latency_calculator_docker_id and pushed to Google Cloud container registry.

Instead of building with Dockerfile, you can also use other tools like bazel to build the trainer, and use it with the Neural Architecture Search service.

Run latency calculator local

Run the following command to see the list of arguments supported by vertex_nas_cli.py run_latency_calculator_local:

python3 vertex_nas_cli.py run_latency_calculator_local -h

Search in local

Run the following command to see the list of arguments supported by vertex_nas_cli.py search_in_local:

python3 vertex_nas_cli.py search_in_local -h

You need to specify either --search_space_module or --prebuilt_search_space so that vertex_nas_cli.py internally generates a random model architecture to use.

This command will run the docker gcr.io/project_id/trainer_docker_id:latest on your local machine with a randomly sampled architecture.

You can pass through the flags to be used by the container after --search_docker_flags. For example, you can pass through --training_data_path and validation_data_path to the container:

python3 vertex_nas_cli.py search_in_local \
--project_id=${PROJECT_ID} \
--trainer_docker_id=${TRAINER_DOCKER_ID} \
--prebuilt_search_space=spinenet \
--use_prebuilt_trainer=True \
--local_output_dir=${JOB_DIR} \
--search_docker_flags \
training_data_path=/test_data/test-coco.tfrecord \
validation_data_path=/test_data/test-coco.tfrecord \
model=retinanet

Run the following command to see the list of arguments supported by vertex_nas_cli.py search:

python3 vertex_nas_cli.py search -h

You need to specify either --search_space_module or --prebuilt_search_space so that vertex_nas_cli.py internally creates search_space_spec.

The machines to run Neural Architecture Search jobs can be specified by --accelerator_type. For more information or to customize for your own needs, like using more GPUs, see add_machine_configurations.

Use the flags with prefix train_ to set the stage-2 training related parameters.

Search Resume

Run the following command to see the list of arguments supported by vertex_nas_cli.py search_resume:

python3 vertex_nas_cli.py search_resume -h

You can resume a previously run search job by passing previous_nas_job_id and optionally previous_latency_job_id. The previous_latency_job_id flag is needed only if your previous search job involved a Google Cloud latency job. If instead of a Google Cloud latency job you used an on-premises latency calculator, then you have to run that on-premises latency calculator job separately again. The previous search job should not itself be a resume job. The region for the search resume job should be the same as for the previous search job. An example search_resume command looks like the following:

python3 vertex_nas_cli.py search_resume \
  --project_id=${PROJECT} \
  --region=${REGION} \
  --job_name="${JOB_NAME}" \
  --previous_nas_job_id=${previous_nas_job_id} \
  --previous_latency_job_id=${previous_latency_job_id} \
  --root_output_dir=${GCS_ROOT_DIR} \
  --max_nas_trial=2 \
  --max_parallel_nas_trial=2 \
  --max_failed_nas_trial=2

List trials

Run the following command to see the list of arguments supported by vertex_nas_cli.py list_trials:

python3 vertex_nas_cli.py list_trials -h

Train

Run the following command to see the list of arguments supported by vertex_nas_cli.py train:

python3 vertex_nas_cli.py train -h

Proxy-task variance measurement

Run the following command to see the list of arguments supported by vertex_nas_cli.py measure_proxy_task_variance:

python3 vertex_nas_cli.py measure_proxy_task_variance -h

Proxy-task model selection

Run the following command to see the list of arguments supported by vertex_nas_cli.py select_proxy_task_models:

python3 vertex_nas_cli.py select_proxy_task_models -h

Run the following command to see the list of arguments supported by vertex_nas_cli.py search_proxy_task:

python3 vertex_nas_cli.py search_proxy_task -h