Proxy 工作設計

透過集合功能整理內容 你可以依據偏好儲存及分類內容。

from google.cloud.visionsolutions.nas.proxy_task import proxy_task_utils

if proxy_task_utils.get_stop_training(
    model_dir,
    end_training_cycle_step=<last-training-step-idx done so far>,
    total_training_steps=<total-training-steps>):
  break

proxy_task_utils.update_trial_training_accuracy_metric(
      model_dir=model_dir,
      accuracy=<latest accuracy value>,
      begin_training_cycle_step=<beginning training step for this cycle>,
      end_training_cycle_step=<end training step for this cycle>,
      training_cycle_time_in_secs=<training cycle time (excluding validation)>,
      total_training_steps=<total-training-steps>)

proxy_task_utils.update_trial_training_latency_metric(
          model_dir=model_dir,
          latency=<measured_latency>)

parser.add_argument(
      "--retrain_use_search_job_checkpoint",
      type=cloud_nas_utils.str_2_bool,
      default=False,
      help="True to use previous NAS search job checkpoint."
  )

if FLAGS.retrain_use_search_job_checkpoint:
    prev_checkpoint_dir = cloud_nas_utils.get_retrain_search_job_model_dir(
        retrain_search_job_trials=FLAGS.retrain_search_job_trials,
        retrain_search_job_dir=FLAGS.retrain_search_job_dir)
    logging.info("Setting checkpoint to %s.", prev_checkpoint_dir)
    # Now set your checkpoint using 'prev_checkpoint_dir'.

DATE="$(date '+%Y%m%d_%H%M%S')"
project_id= project-id>
# You can choose any unique docker id below.
trainer_docker_id=${USER}_trainer_${DATE}
trainer_docker_file= to your trainer dockerfile>
region= job region such as 'us-central1'>
search_space_module= to your search space module>
accelerator_type="NVIDIA_TESLA_V100"
num_gpus=2
# Your bucket should be for your project and in the same region as the job.
root_output_dir=

####### Variance measurement related parameters ######
proxy_task_variance_measurement_docker_id=${USER}_variance_measurement_${DATE}
# Use the service account that you set-up for your project.
service_account= service account>
job_name= job name>
############################################################

python3 vertex_nas_cli.py build \
--project_id=${project_id} \
--region=${region} \
--trainer_docker_id=${trainer_docker_id} \
--trainer_docker_file=${trainer_docker_file} \
--proxy_task_variance_measurement_docker_id=${proxy_task_variance_measurement_docker_id}

# The command below passes 'dummy' arguments for the training docker.
# You need to modify them for your own docker.
python3 vertex_nas_cli.py measure_proxy_task_variance \
--proxy_task_variance_measurement_docker_id=${proxy_task_variance_measurement_docker_id} \
--project_id=${project_id} \
--service_account=${service_account} \
--region=${region} \
--trainer_docker_id=${trainer_docker_id} \
--job_name=${job_name} \
--search_space_module=${search_space_module} \
--accelerator_type=${accelerator_type} \
--num_gpus=${num_gpus} \
--root_output_dir=${root_output_dir} \
--search_docker_flags \
dummy_trainer_flag1="dummy_trainer_val"

= (30*1/30 + 25*1/25 + 20*1/20 + 15*1/15 + 10*(remaining-training-time-fraction)) * full-training-time
= (4 + 10*(0.81)) * full-training-time
~= 12 * full-training-time

DATE="$(date '+%Y%m%d_%H%M%S')"
project_id= project-id>
# You can choose any unique docker id below.
trainer_docker_id=${USER}_trainer_${DATE}
trainer_docker_file= to your trainer dockerfile>
latency_calculator_docker_id=${USER}_model_selection_${DATE}
latency_calculator_docker_file=${USER}_latency_${DATE}
region= job region such as 'us-central1'>
search_space_module= to your search space module>
accelerator_type="NVIDIA_TESLA_V100"
num_gpus=2
# Your bucket should be for your project and in the same region as the job.
root_output_dir=
# Your latency computation device.
target_device_type="CPU"

####### Proxy task model-selection related parameters ######
proxy_task_model_selection_docker_id=${USER}_model_selection_${DATE}
# Use the service account that you set-up for your project.
service_account= service account>
job_name= job name>
# The value below depends on your accelerator quota. By default
# the model-selection job runs 30 trials. However, depending on
# your quota, you can choose to only run 10 trials in parallel at a time.
# However, lowering this number can increase the overall runtime for the job.
max_parallel_nas_trial= parallel trials>
# The value below is the 'metric-id' corresponding to the accuracy ONLY
# metric reported by your trainer. Note that this metric may
# be different from the 'reward'.
accuracy_metric_id=<set accuracy metric id used by your trainer>
latency_metric_id=<set latency metric id used by your trainer>
############################################################

python3 vertex_nas_cli.py build \
--project_id=${project_id} \
--region=${region} \
--trainer_docker_id=${trainer_docker_id} \
--trainer_docker_file=${trainer_docker_file} \
--latency_calculator_docker_id=${latency_calculator_docker_id} \
--latency_calculator_docker_file=${latency_calculator_docker_file} \
--proxy_task_model_selection_docker_id=${proxy_task_model_selection_docker_id}

# The command below passes 'dummy' arguments for trainer-docker
# and latency-docker. You need to modify them for your own docker.
python3 vertex_nas_cli.py select_proxy_task_models \
--service_account=${service_account} \
--proxy_task_model_selection_docker_id=${proxy_task_model_selection_docker_id} \
--project_id=${project_id} \
--region=${region} \
--trainer_docker_id=${trainer_docker_id} \
--job_name=${job_name} \
--max_parallel_nas_trial=${max_parallel_nas_trial} \
--accuracy_metric_id=${accuracy_metric_id} \
--latency_metric_id=${latency_metric_id} \
--search_space_module=${search_space_module} \
--accelerator_type=${accelerator_type} \
--num_gpus=${num_gpus} \
--root_output_dir=${root_output_dir} \
--latency_calculator_docker_id=${latency_calculator_docker_id} \
--latency_docker_flags \
dummy_latency_flag1="dummy_latency_val" \
--target_device_type=${target_device_type} \
--search_docker_flags \
dummy_trainer_flag1="dummy_trainer_val"

gcloud storage cat gs:// to 'model_selection_dir'>/MODEL_SELECTION_STATE.json

{
  "start_num_models": 30,
  "final_num_models": 10,
  "num_models_to_remove_per_iter": 5,
  "accuracy_metric_id": "top_1_accuracy_without_latency",
  "latency_metric_id": "latency_milli_seconds",
  "iterations": [
    {
      "num_trials": 30,
      "trials_to_retrain": [
        "27",
        "16",
        ...,
        "14"
      ],
      "search_job_name": "projects/123456/locations/europe-west4/nasJobs/2111217356469436416",
      "search_job_link": "https://console.cloud.google.com/vertex-ai/locations/europe-west4/training/2111217356469436416/cpu?project=my-project",
      "latency_calculator_job_name": "projects/123456/locations/europe-west4/customJobs/6909239809479278592",
      "latency_calculator_job_link": "https://console.cloud.google.com/vertex-ai/locations/europe-west4/training/6909239809479278592/cpu?project=my-project",
      "desired_training_step_pct": 2.0
    },
    ...,
    {
      "num_trials": 15,
      "trials_to_retrain": [
        "14",
        ...,
        "5"
      ],
      "search_job_name": "projects/123456/locations/europe-west4/nasJobs/7045544066951413760",
      "search_job_link": "https://console.cloud.google.com/vertex-ai/locations/europe-west4/training/7045544066951413760/cpu?project=my-project",
      "latency_calculator_job_name": "projects/123456/locations/europe-west4/customJobs/2790768318993137664",
      "latency_calculator_job_link": "https://console.cloud.google.com/vertex-ai/locations/europe-west4/training/2790768318993137664/cpu?project=my-project",
      "desired_training_step_pct": 28.57936507936508
    },
    {
      "num_trials": 10,
      "trials_to_retrain": [],
      "search_job_name": "projects/123456/locations/europe-west4/nasJobs/2742864796394192896",
      "search_job_link": "https://console.cloud.google.com/vertex-ai/locations/europe-west4/training/2742864796394192896/cpu?project=my-project",
      "latency_calculator_job_name": "projects/123456/locations/europe-west4/customJobs/1490864099985195008",
      "latency_calculator_job_link": "https://console.cloud.google.com/vertex-ai/locations/europe-west4/training/1490864099985195008/cpu?project=my-project",
      "desired_training_step_pct": 101.0
    }
  ]
}

# MNasnet training-data size choices.
MNASNET_TRAINING_DATA_PCT_LIST = [25, 50, 75, 95]

# Training data path regex pattern.
_TRAINING_DATA_PATH_REGEX = r"gs://.*/.*"


def update_mnasnet_proxy_training_data(
    baseline_docker_args_map: Dict[str, Any],
    training_data_pct: int) -> Optional[Dict[str, Any]]:
  """Updates MNasnet baseline docker to use a certain training_data_pct."""
  proxy_task_docker_args_map = copy.deepcopy(baseline_docker_args_map)
  # Imagenet training data path looks like:
  # gs:///train-00[0-7]??-of-01024.
  if not re.match(_TRAINING_DATA_PATH_REGEX,
                  baseline_docker_args_map["training_data_path"]):
    raise ValueError(
        "Training data path %s does not match the desired pattern." %
        baseline_docker_args_map["training_data_path"])

  root_path, _ = baseline_docker_args_map["training_data_path"].rsplit("/", 1)
  if training_data_% == 25:
    proxy_task_docker_args_map["training_data_path"] = os.path.join(
        root_path, "train-00[0-1][0-4]?-of-01024*")
  elif training_data_% == 50:
    proxy_task_docker_args_map["training_data_path"] = os.path.join(
        root_path, "train-00[0-4]??-of-01024*")
  elif training_data_% == 75:
    proxy_task_docker_args_map["training_data_path"] = os.path.join(
        root_path, "train-00[0-6][0-4]?-of-01024*")
  elif training_data_% == 95:
    proxy_task_docker_args_map["training_data_path"] = os.path.join(
        root_path, "train-00[0-8][0-4]?-of-01024*")
  else:
    logging.warning("Mnasnet training_data_% %d is not supported.",
                    training_data_pct)
    return None
  proxy_task_docker_args_map["validation_data_path"] = os.path.join(
      root_path, "train-009[0-4]?-of-01024")
  return proxy_task_docker_args_map


def mnasnet_proxy_task_config_generator(
    baseline_docker_args_map: Dict[str, Any]
) -> List[proxy_task_utils.ProxyTaskConfig]:
  """Returns a list of proxy-task configs to be evaluated for MNasnet.

  Args:
    baseline_docker_args_map: A set of baseline training-docker arguments in
      the form of a dictionary of {'key', val}. The different proxy-task
      configs to try can be built by modifying this baseline.

  Returns:
    A list of proxy-task configs to be evaluated for this
    proxy-task search space.
  """
  proxy_task_config_list = []
  # NOTE: Will not search over model-scale for MNasnet.
  for training_data_% in MNASNET_TRAINING_DATA_PCT_LIST:
    proxy_task_docker_args_map = update_mnasnet_proxy_training_data(
        baseline_docker_args_map=baseline_docker_args_map,
        training_data_pct=training_data_pct)
    if not proxy_task_docker_args_map:
      continue
    proxy_task_name = "mnasnet_proxy_training_data_pct_{}".format(
        training_data_pct)
    proxy_task_config_list.append(
        proxy_task_utils.ProxyTaskConfig(
            name=proxy_task_name, docker_args_map=proxy_task_docker_args_map))
  return proxy_task_config_list

DATE="$(date '+%Y%m%d_%H%M%S')"
project_id= project-id>
# You can choose any unique docker id below.
trainer_docker_id=${USER}_trainer_${DATE}
trainer_docker_file= to your trainer dockerfile>
latency_calculator_docker_id=${USER}_model_selection_${DATE}
latency_calculator_docker_file=${USER}_latency_${DATE}
region= job region such as 'us-central1'>
search_space_module= to your NAS job search space module>
accelerator_type="NVIDIA_TESLA_V100"
num_gpus=2
# Your bucket should be for your project and in the same region as the job.
root_output_dir=
# Your latency computation device.
target_device_type="CPU"

####### Proxy task search related parameters ######
proxy_task_search_controller_docker_id=${USER}_proxy_task_search_${DATE}
job_name= job name>
# Path to your proxy task search space definition. For ex:
# 'proxy_task.proxy_task_search_spaces.mnasnet_proxy_task_config_generator'
proxy_task_config_generator_module= to your proxy task config generator module>
# The previous model-slection job provides the candidate-correlation-models
# and their scores.
proxy_task_model_selection_job_id= job id of your previous model-selection>
# During proxy-task search, the proxy-task training is stopped
# when the following correlation score is achieved.
desired_accuracy_correlation=0.65
# During proxy-task search, the proxy-task training is stopped
# if the runtime exceeds this limit: 4 hrs.
training_time_hrs_limit=4
# The proxy-task is marked a good candidate only if the latency
# correlation is also above the required threshold.
# Note: This won't be used if you do not have a latency job.
desired_latency_correlation=0.65
# Early stop a proxy-task evaluation if you already have a better candidate.
# If False, evaluate all proxy-taask candidates.
early_stop_proxy_task_if_not_best=False
# Use the service account that you set-up for your project.
service_account= service account>
###################################################


python3 vertex_nas_cli.py build \
--project_id=${project_id} \
--region=${region} \
--trainer_docker_id=${trainer_docker_id} \
--trainer_docker_file=${trainer_docker_file} \
--latency_calculator_docker_id=${latency_calculator_docker_id} \
--latency_calculator_docker_file=${latency_calculator_docker_file} \
--proxy_task_search_controller_docker_id=${proxy_task_search_controller_docker_id}

# The command below passes 'dummy' arguments for trainer-docker
# and latency-docker. You need to modify them for your own docker.
python3 vertex_nas_cli.py search_proxy_task \
--service_account=${service_account} \
--proxy_task_search_controller_docker_id=${proxy_task_search_controller_docker_id} \
--proxy_task_config_generator_module=${proxy_task_config_generator_module} \
--proxy_task_model_selection_job_id=${proxy_task_model_selection_job_id} \
--proxy_task_model_selection_job_region=${region} \
--desired_accuracy_correlation={$desired_accuracy_correlation}\
--training_time_hrs_limit=${training_time_hrs_limit} \
--desired_latency_correlation=${desired_latency_correlation} \
--early_stop_proxy_task_if_not_best=${early_stop_proxy_task_if_not_best} \
--project_id=${project_id} \
--region=${region} \
--trainer_docker_id=${trainer_docker_id} \
--job_name=${job_name} \
--search_space_module=${search_space_module} \
--accelerator_type=${accelerator_type} \
--num_gpus=${num_gpus} \
--root_output_dir=${root_output_dir} \
--latency_calculator_docker_id=${latency_calculator_docker_id} \
--latency_docker_flags \
dummy_latency_flag1="dummy_latency_val" \
--target_device_type=${target_device_type} \
--search_docker_flags \
dummy_trainer_flag1="dummy_trainer_val"

gcloud storage cat gs:// to 'search_controller_dir'>/SEARCH_CONTROLLER_STATE.json

{
  "proxy_tasks_map": {
    "mnasnet_proxy_training_data_pct_25": {
      "proxy_task_stats": {
        "training_steps": [
          1249,
          2499,
          ...,
          18749
        ],
        "accuracy_correlation_over_step": [
          -0.06666666666666667,
          -0.6,
          ...,
          0.7857142857142856
        ],
        "accuracy_correlation_p_value_over_step": [
          0.8618005952380953,
          0.016666115520282188,
          ...,
          0.005505952380952381
        ],
        "median_accuracy_over_step": [
          0.011478611268103123,
          0.04956454783678055,
          ...,
          0.32932570576667786
        ],
        "median_training_time_hrs_over_step": [
          0.11611097933475001,
          0.22913257125276987,
          ...,
          1.6682701704073444
        ],
        "latency_correlation": 0.9555555555555554,
        "latency_correlation_p_value": 5.5114638447971785e-06,
        "stopping_state": "Met desired correlation",
        "posted_stop_trials_message": true,
        "final_training_time_in_hours": 1.6675102778428197,
        "final_training_steps": 18512
      },
      "proxy_task_name": "mnasnet_proxy_training_data_pct_25",
      "search_job_name": "projects/123456/locations/europe-west4/nasJobs/4173661476642357248",
      "search_job_link": "https://console.cloud.google.com/vertex-ai/locations/europe-west4/training/4173661476642357248/cpu?project=my-project",
      "latency_calculator_job_name": "projects/123456/locations/europe-west4/customJobs/8785347495069745152",
      "latency_calculator_job_link": "https://console.cloud.google.com/vertex-ai/locations/europe-west4/training/8785347495069745152/cpu?project=my-project"
    },
    ...,
    "mnasnet_proxy_training_data_pct_95": {
      "proxy_task_stats": {
        "training_steps": [
          1249,
          ...,
          18749
        ],
        "accuracy_correlation_over_step": [
          -0.3333333333333333,
          ...,
          0.7857142857142856,
          -5.0
        ],
        "accuracy_correlation_p_value_over_step": [
          0.21637345679012346,
          ...,
          0.005505952380952381,
          -5.0
        ],
        "median_accuracy_over_step": [
          0.01120645459741354,
          ...,
          0.38238024711608887,
          -1.0
        ],
        "median_training_time_hrs_over_step": [
          0.11385884770307843,
          ...,
          1.5466042930547819,
          -1.0
        ],
        "latency_correlation": 0.9555555555555554,
        "latency_correlation_p_value": 5.5114638447971785e-06,
        "stopping_state": "Met desired correlation",
        "posted_stop_trials_message": true,
        "final_training_time_in_hours": 1.533235285929564,
        "final_training_steps": 17108
      },
      "proxy_task_name": "mnasnet_proxy_training_data_pct_95",
      "search_job_name": "projects/123456/locations/europe-west4/nasJobs/2341822328209408000",
      "search_job_link": "https://console.cloud.google.com/vertex-ai/locations/europe-west4/training/2341822328209408000/cpu?project=my-project",
      "latency_calculator_job_name": "projects/123456/locations/europe-west4/customJobs/7575005095213924352",
      "latency_calculator_job_link": "https://console.cloud.google.com/vertex-ai/locations/europe-west4/training/7575005095213924352/cpu?project=my-project"
    }
  },
  "best_proxy_task_name": "mnasnet_proxy_training_data_pct_75"
}

gcloud storage cp gs:// to 'search_controller_dir'>/* /your/local/dir