def run(self, experiment):
if isinstance(experiment, basestring):
experiment = self.db.get_experiment(experiment)
elif not isinstance(experiment, model.Experiment):
raise ValueError("Unknown type of experiment: " +
str(type(experiment)))
self.logger.info("Experiment key: " + experiment.key)
with model.get_db_provider(self.config) as db:
db.start_experiment(experiment)
""" Override env variables with those inside the queued message
"""
env = dict(os.environ)
if 'env' in self.config.keys():
for k, v in self.config['env'].iteritems():
if v is not None:
env[str(k)] = str(v)
fs_tracker.setup_experiment(env, experiment, clean=True)
log_path = fs_tracker.get_artifact_cache('output', experiment.key)
# log_path = os.path.join(model_dir, self.config['log']['name'])
self.logger.debug('Child process environment:')
self.logger.debug(str(env))
sched = BackgroundScheduler()
sched.start()
with open(log_path, 'w') as output_file:
p = subprocess.Popen(
["python", experiment.filename] + experiment.args,
stdout=output_file,
stderr=subprocess.STDOUT,
env=env,
cwd=experiment.artifacts['workspace']['local'])
# simple hack to show what's in the log file
ptail = subprocess.Popen(["tail", "-f", log_path])
sched.add_job(
lambda: db.checkpoint_experiment(experiment),
'interval',
minutes=self.config['saveWorkspaceFrequencyMinutes'])
def kill_if_stopped():
if db.get_experiment(experiment.key,
getinfo=False).status == 'stopped':
p.kill()
sched.add_job(kill_if_stopped, 'interval', seconds=10)
try:
p.wait()
finally:
ptail.kill()
db.finish_experiment(experiment)
sched.shutdown()
def add_experiment(args):
try:
config, python_pkg, e = args
e.pythonenv = add_packages(e.pythonenv, python_pkg)
with model.get_db_provider(config) as db:
db.add_experiment(e)
except BaseException:
traceback.print_exc()
raise
return e
def get_db():
global _db_provider
global _db_provider_timestamp
if not _db_provider or \
not _db_provider_timestamp or \
time.time() - _db_provider_timestamp > DB_PROVIDER_EXPIRATION:
_db_provider = model.get_db_provider(blocking_auth=False)
_db_provider_timestamp = time.time()
return _db_provider
def get_experiment_fitnesses(experiments, optimizer, config, logger):
db_provider = model.get_db_provider()
has_result = [False] * len(experiments)
fitnesses = [0.0] * len(experiments)
try:
term_criterion = config['optimizer']['termination_criterion']
except BaseException:
logger.warn("Cannot find termination criterion in config.yaml, looking"
"in optimizer source code instead")
term_criterion = optimizer.get_configs()['termination_criterion']
skip_gen_thres = term_criterion['skip_gen_thres']
skip_gen_timeout = term_criterion['skip_gen_timeout']
result_timestamp = time.time()
while sum(has_result) < len(experiments):
for i, experiment in enumerate(experiments):
if float(sum(has_result)) / len(experiments) > skip_gen_thres \
and time.time() - result_timestamp > skip_gen_timeout:
logger.warn(
"Skipping to next gen with %s of solutions evaled" %
(float(
sum(has_result)) /
len(experiments)))
has_result = [True] * len(experiments)
break
if has_result[i]:
continue
returned_experiment = db_provider.get_experiment(experiment.key,
getinfo=True)
# try:
# experiment_output = returned_experiment.info['logtail']
# except:
# logger.warn('Cannot access "logtail" ' +
# 'field in experiment.info')
output = db_provider._get_experiment_logtail(returned_experiment)
for line in output:
if line.startswith("Fitness") or line.startswith("fitness"):
try:
fitness = float(line.rstrip().split(':')[1])
assert fitness >= 0.0
except BaseException:
logger.warn('Error parsing or invalid fitness (%s)'
% line)
else:
fitnesses[i] = fitness
has_result[i] = True
result_timestamp = time.time()
break
time.sleep(config['sleep_time'])
return fitnesses
def _list(args, cli_args):
with model.get_db_provider(cli_args.config) as db:
if len(args) == 0:
experiments = db.get_user_experiments()
elif args[0] == 'project':
assert len(args) == 2
experiments = db.get_project_experiments(args[1])
elif args[0] == 'users':
assert len(args) == 1
users = db.get_users()
for u in users:
print(users[u].get('email'))
return
elif args[0] == 'user':
assert len(args) == 2
users = db.get_users()
user_ids = [u for u in users if users[u].get('email') == args[1]]
assert len(user_ids) == 1, \
'The user with email ' + args[1] + \
'not found!'
experiments = db.get_user_experiments(user_ids[0])
elif args[0] == 'all':
assert len(args) == 1
users = db.get_users()
experiments = []
for u in users:
experiments += db.get_user_experiments(u)
else:
get_logger().critical('Unknown command ' + args[0])
return
if cli_args.short:
for e in experiments:
print(e)
return
experiments = [db.get_experiment(e) for e in experiments]
experiments.sort(key=lambda e: -e.time_added)
table = [['Time added', 'Key', 'Project', 'Status']]
for e in experiments:
table.append([
time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(e.time_added)),
e.key,
e.project,
e.status])
print(AsciiTable(table).table)
def main(args=sys.argv[1:]):
parser = argparse.ArgumentParser(description='Studio WebUI server. \
Usage: studio \
<arguments>')
parser.add_argument('--config', help='configuration file', default=None)
# parser.add_argument('--guest',
# help='Guest mode (does not require db credentials)',
# action='store_true')
parser.add_argument('--port',
help='port to run Flask server on',
type=int,
default=5000)
parser.add_argument('--host', help='host name.', default='0.0.0.0')
parser.add_argument('--verbose',
'-v',
help='Verbosity level. Allowed vaules: ' +
'debug, info, warn, error, crit ' +
'or numerical value of logger levels.',
default=None)
args = parser.parse_args(args)
config = model.get_config()
if args.config:
with open(args.config) as f:
config = yaml.load(f)
if args.verbose:
config['verbose'] = args.verbose
# if args.guest:
# config['database']['guest'] = True
global _config
global _db_provider
_config = config
_db_provider = model.get_db_provider(_config, blocking_auth=False)
getlogger().setLevel(model.parse_verbosity(config.get('verbose')))
global _save_auth_cookie
_save_auth_cookie = True
print('Starting Studio UI on port {0}'.format(args.port))
app.run(host=args.host, port=args.port)
def __init__(self, args):
self.config = model.get_config()
if args.config:
if isinstance(args.config, basestring):
with open(args.config) as f:
self.config.update(yaml.load(f))
else:
self.config.update(args.config)
if args.guest:
self.config['database']['guest'] = True
self.db = model.get_db_provider(self.config)
self.logger = logging.getLogger('LocalExecutor')
self.logger.setLevel(model.parse_verbosity(self.config.get('verbose')))
self.logger.debug("Config: ")
self.logger.debug(self.config)
def submit_experiments(
experiments,
config,
runner_args,
logger,
resources_needed):
db = model.get_db_provider(config)
verbose = model.parse_verbosity(config['verbose'])
queue_name = 'local'
if 'queue' in config.keys():
queue_name = config['queue']
if runner_args.queue:
queue_name = runner_args.queue
for e in experiments:
e.pythonenv = add_packages(e.pythonenv, runner_args.python_pkg)
db.add_experiment(e)
logger.info("Added experiment " + e.key)
if runner_args.cloud is not None:
assert runner_args.cloud in ['gcloud', 'gcspot', 'ec2', 'ec2spot']
assert runner_args.queue is None, \
'--queue argument cannot be provided with --cloud argument'
auth_cookie = None if config['database'].get('guest') \
else os.path.join(
auth.token_dir,
config['database']['apiKey']
)
if runner_args.cloud in ['gcloud', 'gcspot']:
queue_name = 'pubsub_' + str(uuid.uuid4())
queue = PubsubQueue(queue_name, verbose=verbose)
worker_manager = GCloudWorkerManager(
auth_cookie=auth_cookie,
zone=config['cloud']['zone']
)
if runner_args.cloud in ['ec2', 'ec2spot']:
queue_name = 'sqs_' + str(uuid.uuid4())
queue = SQSQueue(queue_name, verbose=verbose)
worker_manager = EC2WorkerManager(
auth_cookie=auth_cookie
)
if runner_args.cloud == 'gcloud' or \
runner_args.cloud == 'ec2':
num_workers = int(
runner_args.num_workers) if runner_args.num_workers else 1
for i in range(num_workers):
worker_manager.start_worker(
queue_name, resources_needed,
ssh_keypair=runner_args.ssh_keypair,
timeout=runner_args.cloud_timeout)
else:
assert runner_args.bid is not None
if runner_args.num_workers:
start_workers = runner_args.num_workers
queue_upscaling = False
else:
start_workers = 1
queue_upscaling = True
worker_manager.start_spot_workers(
queue_name,
runner_args.bid,
resources_needed,
start_workers=start_workers,
queue_upscaling=queue_upscaling,
ssh_keypair=runner_args.ssh_keypair,
timeout=runner_args.cloud_timeout)
else:
if queue_name == 'local':
queue = LocalQueue()
queue.clean()
elif queue_name.startswith('sqs_'):
queue = SQSQueue(queue_name, verbose=verbose)
else:
queue = PubsubQueue(
queue_name,
config['database']['projectId'],
verbose=verbose)
for e in experiments:
queue.enqueue(json.dumps({
'experiment': e.__dict__,
'config': config}))
if queue_name == 'local':
worker_args = ['studio-local-worker']
if runner_args.config:
worker_args += ['--config=' + runner_args.config]
if runner_args.guest:
worker_args += ['--guest']
logger.info('worker args: {}'.format(worker_args))
if not runner_args.num_workers or int(runner_args.num_workers) == 1:
local_worker.main(worker_args)
else:
raise NotImplementedError("Multiple local workers are not " +
"implemented yet")
return
def main(args=sys.argv):
logger = logging.getLogger('studio-runner')
parser = argparse.ArgumentParser(
description='Studio runner. \
Usage: studio run <runner_arguments> \
script <script_arguments>')
parser.add_argument('--config', help='configuration file', default=None)
parser.add_argument('--project', help='name of the project', default=None)
parser.add_argument(
'--experiment', '-e',
help='Name of the experiment. If none provided, ' +
'random uuid will be generated',
default=None)
parser.add_argument(
'--guest',
help='Guest mode (does not require db credentials)',
action='store_true')
parser.add_argument(
'--force-git',
help='If run in a git directory, force running the experiment ' +
'even if changes are not commited',
action='store_true')
parser.add_argument(
'--gpus',
help='Number of gpus needed to run the experiment',
default=None)
parser.add_argument(
'--cpus',
help='Number of cpus needed to run the experiment' +
' (used to configure cloud instance)',
default=None)
parser.add_argument(
'--ram',
help='Amount of RAM needed to run the experiment' +
' (used to configure cloud instance)',
default=None)
parser.add_argument(
'--hdd',
help='Amount of hard drive space needed to run the experiment' +
' (used to configure cloud instance)',
default=None)
parser.add_argument(
'--queue', '-q',
help='Name of the remote execution queue',
default=None)
parser.add_argument(
'--cloud',
help='Cloud execution mode. Could be gcloud, ec2 or ec2spot',
default=None)
parser.add_argument(
'--bid',
help='Spot instance price bid, specified in USD or in percentage ' +
'of on-demand instance price. Default is %(default)s',
default='100%')
parser.add_argument(
'--capture-once', '-co',
help='Name of the immutable artifact to be captured. ' +
'It will be captured once before the experiment is run',
default=[], action='append')
parser.add_argument(
'--capture', '-c',
help='Name of the mutable artifact to be captured continuously',
default=[], action='append')
parser.add_argument(
'--reuse', '-r',
help='Name of the artifact from another experiment to use',
default=[], action='append')
parser.add_argument(
'--verbose', '-v',
help='Verbosity level. Allowed values: ' +
'debug, info, warn, error, crit ' +
'or numerical value of logger levels.',
default=None)
parser.add_argument(
'--metric', '-m',
help='Metric to show in the summary of the experiment, ' +
'and to base hyperparameter search on. ' +
'Refers a scalar value in tensorboard log ' +
'example: --metric=val_loss[:final | :min | :max] to report ' +
'validation loss in the end of the keras experiment ' +
'(or smallest or largest throughout the experiment for :min ' +
'and :max respectively)',
default=None)
parser.add_argument(
'--hyperparam', '-hp',
help='Try out multiple values of a certain parameter. ' +
'For example, --hyperparam=learning_rate:0.01:0.1:l10 ' +
'will instantiate 10 versions of the script, replace ' +
'learning_rate with a one of the 10 values for learning ' +
'rate that lies on a log grid from 0.01 to 0.1, create '
'experiments and place them in the queue.',
default=[], action='append')
parser.add_argument(
'--num-workers',
help='Number of local or cloud workers to spin up',
default=None)
parser.add_argument(
'--python-pkg',
help='Python package not present in the current environment ' +
'that is needed for experiment. Only compatible with ' +
'remote and cloud workers for now',
default=[], action='append')
parser.add_argument(
'--ssh-keypair',
help='Name of the SSH keypair used to access the EC2 ' +
'instances directly',
default=None)
parser.add_argument(
'--optimizer', '-opt',
help='Name of optimizer to use, by default is grid search. ' +
'The name of the optimizer must either be in ' +
'studio/optimizer_plugins ' +
'directory or the path to the optimizer source file ' +
'must be supplied. ',
default='grid')
parser.add_argument(
'--cloud-timeout',
help="Time (in seconds) that cloud workers wait for messages. " +
"If negative, " +
"wait for the first message in the queue indefinitely " +
"and shut down " +
"as soon as no new messages are available. " +
"If zero, don't wait at all." +
"Default value is %(default)",
type=int,
default=300)
# detect which argument is the script filename
# and attribute all arguments past that index as related to the script
py_suffix_args = [i for i, arg in enumerate(args) if arg.endswith('.py')]
if len(py_suffix_args) < 1:
print('At least one argument should be a python script ' +
'(end with *.py)')
parser.print_help()
exit()
script_index = py_suffix_args[0]
runner_args = parser.parse_args(args[1:script_index])
exec_filename, other_args = args[script_index], args[script_index + 1:]
# TODO: Queue the job based on arguments and only then execute.
config = model.get_config(runner_args.config)
if runner_args.verbose:
config['verbose'] = runner_args.verbose
verbose = model.parse_verbosity(config['verbose'])
logger.setLevel(verbose)
db = model.get_db_provider(config)
if git_util.is_git() and not git_util.is_clean():
logger.warn('Running from dirty git repo')
if not runner_args.force_git:
logger.error(
'Specify --force-git to run experiment from dirty git repo')
sys.exit(1)
resources_needed = parse_hardware(runner_args, config['cloud'])
logger.debug('resources requested: ')
logger.debug(str(resources_needed))
artifacts = {}
artifacts.update(parse_artifacts(runner_args.capture, mutable=True))
artifacts.update(parse_artifacts(runner_args.capture_once, mutable=False))
artifacts.update(parse_external_artifacts(runner_args.reuse, db))
if any(runner_args.hyperparam):
if runner_args.optimizer is "grid":
experiments = add_hyperparam_experiments(
exec_filename,
other_args,
runner_args,
artifacts,
resources_needed)
submit_experiments(
experiments,
config,
runner_args,
logger,
resources_needed)
else:
opt_modulepath = os.path.join(
os.path.dirname(os.path.abspath(__file__)),
"optimizer_plugins",
runner_args.optimizer + ".py")
# logger.info('optimizer path: %s' % opt_modulepath)
if not os.path.exists(opt_modulepath):
opt_modulepath = os.path.abspath(
os.path.expanduser(runner_args.optimizer))
logger.info('optimizer path: %s' % opt_modulepath)
assert os.path.exists(opt_modulepath)
sys.path.append(os.path.dirname(opt_modulepath))
opt_module = importlib.import_module(
os.path.basename(opt_modulepath.replace(".py", '')))
hyperparam_values, log_scale_dict = get_hyperparam_values(
runner_args)
optimizer = getattr(opt_module, "Optimizer")(hyperparam_values,
log_scale_dict)
while not optimizer.stop():
hyperparam_tuples = optimizer.ask()
experiments = add_hyperparam_experiments(
exec_filename,
other_args,
runner_args,
artifacts,
resources_needed,
optimizer=optimizer,
hyperparam_tuples=hyperparam_tuples)
submit_experiments(
experiments,
config,
runner_args,
logger,
resources_needed)
fitnesses = get_experiment_fitnesses(experiments,
optimizer, config, logger)
optimizer.tell(hyperparam_tuples, fitnesses)
# if config['verbose'] == "info" or config['verbose'] ==
# "debug":
try:
optimizer.disp()
except BaseException:
logger.warn('Optimizer has no disp() method')
else:
experiments = [model.create_experiment(
filename=exec_filename,
args=other_args,
experiment_name=runner_args.experiment,
project=runner_args.project,
artifacts=artifacts,
resources_needed=resources_needed,
metric=runner_args.metric)]
submit_experiments(
experiments,
config,
runner_args,
logger,
resources_needed)
db = None
return
def _kill(args, cli_args):
with model.get_db_provider(cli_args.config) as db:
for e in args:
get_logger().info('Deleting experiment ' + e)
db.delete_experiment(e)
def run(self, experiment):
if isinstance(experiment, six.string_types):
experiment = self.db.get_experiment(experiment)
elif not isinstance(experiment, Experiment):
raise ValueError("Unknown type of experiment: " +
str(type(experiment)))
self.logger.info("Experiment key: " + experiment.key)
with model.get_db_provider(self.config) as db:
db.start_experiment(experiment)
""" Override env variables with those inside the queued message
"""
env = dict(os.environ)
if 'env' in self.config.keys():
for k, v in six.iteritems(self.config['env']):
if v is not None:
env[str(k)] = str(v)
env['PYTHONUNBUFFERED'] = 'TRUE'
fs_tracker.setup_experiment(env, experiment, clean=False)
log_path = fs_tracker.get_artifact_cache('output', experiment.key)
# log_path = os.path.join(model_dir, self.config['log']['name'])
self.logger.debug('Child process environment:')
self.logger.debug(str(env))
sched = BackgroundScheduler()
sched.start()
with open(log_path, 'w') as output_file:
python = 'python'
if experiment.pythonver == 3:
python = 'python3'
cmd = [python, experiment.filename] + experiment.args
cwd = experiment.artifacts['workspace']['local']
container_artifact = experiment.artifacts.get('_singularity')
if container_artifact:
container = container_artifact.get('local')
if not container:
container = container_artifact.get('qualified')
cwd = fs_tracker.get_artifact_cache(
'workspace', experiment.key)
for tag, art in six.iteritems(experiment.artifacts):
local_path = art.get('local')
if not art['mutable'] and os.path.exists(local_path):
os.symlink(art['local'],
os.path.join(os.path.dirname(cwd), tag))
if experiment.filename is not None:
cmd = [
'singularity',
'exec',
container,
] + cmd
else:
cmd = ['singularity', 'run', container]
self.logger.info('Running cmd: \n {} '.format(cmd))
p = subprocess.Popen(cmd,
stdout=output_file,
stderr=subprocess.STDOUT,
env=env,
cwd=cwd)
# simple hack to show what's in the log file
# ptail = subprocess.Popen(["tail", "-f", log_path])
logtail = Pygtail(log_path)
def tail_func():
while logtail:
for line in logtail:
print(line)
time.sleep(0.1)
tail_thread = threading.Thread(target=tail_func)
tail_thread.start()
minutes = 0
if self.config.get('saveWorkspaceFrequency'):
minutes = int(
str2duration(self.config['saveWorkspaceFrequency']).
total_seconds() / 60)
def checkpoint():
try:
db.checkpoint_experiment(experiment)
except BaseException as e:
self.logger.info(e)
sched.add_job(checkpoint, 'interval', minutes=minutes)
metrics_path = fs_tracker.get_artifact_cache(
'_metrics', experiment.key)
minutes = 0
if self.config.get('saveMetricsFrequency'):
minutes = int(
str2duration(self.config['saveMetricsFrequency']).
total_seconds() / 60)
sched.add_job(lambda: save_metrics(metrics_path),
'interval',
minutes=minutes)
def kill_if_stopped():
if db.get_experiment(experiment.key,
getinfo=False).status == 'stopped':
p.kill()
if experiment.max_duration is not None and \
time.time() > experiment.time_started + \
int(str2duration(experiment.max_duration)
.total_seconds()):
p.kill()
sched.add_job(kill_if_stopped, 'interval', seconds=10)
try:
p.wait()
finally:
save_metrics(metrics_path)
sched.shutdown()
logtail = None
db.checkpoint_experiment(experiment)
db.finish_experiment(experiment)
return p.returncode
def worker_loop(queue,
parsed_args,
single_experiment=False,
timeout=0,
verbose=None):
fetch_artifacts = True
logger = logs.getLogger('worker_loop')
hold_period = 4
retval = 0
while True:
msg = queue.dequeue(acknowledge=False, timeout=timeout)
if not msg:
break
# first_exp, ack_key = queue.dequeue(acknowledge=False)
first_exp, ack_key = msg
data_dict = json.loads(sixdecode(first_exp))
experiment_key = data_dict['experiment']['key']
config = data_dict['config']
parsed_args.config = config
if verbose:
config['verbose'] = verbose
else:
verbose = model.parse_verbosity(config.get('verbose'))
logger.setLevel(verbose)
logger.debug('Received message: \n{}'.format(data_dict))
executor = LocalExecutor(parsed_args)
with model.get_db_provider(config) as db:
# experiment = experiment_from_dict(data_dict['experiment'])
def try_get_experiment():
experiment = db.get_experiment(experiment_key)
if experiment is None:
raise ValueError(
'experiment is not found - indicates storage failure')
return experiment
experiment = retry(try_get_experiment,
sleep_time=10,
logger=logger)
if config.get('experimentLifetime') and \
int(str2duration(config['experimentLifetime'])
.total_seconds()) + experiment.time_added < time.time():
logger.info(
'Experiment expired (max lifetime of {} was exceeded)'.
format(config.get('experimentLifetime')))
queue.acknowledge(ack_key)
continue
if allocate_resources(experiment, config, verbose=verbose):
def hold_job():
queue.hold(ack_key, hold_period)
hold_job()
sched = BackgroundScheduler()
sched.add_job(hold_job, 'interval', minutes=hold_period / 2)
sched.start()
try:
python = 'python'
if experiment.pythonver == 3:
python = 'python3'
if '_singularity' not in experiment.artifacts.keys():
pip_diff = pip_needed_packages(experiment.pythonenv,
python)
if any(pip_diff):
logger.info(
'Setting up python packages for experiment')
if pip_install_packages(pip_diff, python,
logger) != 0:
logger.info(
"Installation of all packages together " +
" failed, "
"trying one package at a time")
for pkg in pip_diff:
pip_install_packages([pkg], python, logger)
for tag, art in six.iteritems(experiment.artifacts):
if fetch_artifacts or 'local' not in art.keys():
logger.info('Fetching artifact ' + tag)
if tag == 'workspace':
art['local'] = retry(lambda: db.get_artifact(
art, only_newer=False),
sleep_time=10,
logger=logger)
else:
art['local'] = retry(
lambda: db.get_artifact(art),
sleep_time=10,
logger=logger)
returncode = executor.run(experiment)
if returncode != 0:
retval = returncode
finally:
sched.shutdown()
queue.acknowledge(ack_key)
if single_experiment:
logger.info('single_experiment is True, quitting')
return retval
else:
logger.info('Cannot run experiment ' + experiment.key +
' due lack of resources. Will retry')
time.sleep(config['sleep_time'])
# wait_for_messages(queue, timeout, logger)
# queue = glob.glob(fs_tracker.get_queue_directory() + "/*")
logger.info("Queue in {} is empty, quitting".format(
fs_tracker.get_queue_directory()))
return retval
def get_db():
global _db_provider
if not _db_provider:
_db_provider = model.get_db_provider(blocking_auth=False)
return _db_provider
def main(args=sys.argv[1:]):
logger = logs.getLogger('studio-runner')
parser = argparse.ArgumentParser(
description='Studio runner. \
Usage: studio run <runner_arguments> \
script <script_arguments>')
parser.add_argument('--config', help='configuration file', default=None)
parser.add_argument('--project', help='name of the project', default=None)
parser.add_argument(
'--experiment', '-e',
help='Name of the experiment. If none provided, ' +
'random uuid will be generated',
default=None)
parser.add_argument(
'--guest',
help='Guest mode (does not require db credentials)',
action='store_true')
parser.add_argument(
'--force-git',
help='If run in a git directory, force running the experiment ' +
'even if changes are not commited',
action='store_true')
parser.add_argument(
'--gpus',
help='Number of gpus needed to run the experiment',
type=int,
default=None)
parser.add_argument(
'--cpus',
help='Number of cpus needed to run the experiment' +
' (used to configure cloud instance)',
type=int,
default=None)
parser.add_argument(
'--ram',
help='Amount of RAM needed to run the experiment' +
' (used to configure cloud instance), ex: 10G, 10GB',
default=None)
parser.add_argument(
'--gpuMem',
help='Amount of GPU RAM needed to run the experiment',
default=None)
parser.add_argument(
'--hdd',
help='Amount of hard drive space needed to run the experiment' +
' (used to configure cloud instance), ex: 10G, 10GB',
default=None)
parser.add_argument(
'--queue', '-q',
help='Name of the remote execution queue',
default=None)
parser.add_argument(
'--cloud',
help='Cloud execution mode. Could be gcloud, gcspot, ec2 or ec2spot',
default=None)
parser.add_argument(
'--bid',
help='Spot instance price bid, specified in USD or in percentage ' +
'of on-demand instance price. Default is %(default)s',
default='100%')
parser.add_argument(
'--capture-once', '-co',
help='Name of the immutable artifact to be captured. ' +
'It will be captured once before the experiment is run',
default=[], action='append')
parser.add_argument(
'--capture', '-c',
help='Name of the mutable artifact to be captured continuously',
default=[], action='append')
parser.add_argument(
'--reuse', '-r',
help='Name of the artifact from another experiment to use',
default=[], action='append')
parser.add_argument(
'--verbose', '-v',
help='Verbosity level. Allowed values: ' +
'debug, info, warn, error, crit ' +
'or numerical value of logger levels.',
default=None)
parser.add_argument(
'--metric',
help='Metric to show in the summary of the experiment, ' +
'and to base hyperparameter search on. ' +
'Refers a scalar value in tensorboard log ' +
'example: --metric=val_loss[:final | :min | :max] to report ' +
'validation loss in the end of the keras experiment ' +
'(or smallest or largest throughout the experiment for :min ' +
'and :max respectively)',
default=None)
parser.add_argument(
'--hyperparam', '-hp',
help='Try out multiple values of a certain parameter. ' +
'For example, --hyperparam=learning_rate:0.01:0.1:l10 ' +
'will instantiate 10 versions of the script, replace ' +
'learning_rate with a one of the 10 values for learning ' +
'rate that lies on a log grid from 0.01 to 0.1, create '
'experiments and place them in the queue.',
default=[], action='append')
parser.add_argument(
'--num-workers',
help='Number of local or cloud workers to spin up',
type=int,
default=None)
parser.add_argument(
'--python-pkg',
help='Python package not present in the current environment ' +
'that is needed for experiment. Only compatible with ' +
'remote and cloud workers for now',
default=[], action='append')
parser.add_argument(
'--ssh-keypair',
help='Name of the SSH keypair used to access the EC2 ' +
'instances directly',
default=None)
parser.add_argument(
'--optimizer', '-opt',
help='Name of optimizer to use, by default is grid search. ' +
'The name of the optimizer must either be in ' +
'studio/optimizer_plugins ' +
'directory or the path to the optimizer source file ' +
'must be supplied. ',
default='grid')
parser.add_argument(
'--cloud-timeout',
help="Time (in seconds) that cloud workers wait for messages. " +
"If negative, " +
"wait for the first message in the queue indefinitely " +
"and shut down " +
"as soon as no new messages are available. " +
"If zero, don't wait at all." +
"Default value is %(default)d",
type=int,
default=300)
parser.add_argument(
'--user-startup-script',
help='Path of script to run immediately ' +
'before running the remote worker',
default=None)
parser.add_argument(
'--branch',
help='Branch of studioml to use when running remote worker, useful ' +
'for debugging pull requests. Default is current',
default=None)
parser.add_argument(
'--max-duration',
help='Max experiment runtime (i.e. time after which experiment ' +
'should be killed no matter what.). Examples of values ' +
'might include 5h, 48h2m10s',
default=None)
parser.add_argument(
'--lifetime',
help='Max experiment lifetime (i.e. wait time after which ' +
'experiment loses relevance and should not be started)' +
' Examples include 240h30m10s',
default=None)
parser.add_argument(
'--container',
help='Singularity container in which experiment should be run. ' +
'Assumes that container has all dependencies installed',
default=None
)
parser.add_argument(
'--port',
help='Ports to open on a cloud instance',
default=[], action='append'
)
# detect which argument is the script filename
# and attribute all arguments past that index as related to the script
(runner_args, other_args) = parser.parse_known_args(args)
py_suffix_args = [i for i, arg in enumerate(args) if arg.endswith('.py')
or '::' in arg]
rerun = False
if len(py_suffix_args) < 1:
print('None of the arugments end with .py')
if len(other_args) == 0:
print("Trying to run a container job")
assert runner_args.container is not None
exec_filename = None
elif len(other_args) == 1:
print("Treating last argument as experiment key to rerun")
rerun = True
experiment_key = args[-1]
else:
print("Too many extra arguments - should be either none " +
"for container job or one for experiment re-run")
sys.exit(1)
else:
script_index = py_suffix_args[0]
exec_filename, other_args = args[script_index], args[script_index + 1:]
runner_args = parser.parse_args(args[:script_index])
# TODO: Queue the job based on arguments and only then execute.
config = model.get_config(runner_args.config)
if runner_args.verbose:
config['verbose'] = runner_args.verbose
if runner_args.guest:
config['database']['guest'] = True
if runner_args.container:
runner_args.capture_once.append(
runner_args.container + ':_singularity')
verbose = model.parse_verbosity(config['verbose'])
logger.setLevel(verbose)
if git_util.is_git() and not git_util.is_clean() and not rerun:
logger.warn('Running from dirty git repo')
if not runner_args.force_git:
logger.error(
'Specify --force-git to run experiment from dirty git repo')
sys.exit(1)
resources_needed = parse_hardware(runner_args, config['resources_needed'])
logger.debug('resources requested: ')
logger.debug(str(resources_needed))
artifacts = {}
artifacts.update(parse_artifacts(runner_args.capture, mutable=True))
artifacts.update(parse_artifacts(runner_args.capture_once, mutable=False))
with model.get_db_provider(config) as db:
artifacts.update(parse_external_artifacts(runner_args.reuse, db))
if runner_args.branch:
config['cloud']['branch'] = runner_args.branch
if runner_args.user_startup_script:
config['cloud']['user_startup_script'] = \
runner_args.user_startup_script
if runner_args.lifetime:
config['experimentLifetime'] = runner_args.lifetime
if any(runner_args.hyperparam):
if runner_args.optimizer is "grid":
experiments = add_hyperparam_experiments(
exec_filename,
other_args,
runner_args,
artifacts,
resources_needed,
logger)
queue_name = submit_experiments(
experiments,
config=config,
logger=logger,
queue_name=runner_args.queue,
cloud=runner_args.cloud)
spin_up_workers(
runner_args,
config,
resources_needed,
queue_name=queue_name,
verbose=verbose)
else:
opt_modulepath = os.path.join(
os.path.dirname(os.path.abspath(__file__)),
"optimizer_plugins",
runner_args.optimizer + ".py")
if not os.path.exists(opt_modulepath):
opt_modulepath = os.path.abspath(
os.path.expanduser(runner_args.optimizer))
logger.info('optimizer path: %s' % opt_modulepath)
assert os.path.exists(opt_modulepath)
sys.path.append(os.path.dirname(opt_modulepath))
opt_module = importlib.import_module(
os.path.basename(opt_modulepath.replace(".py", '')))
h = HyperparameterParser(runner_args, logger)
hyperparams = h.parse()
optimizer = getattr(
opt_module,
"Optimizer")(
hyperparams,
config['optimizer'],
logger)
workers_started = False
queue_name = runner_args.queue
while not optimizer.stop():
hyperparam_pop = optimizer.ask()
hyperparam_tuples = h.convert_to_tuples(hyperparam_pop)
experiments = add_hyperparam_experiments(
exec_filename,
other_args,
runner_args,
artifacts,
resources_needed,
logger,
optimizer=optimizer,
hyperparam_tuples=hyperparam_tuples)
queue_name = submit_experiments(
experiments,
config=config,
logger=logger,
cloud=runner_args.cloud,
queue_name=queue_name)
if not workers_started:
spin_up_workers(
runner_args,
config,
resources_needed,
queue_name=queue_name,
verbose=verbose)
workers_started = True
fitnesses, behaviors = get_experiment_fitnesses(
experiments, optimizer, config, logger)
# for i, hh in enumerate(hyperparam_pop):
# print fitnesses[i]
# for hhh in hh:
# print hhh
try:
optimizer.tell(hyperparam_pop, fitnesses, behaviors)
except BaseException:
optimizer.tell(hyperparam_pop, fitnesses)
try:
optimizer.disp()
except BaseException:
logger.warn('Optimizer has no disp() method')
else:
if rerun:
with model.get_db_provider(config) as db:
experiment = db.get_experiment(experiment_key)
new_key = runner_args.experiment if runner_args.experiment \
else experiment_key + '_rerun' + str(uuid.uuid4())
experiment.key = new_key
for _, art in six.iteritems(experiment.artifacts):
art['mutable'] = False
experiments = [experiment]
else:
experiments = [create_experiment(
filename=exec_filename,
args=other_args,
experiment_name=runner_args.experiment,
project=runner_args.project,
artifacts=artifacts,
resources_needed=resources_needed,
metric=runner_args.metric,
max_duration=runner_args.max_duration,
)]
queue_name = submit_experiments(
experiments,
config=config,
logger=logger,
cloud=runner_args.cloud,
queue_name=runner_args.queue)
spin_up_workers(
runner_args,
config,
resources_needed,
queue_name=queue_name,
verbose=verbose)
return
def _stop(args, cli_args):
with model.get_db_provider(cli_args.config) as db:
for e in args:
get_logger().info('Stopping experiment ' + e)
db.stop_experiment(e)
def get_experiment_fitnesses(experiments, optimizer, config, logger):
with model.get_db_provider() as db:
progbar = Progbar(len(experiments), interval=0.0)
logger.info("Waiting for fitnesses from %s experiments" %
len(experiments))
bad_line_dicts = [dict() for x in range(len(experiments))]
has_result = [False for i in range(len(experiments))]
fitnesses = [0.0 for i in range(len(experiments))]
behaviors = [None for i in range(len(experiments))]
term_criterion = config['optimizer']['termination_criterion']
skip_gen_thres = term_criterion['skip_gen_thres']
skip_gen_timeout = term_criterion['skip_gen_timeout']
result_timestamp = time.time()
while sum(has_result) < len(experiments):
for i, experiment in enumerate(experiments):
if float(sum(has_result)) / len(experiments) >= skip_gen_thres\
and time.time() - result_timestamp > skip_gen_timeout:
logger.warn(
"Skipping to next gen with %s of solutions evaled" %
(float(
sum(has_result)) /
len(experiments)))
has_result = [True] * len(experiments)
break
if has_result[i]:
continue
returned_experiment = db.get_experiment(experiment.key,
getinfo=True)
# try:
# experiment_output = returned_experiment.info['logtail']
# except BaseException:
# logger.warn('Cannot access "logtail" in experiment.info')
output = db._get_experiment_logtail(
returned_experiment)
if output is None:
continue
for j, line in enumerate(output):
if line.startswith(
"Traceback (most recent call last):") and \
j not in bad_line_dicts[i]:
logger.warn("Experiment %s: error"
" discovered in output" %
returned_experiment.key)
logger.warn("".join(output[j:]))
bad_line_dicts[i][j] = True
if line.startswith("Behavior") or \
line.startswith("behavior"):
try:
behavior = eval(line.rstrip().split(':')[1])
if isinstance(behavior, np.ndarray):
pass
elif isinstance(behavior, list):
behavior = np.array(behavior)
else:
raise
except BaseException:
if j not in bad_line_dicts[i]:
logger.warn(
'Experiment %s: error parsing or invalid'
' behavior' %
returned_experiment.key)
logger.warn(line)
bad_line_dicts[i][j] = True
else:
behaviors[i] = behavior
if line.startswith("Fitness") or \
line.startswith("fitness"):
try:
fitness = float(line.rstrip().split(':')[1])
# assert fitness >= 0.0
except BaseException:
if j not in bad_line_dicts[i]:
logger.warn(
'Experiment %s: error parsing or invalid'
' fitness' %
returned_experiment.key)
logger.warn(line)
bad_line_dicts[i][j] = True
else:
if fitness < 0.0:
logger.warn('Experiment %s: returned'
' fitness is less than zero,'
' setting it to zero' %
returned_experiment.key)
fitness = 0.0
fitnesses[i] = fitness
has_result[i] = True
progbar.add(1)
result_timestamp = time.time()
break
time.sleep(config['sleep_time'])
print
return fitnesses, behaviors
def worker_loop(queue,
parsed_args,
setup_pyenv=False,
single_experiment=False,
fetch_artifacts=False,
timeout=0):
logger = logging.getLogger('worker_loop')
hold_period = 4
while queue.has_next():
first_exp, ack_key = queue.dequeue(acknowledge=False)
experiment_key = json.loads(first_exp)['experiment']['key']
config = json.loads(first_exp)['config']
parsed_args.config = config
verbose = model.parse_verbosity(config.get('verbose'))
logger.setLevel(verbose)
logger.debug(
'Received experiment {} with config {} from the queue'.format(
experiment_key, config))
executor = LocalExecutor(parsed_args)
with model.get_db_provider(config) as db:
experiment = db.get_experiment(experiment_key)
if allocate_resources(experiment, config, verbose=verbose):
def hold_job():
queue.hold(ack_key, hold_period)
hold_job()
sched = BackgroundScheduler()
sched.add_job(hold_job, 'interval', minutes=hold_period / 2)
sched.start()
try:
if setup_pyenv:
logger.info(
'Setting up python packages for experiment')
for pkg in experiment.pythonenv:
pipp = subprocess.Popen(['pip', 'install', pkg],
stdout=subprocess.PIPE,
stderr=subprocess.STDOUT)
pipout, _ = pipp.communicate()
logger.info("pip output: \n" + pipout)
for tag, art in experiment.artifacts.iteritems():
if fetch_artifacts or 'local' not in art.keys():
logger.info('Fetching artifact ' + tag)
if tag == 'workspace':
art['local'] = db.get_artifact(
art, only_newer=False)
else:
art['local'] = db.get_artifact(art)
executor.run(experiment)
finally:
sched.shutdown()
queue.acknowledge(ack_key)
if single_experiment:
logger.info('single_experiment is True, quitting')
return
else:
logger.info('Cannot run experiment ' + experiment.key +
' due lack of resources. Will retry')
time.sleep(config['sleep_time'])
wait_for_messages(queue, timeout, logger)
# queue = glob.glob(fs_tracker.get_queue_directory() + "/*")
logger.info("Queue in {} is empty, quitting".format(
fs_tracker.get_queue_directory()))