gcloud alpha ai custom-jobs local-run - run a custom training locally
gcloud alpha ai custom-jobs local-run --executor-image-uri=IMAGE_URI [--extra-dirs=[EXTRA_DIR,...]] [--extra-packages=[PACKAGE,...]] [--gpu] [--local-package-path=LOCAL_PATH] [--output-image-uri=OUTPUT_IMAGE] [--requirements=[REQUIREMENTS,...]] [--service-account-key-file=ACCOUNT_KEY_FILE] [--python-module=PYTHON_MODULE | --script=SCRIPT] [GCLOUD_WIDE_FLAG ...] [-- ARGS ...]
(ALPHA) Packages your training code into a Docker image and executes it locally.
You should execute this command in the top folder which includes all the code and resources you want to pack and run, or specify the 'work-dir' flag to point to it. Any other path you specified via flags should be a relative path to the work-dir and under it; otherwise it will be unaccessible.
Supposing your directories are like the following structures:
/root - my_project - my_training - task.py - util.py - setup.py - other_modules - some_module.py - dataset - small.dat - large.dat - config - dep - foo.tar.gz - bar.whl - requirements.txt - another_project - something
If you set 'my_project' as the package, then you should execute the task.py by specifying "--script=my_training/task.py" or "--python-module=my_training.task", the 'requirements.txt' will be processed. And you will also be able to install extra packages by, e.g. specifying "--extra-packages=dep/foo.tar.gz,bar.whl" or include extra directories, e.g. specifying "--extra-dirs=dataset,config".
If you set 'my_training' as the package, then you should execute the task.py by specifying "--script=task.py" or "--python-module=task", the 'setup.py' will be processed. However, you won't be able to access any other files or directories that are not in 'my_training' folder.
See more details in the HELP info of the corresponding flags.
To execute an python module with required dependencies, run:
$ gcloud alpha ai custom-jobs local-run \ --python-module=my_training.task \ --executor-image-uri=gcr.io/my/image \ --requirements=pandas,scipy>=1.3.0
To execute a python script using local GPU, run:
$ gcloud alpha ai custom-jobs local-run \ --script=my_training/task.py \ --executor-image-uri=gcr.io/my/image --gpu
To execute an arbitrary script with custom arguments, run:
$ gcloud alpha ai custom-jobs local-run --script=my_run.sh \ --executor-image-uri=gcr.io/my/image -- --my-arg bar \ --enable_foo
To run an existing container training without building new image, run:
$ gcloud alpha ai custom-jobs local-run \ --executor-image-uri=gcr.io/my/custom-training-image
- [-- ARGS ...]
Additional user arguments to be forwarded to your application.
The '--' argument must be specified between gcloud specific args on the left and ARGS on the right. Example:
$ gcloud alpha ai custom-jobs local-run --script=my_run.sh \ --base-image=gcr.io/my/image -- --my-arg bar --enable_foo
- --executor-image-uri=IMAGE_URI
URI or ID of the container image in either the Container Registry or local that will run the application. See https://cloud.google.com/vertex-ai/docs/training/pre-built-containers for available pre-built container images provided by Vertex AI for training.
- --extra-dirs=[EXTRA_DIR,...]
Extra directories under the working directory to include, besides the one that contains the main executable.
By default, only the parent directory of the main script or python module is copied to the container. For example, if the module is "training.task" or the script is "training/task.py", the whole "training" directory, including its sub-directories, will always be copied to the container. You may specify this flag to also copy other directories if necessary.
Note: if no parent is specified in 'python_module' or 'scirpt', the whole working directory is copied, then you don't need to specify this flag.
- --extra-packages=[PACKAGE,...]
Local paths to Python archives used as training dependencies in the image container. These can be absolute or relative paths. However, they have to be under the work_dir; Otherwise, this tool will not be able to access it.
Example: 'dep1.tar.gz, ./downloads/dep2.whl'
- --gpu
Enable to use GPU.
- --local-package-path=LOCAL_PATH
local path of the directory where the python-module or script exists. If not specified, it use the directory where you run the this command.
Only the contents of this directory will be accessible to the built container image.
- --output-image-uri=OUTPUT_IMAGE
Uri of the custom container image to be built with the your application packed in.
- --requirements=[REQUIREMENTS,...]
Python dependencies from PyPI to be used when running the application. If this is not specified, and there is no "setup.py" or "requirements.txt" in the working directory, your application will only have access to what exists in the base image with on other dependencies.
Example: 'tensorflow-cpu, pandas==1.2.0, matplotlib>=3.0.2'
- --service-account-key-file=ACCOUNT_KEY_FILE
The JSON file of a Google Cloud service account private key. When specified, the corresponding service account will be used to authenticate the local container to access Google Cloud services. Note that the key file won't be copied to the container, it will be mounted during running time.
- At most one of these can be specified:
- --python-module=PYTHON_MODULE
Name of the python module to execute, in 'trainer.train' or 'train' format. Its path should be relative to the work_dir.
- --script=SCRIPT
The relative path of the file to execute. Accepets a Python file or an arbitrary bash script. This path should be relative to the work_dir.
These flags are available to all commands: --access-token-file, --account, --billing-project, --configuration, --flags-file, --flatten, --format, --help, --impersonate-service-account, --log-http, --project, --quiet, --trace-token, --user-output-enabled, --verbosity.
Run $ gcloud help for details.
This command is currently in alpha and might change without notice. If this command fails with API permission errors despite specifying the correct project, you might be trying to access an API with an invitation-only early access allowlist. These variants are also available:
$ gcloud ai custom-jobs local-run
$ gcloud beta ai custom-jobs local-run