def run(self, args=None):
"""Main entrypoint."""
parser = argparse.ArgumentParser()
parser.add_argument(
'-c',
'--config',
default=None,
help=('Configuration as a file or a JSON string. '
'Setting "-" will read from the standard input.'))
parser.add_argument(
'-s',
'--storage_config',
default=None,
help=
('Configuration of available storages as a file or a JSON string. '
'Setting "-" will read from the standard input.'))
parser.add_argument(
'-ms',
'--model_storage',
default=None,
help='Model storage in the form <storage_id>:[<path>].')
parser.add_argument(
'-msr',
'--model_storage_read',
default=None,
help=(
'Model storage to read from, in the form <storage_id>:[<path>] '
'(defaults to model_storage).'))
parser.add_argument(
'-msw',
'--model_storage_write',
default=None,
help=(
'Model storage to write to, in the form <storage_id>:[<path>] '
'(defaults to model_storage).'))
parser.add_argument('-m',
'--model',
default=None,
help='Model to load.')
parser.add_argument(
'-g',
'--gpuid',
default="0",
help=
"Comma-separated list of 1-indexed GPU identifiers (0 for CPU).")
parser.add_argument('-t',
'--task_id',
default=None,
help="Identifier of this run.")
parser.add_argument(
'-i',
'--image',
default="?",
help="Full URL (registry/image:tag) of the image used for this run."
)
parser.add_argument('-b',
'--beat_url',
default=None,
help=("Endpoint that listens to beat requests "
"(push notifications of activity)."))
parser.add_argument('-bi',
'--beat_interval',
default=30,
type=int,
help="Interval of beat requests in seconds.")
parser.add_argument('--no_push',
default=False,
action='store_true',
help='Do not push model.')
subparsers = parser.add_subparsers(help='Run type', dest='cmd')
parser_train = subparsers.add_parser('train', help='Run a training.')
parser_trans = subparsers.add_parser('trans',
help='Run a translation.')
parser_trans.add_argument('-i',
'--input',
required=True,
nargs='+',
help='Input file.')
parser_trans.add_argument('-o',
'--output',
required=True,
nargs='+',
help='Output file.')
parser_trans.add_argument('--as_release',
default=False,
action='store_true',
help='Translate from a released model.')
parser_release = subparsers.add_parser(
'release', help='Release a model for serving.')
parser_release.add_argument(
'-d',
'--destination',
default=None,
help='Released model storage (defaults to the model storage).')
parser_serve = subparsers.add_parser('serve', help='Serve a model.')
parser_serve.add_argument('-hs',
'--host',
default="0.0.0.0",
help='Serving hostname.')
parser_serve.add_argument('-p',
'--port',
type=int,
default=4000,
help='Serving port.')
parser_preprocess = subparsers.add_parser(
'preprocess', help='Sample and preprocess corpus.')
parser_preprocess.add_argument('--build_model',
default=False,
action='store_true',
help='Preprocess data into a model.')
parser.build_vocab = subparsers.add_parser('buildvocab',
help='Build vocabularies.')
args = parser.parse_args(args=args)
if args.config is None and args.model is None:
parser.error(
'at least one of --config or --model options must be set')
if args.model_storage_read is None:
args.model_storage_read = args.model_storage
if args.model_storage_write is None:
args.model_storage_write = args.model_storage
if (not self._stateless
and (args.cmd != 'preprocess' or args.build_model)
and (args.model_storage_write is None
or args.model_storage_write is None)):
parser.error('Missing model storage argument')
if args.task_id is None:
args.task_id = str(uuid.uuid4())
# for backward compatibility - convert singleton in int
args.gpuid = args.gpuid.split(',')
args.gpuid = [int(g) for g in args.gpuid]
if len(args.gpuid) == 1:
args.gpuid = args.gpuid[0]
start_beat_service(os.uname()[1],
args.beat_url,
args.task_id,
interval=args.beat_interval)
config = load_config(args.config) if args.config is not None else {}
parent_model = args.model or config.get('model')
storage = StorageClient(tmp_dir=self._tmp_dir,
config=load_config(args.storage_config)
if args.storage_config else None)
if parent_model is not None and not self._stateless:
# Download model locally and merge the configuration.
remote_model_path = storage.join(args.model_storage_read,
parent_model)
model_path = os.path.join(self._models_dir, parent_model)
fetch_model(storage, remote_model_path, model_path)
with open(os.path.join(model_path, 'config.json'),
'r') as config_file:
model_config = json.load(config_file)
if 'modelType' not in model_config:
if parent_model.endswith('_release'):
model_config['modelType'] = 'release'
else:
model_config['modelType'] = 'checkpoint'
config = merge_config(copy.deepcopy(model_config), config)
else:
model_path = None
model_config = None
if args.cmd == 'train':
if (parent_model is not None and config['modelType']
not in ('checkpoint', 'base', 'preprocess')):
raise ValueError(
'cannot train from a model that is not a training checkpoint, '
'a base model, or a preprocess model')
self.train_wrapper(args.task_id,
config,
storage,
args.model_storage_write,
args.image,
parent_model=parent_model,
model_path=model_path,
model_config=model_config,
gpuid=args.gpuid,
push_model=not args.no_push)
elif args.cmd == 'buildvocab':
self.build_vocab(args.task_id,
config,
storage,
args.model_storage_write,
args.image,
push_model=not args.no_push)
elif args.cmd == 'trans':
if (not self._stateless and
(parent_model is None or config['modelType'] != 'checkpoint')):
raise ValueError('translation requires a training checkpoint')
self.trans_wrapper(config,
model_path,
storage,
args.input,
args.output,
as_release=args.as_release,
gpuid=args.gpuid)
elif args.cmd == 'release':
if (not self._stateless and
(parent_model is None or config['modelType'] != 'checkpoint')):
raise ValueError('releasing requires a training checkpoint')
if args.destination is None:
args.destination = args.model_storage_write
self.release_wrapper(config,
model_path,
storage,
args.image,
args.destination,
gpuid=args.gpuid,
push_model=not args.no_push)
elif args.cmd == 'serve':
if (not self._stateless and
(parent_model is None or config['modelType'] != 'release')):
raise ValueError('serving requires a released model')
self.serve_wrapper(config,
model_path,
args.host,
args.port,
gpuid=args.gpuid)
elif args.cmd == 'preprocess':
if not args.build_model:
self.preprocess(config, storage)
else:
if (parent_model is not None
and config['modelType'] not in ('checkpoint', 'base')):
raise ValueError(
'cannot preprocess from a model that is not a training '
'checkpoint or a base model')
self.preprocess_into_model(args.task_id,
config,
storage,
args.model_storage_write,
args.image,
parent_model=parent_model,
model_path=model_path,
push_model=not args.no_push)
def run(self):
"""Main entrypoint."""
parser = argparse.ArgumentParser()
parser.add_argument(
'-c',
'--config',
default=None,
help=('Configuration as a file or a JSON string. '
'Setting "-" will read from the standard input.'))
parser.add_argument(
'-s',
'--storage_config',
default=None,
help=
('Configuration of available storages as a file or a JSON string. '
'Setting "-" will read from the standard input.'))
parser.add_argument(
'-ms',
'--model_storage',
help='Model storage in the form <storage_id>:[<path>].')
parser.add_argument('-m',
'--model',
default=None,
help='Model to load.')
parser.add_argument(
'-g',
'--gpuid',
default="0",
help=
"Comma-separated list of 1-indexed GPU identifiers (0 for CPU).")
parser.add_argument('-t',
'--task_id',
default=None,
help="Identifier of this run.")
parser.add_argument(
'-i',
'--image',
default="?",
help="Full URL (registry/image:tag) of the image used for this run."
)
parser.add_argument('-b',
'--beat_url',
default=None,
help=("Endpoint that listens to beat requests "
"(push notifications of activity)."))
parser.add_argument('-bi',
'--beat_interval',
default=30,
type=int,
help="Interval of beat requests in seconds.")
subparsers = parser.add_subparsers(help='Run type', dest='cmd')
parser_train = subparsers.add_parser('train', help='Run a training.')
parser_trans = subparsers.add_parser('trans',
help='Run a translation.')
parser_trans.add_argument('-i',
'--input',
required=True,
help='Input file.')
parser_trans.add_argument('-o',
'--output',
required=True,
help='Output file.')
parser_serve = subparsers.add_parser('serve', help='Serve a model.')
parser_serve.add_argument('-hs',
'--host',
default="0.0.0.0",
help='Serving hostname.')
parser_serve.add_argument('-p',
'--port',
type=int,
default=4000,
help='Serving port.')
parser.build_vocab = subparsers.add_parser(
'preprocess', help='Sample and preprocess corpus.')
args = parser.parse_args()
if args.config is None and args.model is None:
parser.error(
'at least one of --config or --model options must be set')
if not self._stateless and args.cmd != 'preprocess' and not args.model_storage:
parser.error('argument -ms/--model_storage is required')
if args.task_id is None:
args.task_id = str(uuid.uuid4())
# for backward compatibility - convert singleton in int
args.gpuid = args.gpuid.split(',')
args.gpuid = [int(g) for g in args.gpuid]
if len(args.gpuid) == 1:
args.gpuid = args.gpuid[0]
start_beat_service(os.uname()[1],
args.beat_url,
args.task_id,
interval=args.beat_interval)
config = load_config(args.config) if args.config is not None else {}
parent_model = args.model or config.get('model')
storage = StorageClient(tmp_dir=self._tmp_dir,
config=load_config(args.storage_config)
if args.storage_config else None)
if parent_model is not None and not self._stateless:
# Download model locally and merge the configuration.
remote_model_path = storage.join(args.model_storage, parent_model)
model_path = os.path.join(self._models_dir, parent_model)
fetch_model(storage, remote_model_path, model_path)
with open(os.path.join(model_path, 'config.json'),
'r') as config_file:
model_config = json.load(config_file)
config = merge_config(model_config, config)
else:
model_path = None
if args.cmd == 'train':
self.train_wrapper(args.task_id,
config,
storage,
args.model_storage,
args.image,
parent_model=parent_model,
model_path=model_path,
gpuid=args.gpuid)
elif args.cmd == 'trans':
if parent_model is None:
raise ValueError('translation requires a model')
self.trans_wrapper(config,
model_path,
storage,
args.input,
args.output,
gpuid=args.gpuid)
elif args.cmd == 'serve':
if parent_model is None:
raise ValueError('serving requires a model')
self.serve_wrapper(config,
model_path,
args.host,
args.port,
gpuid=args.gpuid)
elif args.cmd == 'preprocess':
self.preprocess(config, storage)
def run(self):
"""Main entrypoint."""
parser = argparse.ArgumentParser()
parser.add_argument(
'-c',
'--config',
default=None,
help=('Configuration as a file or a JSON string. '
'Setting "-" will read from the standard input.'))
parser.add_argument(
'-s',
'--storage_config',
default=None,
help=
('Configuration of available storages as a file or a JSON string. '
'Setting "-" will read from the standard input.'))
parser.add_argument(
'-ms',
'--model_storage',
required=not self._stateless,
help='Model storage in the form <storage_id>:[<path>].')
parser.add_argument('-m',
'--model',
default=None,
help='Model to load.')
parser.add_argument('-g',
'--gpuid',
default=0,
type=int,
help="1-indexed GPU identifier (0 for CPU).")
parser.add_argument('-t',
'--task_id',
default=None,
help="Identifier of this run.")
parser.add_argument(
'-i',
'--image',
default="?",
help="Full URL (registry/image:tag) of the image used for this run."
)
parser.add_argument('-b',
'--beat_url',
default=None,
help=("Endpoint that listens to beat requests "
"(push notifications of activity)."))
parser.add_argument('-bi',
'--beat_interval',
default=30,
type=int,
help="Interval of beat requests in seconds.")
subparsers = parser.add_subparsers(help='Run type', dest='cmd')
parser_train = subparsers.add_parser('train', help='Run a training.')
parser_trans = subparsers.add_parser('trans',
help='Run a translation.')
parser_trans.add_argument('-i',
'--input',
required=True,
help='Input file.')
parser_trans.add_argument('-o',
'--output',
required=True,
help='Output file.')
args = parser.parse_args()
if args.config is None and args.model is None:
parser.error(
'at least one of --config or --model options must be set')
if args.task_id is None:
args.task_id = str(uuid.uuid4())
start_beat_service(os.uname()[1],
args.beat_url,
args.task_id,
interval=args.beat_interval)
config = load_config(args.config) if args.config is not None else {}
parent_model = args.model or config.get('model')
storage = StorageClient(config=load_config(args.storage_config)
if args.storage_config else None)
if parent_model is not None and not self._stateless:
# Download model locally and merge the configuration.
remote_model_path = storage.join(args.model_storage, parent_model)
model_path = os.path.join(self._models_dir, parent_model)
fetch_model(storage, remote_model_path, model_path)
with open(os.path.join(model_path, 'config.json'),
'r') as config_file:
model_config = json.load(config_file)
config = merge_config(model_config, config)
else:
model_path = None
if args.cmd == 'train':
self.train_wrapper(args.task_id,
config,
storage,
args.model_storage,
args.image,
model_path=model_path,
gpuid=args.gpuid)
elif parent_model is None:
raise ValueError('translation requires a model')
elif args.cmd == 'trans':
self.trans_wrapper(config,
model_path,
storage,
args.input,
args.output,
gpuid=args.gpuid)