def run(args):
"""main pbt entry point, receives *args* with sys.argv format"""
if len(args) == 1:
print(format_usage(args[0]))
else:
_, command, *command_args = args
pbt.run(command, command_args)
def run(models, config_path=None, pbt_path=None, basepath=None):
if config_path is None:
parser = argparse.ArgumentParser("Train PyTorch model")
parser.add_argument(
'--pbt', default='', help="Path to pbt config file")
parser.add_argument(
'--basepath', default='', help="Path to project root directory")
parser.add_argument('config', help="Path to model config file")
args = parser.parse_args()
pbt_path = args.pbt
basepath = args.basepath or basepath
config_path = args.config
model_args = config.load(config_path)
model_name = model_args['model']
del model_args['model']
if "basepath" not in model_args and basepath:
model_args["basepath"] = basepath
Model = models[model_name]
if pbt_path:
import pbt
print('Running HPO on model with parameters {}'.format(model_args))
model = PbtModel(Model, model_args)
pbt.run(model, pbt_path)
else:
print('Running model with parameters {}'.format(model_args))
if "distributed" in model_args:
distributed.train(Model, model_args)
else:
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = ','.join(
map(str, model_args['gpus']))
model = Model(**model_args)
model.train()
return model
def run(args):
"""main pbt entry point, receives *args* with sys.argv format"""
if len(args) == 1:
print(format_usage(args[0]))
else:
_, command, *command_args = args
try:
pbt.run(command, command_args)
except pbt.ProjectNotFoundError:
print("Running a project command outside a pbt project")
