/aiplatgo

CLI wrapper for gcloud ai-platform

Primary LanguagePython

aiplatgo

A convince wrapper for:

Because ...

"For convenience, it's useful to define your configuration values as environment variables. ... Even if you use a YAML file, certain details must be supplied as command-line flags."

is silly.

aiplatgo's config file to replaces the combination of CLI-flags and config file natively offered by gcloud ai platform. Moreover:

  • The config file can contain all possible flags for training or prediction local or platform. Only relevant flags will be submitted .
  • A parser for parsing user arguments passed through the CLI.
  • There are configurations to automatically handle naming conventions
  • platform job directories can automatically be prefixed with gs://bucket/folder
  • Underscores are accepted: ie arg_name for arg-name
  • There are additional aliases for some args

MINIMAL EXAMPLES

# aiplatgo config for auto-naming conventions, gs-bucket-info, etc...
config:
    version: 1
    bucket: dev-ai-platform
    folder: gcs/folder/dev


# google ai platform flags with values (ie --module-name MODULE_NAME )
args:
  package: trainer
  module: trainer.task
  region: us-central1


# google ai platform flags without values (ie --stream-logs )
flags:
  - distributed
  - stream-logs


# user-args
user:
    user_arg_1: 1
    user_arg_2: 2


# local-user-args (only added with `aiplatgo local ...`)
local:
    local_user_arg_1: 1
    local_user_arg_2: 2

# job-user-args (only added with `aiplatgo <train|predict> ...`)
job:
    local_user_arg_1: 1
    local_user_arg_2: 2

The examples below assume the above config file is named config.yaml, and use the --echo/-e-flag which prints out the command without executing:

# local train
aiplatgo local train config --echo
# output
gcloud ai-platform local train --package-path trainer --module-name trainer.task --job-dir v1/output --distributed --  --user_arg_1 1 --user_arg_2 2
# platform training (recommended: see below to specify job-name in config file)
aiplatgo train --name my_job config --echo
# output
gcloud ai-platform jobs submit training my_job --package-path trainer --module-name trainer.task --region us-central1 --job-dir gs://dev-ai-platform/output --staging-bucket gs://dev-ai-platform --stream-logs --  --user_arg_1 1 --user_arg_2 2

With command-line args/kwargs:

  • key=value updates the args will update the args config, while (config|flags|user).key=value will update the config/flags/user config.
  • flag_name, ie word not followed by =value, becomes a cli flag (ex stream-logs becomes --stream-logs)
aiplatgo local train config config.version=1234 worker-count=4 --echo
# output
gcloud ai-platform local train --package-path trainer --module-name trainer.task --worker-count 2 --job-dir my_default_job_name/v1234/output --distributed --  --user_arg_1 1 --user_arg_2 2

Adding a default_job_name:

# filename: config.yaml`
config:
    name: my_default_job_name
    ...
aiplatgo train config --echo
# output
gcloud ai-platform jobs submit training my_default_job_name --package-path trainer --module-name trainer.task --region us-central1 --job-dir gs://dev-ai-platform/output --staging-bucket gs://dev-ai-platform --stream-logs --  --user_arg_1 1 --user_arg_2 2

Note: you can use . to skip a config file, only using command-line args/kwargs.

PARSER

gcloud ai-platform passes user args as a flat list of key-value pairs. For example:

['key1','value1','key2','value2',...,'key(2n)','value(2n+1)']

aiplatgo.parser.Parser combines these together into a key-value dictionary. Moreover, it allows you to provide a config file that groups args into method specific groups and add default values.

Here is an example config file and python script:

# args_config.yaml
defaults:
    # these defaults are passed for every method using the argument
    train_model: 'true'
    score_model: 'false' 
    something_else: 1234

training_objects:
    - loss_func
    - optimizer
    - nb_classes
    - acc_weights
    - metric
    # the : can be used to provide per-method default values
    - loss_weights: null
    - label_smoothing: null
    - from_logits: false

build_model:
    - nb_classes
    - from_logits
    - size
    - cropping
    - float_cropping
    - in_ch
def training_objects(loss_func, optimizer, nb_classes, acc_weights, metric, loss_weights, label_smoothing, from_logits):
    # get optimizer,loss,metrics
    pass

def build_model(
        optimizer,
        loss,
        metrics,
        nb_classes,
        from_logits,
        size,
        cropping,
        float_cropping,
        in_ch):
    # build model 
    pass

@click.pass_context
def run(ctx):
    parser=Parser('args_config.yaml',config_list=ctx.args)
    
    train_model=parser.get('train_model')
    score_model=parser.get('score_model')
    
    ...

    optimizer,loss,metrics=training_objects(**parser.get('training_objects'))

    model=build_model(
        optimizer,
        loss,
        metrics,
        **parser.args('build_model'))

CLI DOCS

aiplatgo local --help
Usage: aiplatgo local [OPTIONS] VERB [CONFIG]

  run local training or prediction:  
    * verb: "train" or "predict"  
    * config<optional>: yaml config. default or "." uses kwargs only
    * additional args become flags <ie> "arg_name" => "--arg_name"
    * additional kwargs become kw-flags <ie> "arg_name=123" => "--arg_name 123"

Options:
  --echo BOOLEAN  if true print command without executing
  --help          Show this message and exit.
 
aiplatgo train --help
Usage: aiplatgo train [OPTIONS] [CONFIG]

  submit training job to platform:  * config<optional>: yaml config. default
  or "." uses kwargs only  * additional args become flags <ie> "arg_name" =>
  "--arg_name"  * additional kwargs become kw-flags <ie> "arg_name=123" => "
  --arg_name 123"

Options:
  -n, --name TEXT  by default uses `config.name` for the job name. pass --name
                   to override

  -e, --echo       if true print command without executing
  -t, --timestamp  append timestamp (YYYYMMDD_HMS) to job name
  --help           Show this message and exit.
 
aiplatgo predict --help
Usage: aiplatgo predict [OPTIONS] [CONFIG]

  submit prediction job to platform:  * config<optional>: yaml config.
  default or "." uses kwargs only  * additional args become flags <ie>
  "arg_name" => "--arg_name"  * additional kwargs become kw-flags <ie>
  "arg_name=123" => "--arg_name 123"

Options:
  -n, --name TEXT  by default uses `config.name` for the job name. pass --name
                   to override

  -e, --echo       if true print command without executing
  -t, --timestamp  append timestamp (YYYYMMDD_HMS) to job name
  --help           Show this message and exit.