gcloud alpha ml-engine local train

gcloud alpha ml-engine local train - run an AI Platform training job locally

gcloud alpha ml-engine local train --module-name=MODULE_NAME [--distributed] [--evaluator-count=EVALUATOR_COUNT] [--job-dir=JOB_DIR] [--package-path=PACKAGE_PATH] [--parameter-server-count=PARAMETER_SERVER_COUNT] [--start-port=START_PORT; default=27182] [--worker-count=WORKER_COUNT] [GCLOUD_WIDE_FLAG …] [-- USER_ARGS …]

This is especially useful in the case of testing distributed models, as it allows you to validate that you are properly interacting with the AI Platform cluster configuration. If your model expects a specific number of parameter servers or workers (i.e. you expect to use the CUSTOM machine type), use the --parameter-server-count and --worker-count flags to further specify the desired cluster configuration, just as you would in your cloud training job configuration:

gcloud alpha ml-engine local train --module-name trainer.task --package-path /path/to/my/code/trainer --distributed --parameter-server-count 4 --worker-count 8

Unlike submitting a training job, the --package-path parameter can be omitted, and will use your current working directory.

AI Platform Training sets a TF_CONFIG environment variable on each VM in your training job. You can use TF_CONFIG to access the cluster description and the task description for each VM.

Learn more about TF_CONFIG: https://cloud.google.com/ai-platform/training/docs/distributed-training-details.

[-- USER_ARGS …]

Additional user arguments to be forwarded to user code. Any relative paths will be relative to the parent directory of --package-path. The '--' argument must be specified between gcloud specific args on the left and USER_ARGS on the right.

--module-name=MODULE_NAME

Name of the module to run.

--distributed

Runs the provided code in distributed mode by providing cluster configurations as environment variables to subprocesses

--evaluator-count=EVALUATOR_COUNT

Number of evaluators with which to run. Ignored if --distributed is not specified. Default: 0

--job-dir=JOB_DIR

Cloud Storage path or local_directory in which to store training outputs and other data needed for training.

This path will be passed to your TensorFlow program as the --job-dir command-line arg. The benefit of specifying this field is that AI Platform will validate the path for use in training. However, note that your training program will need to parse the provided --job-dir argument.

--package-path=PACKAGE_PATH

Path to a Python package to build. This should point to a local directory containing the Python source for the job. It will be built using setuptools (which must be installed) using its parent directory as context. If the parent directory contains a setup.py file, the build will use that; otherwise, it will use a simple built-in one.

--parameter-server-count=PARAMETER_SERVER_COUNT

Number of parameter servers with which to run. Ignored if --distributed is not specified. Default: 2

--start-port=START_PORT; default=27182

Start of the range of ports reserved by the local cluster. This command will use a contiguous block of ports equal to parameter-server-count + worker-count + 1.

If --distributed is not specified, this flag is ignored.

--worker-count=WORKER_COUNT

Number of workers with which to run. Ignored if --distributed is not specified. Default: 2

Run $ gcloud help for details.