def _validate_v2(config, path):
base_path = Path(path).parent
try:
conf = ExperimentConfig(_base_path=base_path, **config)
return conf.json()
except Exception as e:
print_error(f'Config V2 validation failed: {repr(e)}')
def create_experiment(args):
'''start a new experiment'''
experiment_id = ''.join(random.sample(string.ascii_letters + string.digits, 8))
config_path = os.path.abspath(args.config)
if not os.path.exists(config_path):
print_error('Please set correct config path!')
exit(1)
config_yml = get_yml_content(config_path)
try:
config = ExperimentConfig(_base_path=Path(config_path).parent, **config_yml)
config_v2 = config.json()
except Exception as error_v2:
print_warning('Validation with V2 schema failed. Trying to convert from V1 format...')
try:
validate_all_content(config_yml, config_path)
except Exception as error_v1:
print_error(f'Convert from v1 format failed: {repr(error_v1)}')
print_error(f'Config in v2 format validation failed: {repr(error_v2)}')
exit(1)
from nni.experiment.config import convert
config_v2 = convert.to_v2(config_yml).json()
try:
if getattr(config_v2['trainingService'], 'platform', None) in k8s_training_services:
launch_experiment(args, config_yml, 'new', experiment_id, 1)
else:
launch_experiment(args, config_v2, 'new', experiment_id, 2)
except Exception as exception:
restServerPid = Experiments().get_all_experiments().get(experiment_id, {}).get('pid')
if restServerPid:
kill_command(restServerPid)
print_error(exception)
exit(1)
def test_all():
minimal = ExperimentConfig(**minimal_json)
assert minimal.json() == minimal_canon
assert minimal_class.json() == minimal_canon_2
detailed = ExperimentConfig.load(expand_path('assets/config.yaml'))
assert detailed.json() == detailed_canon
def create_experiment(args):
'''start a new experiment'''
experiment_id = ''.join(random.sample(string.ascii_letters + string.digits, 8))
config_path = os.path.abspath(args.config)
if not os.path.exists(config_path):
print_error('Please set correct config path!')
exit(1)
experiment_config = get_yml_content(config_path)
try:
validate_all_content(experiment_config, config_path)
except Exception:
print_warning('Validation with V1 schema failed. Trying to convert from V2 format...')
try:
config = ExperimentConfig(**experiment_config)
experiment_config = convert.to_v1_yaml(config)
except Exception as e:
print_error(f'Conversion from v2 format failed: {repr(e)}')
try:
validate_all_content(experiment_config, config_path)
except Exception as e:
print_error(f'Config in v1 format validation failed. {repr(e)}')
exit(1)
try:
launch_experiment(args, experiment_config, 'new', experiment_id)
except Exception as exception:
restServerPid = Experiments().get_all_experiments().get(experiment_id, {}).get('pid')
if restServerPid:
kill_command(restServerPid)
print_error(exception)
exit(1)
def create_experiment(args):
'''start a new experiment'''
experiment_id = ''.join(
random.sample(string.ascii_letters + string.digits, 8))
nni_config = Config(experiment_id)
nni_config.set_config('experimentId', experiment_id)
config_path = os.path.abspath(args.config)
if not os.path.exists(config_path):
print_error('Please set correct config path!')
exit(1)
experiment_config = get_yml_content(config_path)
try:
config = ExperimentConfig(**experiment_config)
experiment_config = convert.to_v1_yaml(config)
except Exception:
pass
try:
validate_all_content(experiment_config, config_path)
except Exception as e:
print_error(e)
exit(1)
nni_config.set_config('experimentConfig', experiment_config)
nni_config.set_config('restServerPort', args.port)
try:
launch_experiment(args, experiment_config, 'new', experiment_id)
except Exception as exception:
nni_config = Config(experiment_id)
restServerPid = nni_config.get_config('restServerPid')
if restServerPid:
kill_command(restServerPid)
print_error(exception)
exit(1)
def test_remote():
config = ExperimentConfig(**minimal_json)
assert config.json() == minimal_canon
assert minimal_class.json() == minimal_canon
config = ExperimentConfig(**detailed_json)
assert config.json() == detailed_canon
def create_experiment(args):
# to make it clear what are inside args
config_file = Path(args.config)
port = args.port
debug = args.debug
url_prefix = args.url_prefix
foreground = args.foreground
# it should finally be done in nnictl main function
# but for now don't break routines without logging support
init_logger_for_command_line()
logging.getLogger('nni').setLevel(logging.INFO)
if not config_file.is_file():
_logger.error(f'"{config_file}" is not a valid file.')
exit(1)
with config_file.open() as config:
config_content = yaml.safe_load(config)
v1_platform = config_content.get('trainingServicePlatform')
if v1_platform:
can_convert = True
if v1_platform == 'adl':
can_convert = False
if v1_platform in ['kubeflow', 'frameworkcontroller']:
reuse = config_content.get(v1_platform + 'Config', {}).get('reuse')
can_convert = (
reuse != False
) # if user does not explicitly specify it, convert to reuse mode
if not can_convert:
legacy_launcher.create_experiment(args)
exit()
try:
v2_config = convert.to_v2(config_content)
except Exception:
_logger.error(
'You are using legacy config format with incorrect fields or values, '
'to get more accurate error message please update it to the new format.'
)
_logger.error(
'Reference: https://nni.readthedocs.io/en/stable/reference/experiment_config.html'
)
exit(1)
_logger.warning(
f'You are using legacy config file, please update it to latest format:'
)
# use `print` here because logging will add timestamp and make it hard to copy paste
print(Fore.YELLOW + '=' * 80 + Fore.RESET)
print(yaml.dump(v2_config, sort_keys=False).strip())
print(Fore.YELLOW + '=' * 80 + Fore.RESET)
print(
Fore.YELLOW +
'Reference: https://nni.readthedocs.io/en/stable/reference/experiment_config.html'
+ Fore.RESET)
utils.set_base_path(config_file.parent)
config = ExperimentConfig(**v2_config)
utils.unset_base_path()
else:
config = ExperimentConfig.load(config_file)
if config.use_annotation:
path = Path(tempfile.gettempdir(), getuser(), 'nni', 'annotation')
path.mkdir(parents=True, exist_ok=True)
path = tempfile.mkdtemp(dir=path)
code_dir = expand_annotations(config.trial_code_directory, path)
config.trial_code_directory = code_dir
config.search_space = generate_search_space(code_dir)
assert config.search_space, 'ERROR: Generated search space is empty'
config.use_annotation = False
exp = Experiment(config)
exp.url_prefix = url_prefix
run_mode = RunMode.Foreground if foreground else RunMode.Detach
exp.start(port, debug, run_mode)
_logger.info(
f'To stop experiment run "nnictl stop {exp.id}" or "nnictl stop --all"'
)
_logger.info(
'Reference: https://nni.readthedocs.io/en/stable/Tutorial/Nnictl.html')
import json
from pathlib import Path
import yaml
from nni.experiment.config import ExperimentConfig, AlgorithmConfig, LocalConfig
## template ##
config = ExperimentConfig(search_space_file='',
trial_command='echo hello',
trial_concurrency=1,
tuner=AlgorithmConfig(name='randomm'),
training_service=LocalConfig())
space_correct = {
'pool_type': {
'_type': 'choice',
'_value': ['max', 'min', 'avg']
},
'学习率': {
'_type': 'loguniform',
'_value': [1e-7, 0.1]
}
}
# FIXME
# PyYAML 6.0 (YAML 1.1) does not support tab and scientific notation
# JSON does not support comment and extra comma
# So some combinations will fail to load
formats = [
## minimal config ##
minimal_json = {
'searchSpace': {'a': 1},
'trialCommand': 'python main.py',
'trialConcurrency': 2,
'tuner': {
'name': 'random',
},
'trainingService': {
'platform': 'local',
},
}
minimal_class = ExperimentConfig('local')
minimal_class.search_space = {'a': 1}
minimal_class.trial_command = 'python main.py'
minimal_class.trial_concurrency = 2
minimal_class.tuner.name = 'random'
minimal_canon = {
'searchSpace': {'a': 1},
'trialCommand': 'python main.py',
'trialCodeDirectory': os.path.realpath('.'),
'trialConcurrency': 2,
'useAnnotation': False,
'debug': False,
'logLevel': 'info',
'experimentWorkingDirectory': str(Path.home() / 'nni-experiments'),
'tuner': {'name': 'random'},
'user': '******',
'password': '******',
},
],
},
}
minimal_class = ExperimentConfig(
search_space = {'a': 1},
trial_command = 'python main.py',
trial_concurrency = 2,
tuner = AlgorithmConfig(
name = 'random',
),
training_service = RemoteConfig(
machine_list = [
RemoteMachineConfig(
host = '1.2.3.4',
user = '******',
password = '******',
),
],
),
)
minimal_canon = {
'searchSpace': {'a': 1},
'trialCommand': 'python main.py',
'trialCodeDirectory': os.path.realpath('.'),
'trialConcurrency': 2,
'useAnnotation': False,