def _test_serving(self, data_in, expected_data_out, wrapper=None):
self.port = randint(5000, 9000)
server_experimentid = 'test_serving_' + str(uuid.uuid4())
with get_local_queue_lock():
args = [
'studio', 'run', '--force-git', '--verbose=debug',
'--experiment=' + server_experimentid,
'--config=' + self.get_config_path(), 'studio::serve_main',
'--port=' + str(self.port), '--host=localhost'
]
if wrapper:
args.append('--wrapper=' + wrapper)
subprocess.Popen(args, cwd=os.path.dirname(__file__))
time.sleep(60)
try:
retval = requests.post(url='http://localhost:' + str(self.port),
json=data_in)
data_out = retval.json()
assert data_out == expected_data_out
finally:
with model.get_db_provider(model.get_config(
self.get_config_path())) as db:
db.stop_experiment(server_experimentid)
time.sleep(20)
db.delete_experiment(server_experimentid)
def test_experiment_lifetime(self):
my_path = os.path.dirname(os.path.realpath(__file__))
logger = logs.getLogger('test_experiment_lifetime')
logger.setLevel(10)
config_name = os.path.join(my_path, 'test_config_http_client.yaml')
key = 'test_experiment_lifetime' + str(uuid.uuid4())
with model.get_db_provider(model.get_config(config_name)) as db:
try:
db.delete_experiment(key)
except Exception:
pass
p = subprocess.Popen(['studio', 'run',
'--config=' + config_name,
'--experiment=' + key,
'--force-git',
'--verbose=debug',
'--lifetime=-10m',
'stop_experiment.py'],
stdout=subprocess.PIPE,
stderr=subprocess.STDOUT,
cwd=my_path)
pout, _ = p.communicate()
if pout:
logger.debug("studio run output: \n" + pout.decode())
db.delete_experiment(key)
def getResultsWithTimeout(self, timeout=0):
total_sleep_time = 0
sleep_time = 1
while True:
with model.get_db_provider(self.config) as db:
if self.resumable:
experiment_keys = db.get_project_experiments(
self.project_name).keys()
else:
experiment_keys = self.submitted
for key in experiment_keys:
e = db.get_experiment(key)
if e is not None and e.status == 'finished':
self.logger.debug(
'Experiment {} finished, getting results' .format(
e.key))
with open(db.get_artifact(e.artifacts['retval']),
'rb') as f:
data = pickle.load(f)
if not self.resumable:
self.submitted.remove(e.key)
else:
db.delete_experiment(e.key)
return (e.key, data)
if timeout == 0 or \
(timeout > 0 and total_sleep_time > timeout):
return None
time.sleep(sleep_time)
total_sleep_time += sleep_time
def get_store(self, config_name='test_config.yaml'):
config_file = os.path.join(os.path.dirname(os.path.realpath(__file__)),
config_name)
with open(config_file) as f:
config = yaml.load(f, Loader=yaml.SafeLoader)
return model.get_db_provider(config).store
def __enter__(self):
with model.get_db_provider(self.config):
pass
if self.wm:
self.logger.debug('Spinning up cloud workers')
if self.use_spot:
self.wm.start_spot_workers(
self.queue_name,
self.bid,
self.resources_needed,
start_workers=self.num_workers,
queue_upscaling=self.queue_upscaling,
ssh_keypair=self.ssh_keypair,
timeout=self.cloud_timeout)
else:
for i in range(self.num_workers):
self.wm.start_worker(
self.queue_name,
self.resources_needed,
ssh_keypair=self.ssh_keypair,
timeout=self.cloud_timeout)
self.p = None
else:
self.logger.debug('Starting local worker')
self.p = subprocess.Popen([
'studio-local-worker',
'--verbose=%s' % self.config['verbose'],
'--timeout=' + str(self.cloud_timeout)],
close_fds=True)
return self
def get_firebase_provider(self, config_name='test_config.yaml'):
config_file = os.path.join(os.path.dirname(os.path.realpath(__file__)),
config_name)
with open(config_file) as f:
config = yaml.load(f)
return model.get_db_provider(config)
def _create_artifacts(
self,
client_code_file,
args_file,
workspace_new,
files):
artifacts = {
'retval': {
'mutable': True,
'unpack': True
},
'clientscript': {
'mutable': False,
'local': client_code_file,
'unpack': True
},
'args': {
'mutable': False,
'local': args_file,
'unpack': True
},
'workspace': {
'mutable': False,
'local': workspace_new,
'unpack': True
}
}
for tag, name in six.iteritems(files):
artifacts[tag] = {}
url_schema = re.compile('^https{0,1}://')
s3_schema = re.compile('^s3://')
gcs_schema = re.compile('^gs://')
studio_schema = re.compile(
'studio://(?P<experiment>.+)/(?P<artifact>.+)')
if url_schema.match(name):
artifacts[tag]['url'] = name
artifacts[tag]['unpack'] = False
elif s3_schema.match(name) or gcs_schema.match(name):
artifacts[tag]['qualified'] = name
artifacts[tag]['unpack'] = False
elif studio_schema.match(name):
ext_experiment_key = studio_schema.match(
name).group('experiment')
ext_tag = studio_schema.match(name).group('artifact')
with model.get_db_provider(self.config) as db:
ext_experiment = db.get_experiment(ext_experiment_key)
artifacts[tag]['key'] = \
ext_experiment.artifacts[ext_tag]['key']
artifacts[tag]['unpack'] = True
else:
artifacts[tag]['local'] = os.path.abspath(
os.path.expanduser(name))
artifacts[tag]['unpack'] = True
artifacts[tag]['mutable'] = False
return artifacts
def get_provider(self, config_name=None):
config_name = config_name if config_name else \
self.get_default_config_name()
config_file = os.path.join(os.path.dirname(os.path.realpath(__file__)),
config_name)
with open(config_file) as f:
config = yaml.load(f)
return model.get_db_provider(config)
def getResultsWithTimeout(self, timeout=0):
total_sleep_time = 0
sleep_time = self.sleep_time
assert self.resumable is False
while True:
with model.get_db_provider(self.config) as db:
for key, submitted_time in six.iteritems(self.submitted):
try:
e = db.get_experiment(key)
if e is not None:
retval_path = db.get_artifact(
e.artifacts['retval'])
if os.path.exists(retval_path) and \
os.path.getmtime(retval_path) > submitted_time:
with open(retval_path, 'rb') as f:
data = pickle.load(f)
del self.submitted[e.key]
return (e.key, data)
except BaseException as e:
self.logger.debug(
"Getting result failed due to exception:")
self.logger.debug(e)
'''
if e is not None and e.status == 'finished':
self.logger.debug(
'Experiment {} finished, getting results' .format(
e.key))
with open(db.get_artifact(e.artifacts['retval']),
'rb') as f:
data = pickle.load(f)
if not self.resumable:
self.submitted.remove(e.key)
else:
db.delete_experiment(e.key)
return (e.key, data)
'''
if timeout == 0 or \
(timeout > 0 and total_sleep_time > timeout):
return None
if self.p is not None:
assert self.p.poll() is None, \
"Executor process died, no point in waiting for results"
time.sleep(sleep_time)
total_sleep_time += sleep_time
def get_db():
global _config
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(_config, blocking_auth=False)
_db_provider_timestamp = time.time()
return _db_provider
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='localhost')
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, Loader=yaml.FullLoader)
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)
getlogger().setLevel(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 _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.keys():
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 test_stop_experiment(self):
my_path = os.path.dirname(os.path.realpath(__file__))
logger = logs.getLogger('test_stop_experiment')
logger.setLevel(10)
config_name = os.path.join(my_path, 'test_config_http_client.yaml')
key = 'test_stop_experiment' + str(uuid.uuid4())
with model.get_db_provider(model.get_config(config_name)) as db:
try:
db.delete_experiment(key)
except Exception:
pass
p = subprocess.Popen(['studio', 'run',
'--config=' + config_name,
'--experiment=' + key,
'--force-git',
'--verbose=debug',
'stop_experiment.py'],
stdout=subprocess.PIPE,
stderr=subprocess.STDOUT,
cwd=my_path)
# wait till experiment spins up
experiment = None
while experiment is None or experiment.status == 'waiting':
time.sleep(1)
try:
experiment = db.get_experiment(key)
except BaseException:
pass
logger.info('Stopping experiment')
db.stop_experiment(key)
pout, _ = p.communicate()
if pout:
logger.debug("studio run output: \n" + pout.decode())
db.delete_experiment(key)
def __enter__(self):
with model.get_db_provider(self.config):
pass
self.p = None
if self.wm:
self.logger.debug('Spinning up cloud workers')
if self.use_spot:
self.wm.start_spot_workers(
self.queue_name,
self.bid,
self.resources_needed,
start_workers=self.num_workers,
queue_upscaling=self.queue_upscaling,
ssh_keypair=self.ssh_keypair,
timeout=self.cloud_timeout)
else:
for i in range(self.num_workers):
self.wm.start_worker(
self.queue_name,
self.resources_needed,
ssh_keypair=self.ssh_keypair,
timeout=self.cloud_timeout)
elif self.queue_name is None or self.queue_name == 'local':
self.logger.debug('Starting local worker')
self.p = subprocess.Popen([
'studio-local-worker',
'--verbose=%s' % self.config['verbose'],
'--timeout=' + str(self.cloud_timeout)],
close_fds=True)
# yet another case is when queue name is specified, but
# cloud is not - that means running on a separately
# managed server that listens to the queue
#
# The contract is queue_name that starts with sqs or ec2
# is an SQS queue, otherwise, it is a PubSub queue
return self
def stubtest_worker(testclass,
experiment_name,
runner_args,
config_name,
test_script,
expected_output,
script_args=[],
queue=LocalQueue(),
wait_for_experiment=True,
delete_when_done=True,
test_output=True):
my_path = os.path.dirname(os.path.realpath(__file__))
config_name = os.path.join(my_path, config_name)
logger = logging.getLogger('stubtest_worker')
logger.setLevel(10)
queue.clean()
with model.get_db_provider(model.get_config(config_name)) as db:
try:
db.delete_experiment(experiment_name)
except Exception:
pass
p = subprocess.Popen(['studio', 'run'] + runner_args + [
'--config=' + config_name, '--verbose=debug', '--force-git',
'--experiment=' + experiment_name, test_script
] + script_args,
stdout=subprocess.PIPE,
stderr=subprocess.STDOUT,
close_fds=True,
cwd=my_path)
pout, _ = p.communicate()
if pout:
logger.debug("studio run output: \n" + pout)
db = model.get_db_provider(model.get_config(config_name))
experiments = [
e for e in db.get_user_experiments() if e.startswith(experiment_name)
]
assert len(experiments) == 1
experiment_name = experiments[0]
try:
# test saved arguments
keybase = "/experiments/" + experiment_name
saved_args = db._get(keybase + '/args')
if saved_args is not None:
testclass.assertTrue(len(saved_args) == len(script_args))
for i in range(len(saved_args)):
testclass.assertTrue(saved_args[i] == script_args[i])
testclass.assertTrue(db._get(keybase + '/filename') == test_script)
else:
testclass.assertTrue(script_args is None or len(script_args) == 0)
experiment = db.get_experiment(experiment_name)
if wait_for_experiment:
while not experiment.status == 'finished':
time.sleep(1)
experiment = db.get_experiment(experiment_name)
if test_output:
with open(db.store.get_artifact(experiment.artifacts['output']),
'r') as f:
data = f.read()
split_data = data.strip().split('\n')
testclass.assertEquals(split_data[-1], expected_output)
check_workspace(testclass, db, experiment_name)
if delete_when_done:
db.delete_experiment(experiment_name)
return db
except Exception as e:
print("Exception {} raised during test".format(e))
print("worker output: \n {}".format(pout))
print("Exception trace:")
print(traceback.format_exc())
raise e
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 get_db_provider(self):
config = model.get_config('test_config_http_client.yaml')
config['database']['serverUrl'] = 'http://localhost:' + str(self.port)
return model.get_db_provider(config)
def _get_provider():
config = _get_config()
return model.get_db_provider(config)
def main(args=sys.argv[1:]):
logger = logs.get_logger('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))
# Set up default artifacts:
# note that their "local" paths will be updated
# on Experiment creation,
# but they must have "local" field defined
# to have storage credentials set up properly.
artifacts = {
'workspace': {
'mutable': False,
'local': os.getcwd(),
'unpack': True
},
'modeldir': {
'mutable': True,
'local': '',
'unpack': True
},
'retval': {
'mutable': True,
'local': '',
'unpack': True
},
'output': {
'mutable': True,
'local': '',
'unpack': True
},
'tb': {
'mutable': True,
'local': '',
'unpack': True
},
'_metrics': {
'mutable': True,
'local': '',
'unpack': True
},
'_metadata': {
'mutable': True,
'local': '',
'unpack': True
}
}
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))
logger.debug("Task artifacts: %s", repr(artifacts))
storage_creds = config.get('storage', {}).get(KEY_CREDENTIALS, None)
_setup_artifacts_creds(artifacts, storage_creds)
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
queueLifetime = None
if any(runner_args.hyperparam):
if runner_args.optimizer == "grid":
experiments = _add_hyperparam_experiments(exec_filename,
other_args, runner_args,
artifacts,
resources_needed, logger)
queue = model.get_queue(queue_name=runner_args.queue,
cloud=runner_args.cloud,
config=config,
close_after=queueLifetime,
logger=logger,
verbose=verbose)
queue_name = submit_experiments(experiments,
config=config,
logger=logger,
queue=queue)
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 = model.get_queue(queue_name=queue_name,
cloud=runner_args.cloud,
config=config,
close_after=queueLifetime,
logger=logger,
verbose=verbose)
queue_name = submit_experiments(experiments,
config=config,
logger=logger,
queue=queue)
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)
try:
optimizer.tell(hyperparam_pop, fitnesses, behaviors)
except BaseException:
util.check_for_kb_interrupt()
optimizer.tell(hyperparam_pop, fitnesses)
try:
optimizer.disp()
except BaseException:
util.check_for_kb_interrupt()
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,
dependency_policy=StudioDependencyPolicy())
]
queue = model.get_queue(queue_name=runner_args.queue,
cloud=runner_args.cloud,
config=config,
close_after=queueLifetime,
logger=logger,
verbose=verbose)
queue_name = submit_experiments(experiments,
config=config,
logger=logger,
queue=queue)
spin_up_workers(runner_args,
config,
resources_needed,
queue_name=queue_name,
verbose=verbose)
return
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)
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:
util.check_for_kb_interrupt()
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:
util.check_for_kb_interrupt()
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'])
return fitnesses, behaviors
def worker_loop(queue,
parsed_args,
single_experiment=False,
timeout=0,
verbose=None):
fetch_artifacts = True
logger = logs.get_logger('worker_loop')
hold_period = 4
retval = 0
while True:
msg = queue.dequeue(acknowledge=False, timeout=timeout)
if not msg:
break
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', None))
logger.setLevel(verbose)
logger.debug('Received message: \n{}'.format(data_dict))
executor = LocalExecutor(queue, 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', None) and \
int(str2duration(config['experimentLifetime'])
.total_seconds()) + experiment.time_added < time.time():
logger.info(
'Experiment expired (max lifetime of {0} was exceeded)'.
format(config.get('experimentLifetime', None)))
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[0] == '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, item in experiment.artifacts.items():
art: Artifact = item
if fetch_artifacts or art.local_path is None:
get_only_newer: bool = True
if tag == 'workspace':
get_only_newer = False
if not art.is_mutable:
logger.info('Fetching artifact ' + tag)
art.local_path = retry(lambda: db.get_artifact(
art, only_newer=get_only_newer),
sleep_time=10,
logger=logger)
else:
logger.info('Skipping mutable artifact ' + tag)
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')
# Debounce failed requests we cannot service yet
time.sleep(config.get('sleep_time', 5))
logger.info("Queue in {0} is empty, quitting".format(
fs_tracker.get_queue_directory()))
return retval
def stubtest_worker(
testclass,
experiment_name,
runner_args,
config_name,
test_script,
expected_output,
script_args=[],
queue=LocalQueue(),
wait_for_experiment=True,
delete_when_done=True,
test_output=True,
test_workspace=True):
my_path = os.path.dirname(os.path.realpath(__file__))
config_name = os.path.join(my_path, config_name)
logger = logs.getLogger('stubtest_worker')
logger.setLevel(10)
queue.clean()
with model.get_db_provider(model.get_config(config_name)) as db:
try:
db.delete_experiment(experiment_name)
except Exception:
pass
os.environ['PYTHONUNBUFFERED'] = 'True'
p = subprocess.Popen(['studio', 'run'] + runner_args +
['--config=' + config_name,
'--verbose=debug',
'--force-git',
'--experiment=' + experiment_name,
test_script] + script_args,
stdout=subprocess.PIPE,
stderr=subprocess.STDOUT,
close_fds=True,
cwd=my_path)
pout, _ = p.communicate()
if pout:
logger.debug("studio run output: \n" + sixdecode(pout))
splitpout = sixdecode(pout).split('\n')
experiments = [line.split(' ')[-1] for line in splitpout
if line.startswith('studio run: submitted experiment')]
logger.debug("added experiments: {}".format(experiments))
db = model.get_db_provider(model.get_config(config_name))
experiment_name = experiments[0]
try:
experiment = db.get_experiment(experiment_name)
if wait_for_experiment:
while not experiment or not experiment.status == 'finished':
experiment = db.get_experiment(experiment_name)
if test_output:
with open(db.get_artifact(experiment.artifacts['output']),
'r') as f:
data = f.read()
split_data = data.strip().split('\n')
print(data)
testclass.assertEquals(split_data[-1], expected_output)
if test_workspace:
check_workspace(testclass, db, experiment_name)
if delete_when_done:
retry(lambda: db.delete_experiment(experiment_name), sleep_time=10)
return db
except Exception as e:
print("Exception {} raised during test".format(e))
print("worker output: \n {}".format(pout))
print("Exception trace:")
print(traceback.format_exc())
raise 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)
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[0] == '3':
python = 'python3'
python = which(python)
cmd = [python, experiment.filename] + experiment.args
cwd = experiment.artifacts['workspace'].local_path
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: {0} in {1}'.format(cmd, cwd))
p = subprocess.Popen(cmd,
stdout=subprocess.PIPE,
stderr=subprocess.STDOUT,
env=env,
cwd=cwd,
text=True)
def kill_subprocess():
p.kill()
def get_duration(tag: str):
value = self.config.get(tag, '0m')
return int(str2duration(value).total_seconds() / 60)
def checkpoint():
try:
db.checkpoint_experiment(experiment)
except BaseException as e:
self.logger.info(e)
check_for_kb_interrupt()
minutes = get_duration('saveWorkspaceFrequency')
sched.add_job(checkpoint, 'interval', minutes=minutes)
metrics_path = fs_tracker.get_artifact_cache(
'_metrics', experiment.key)
minutes = get_duration('saveMetricsFrequency')
sched.add_job(lambda: save_metrics(metrics_path),
'interval',
minutes=minutes)
def kill_if_stopped():
try:
db_expr = db.get_experiment(experiment.key,
getinfo=False)
except:
check_for_kb_interrupt()
db_expr = None
# Transient issues with getting experiment data might
# result in a None value being returned, as result
# leave the experiment running because we wont be able to
# do anything else even if this experiment is stopped
# in any event if the experiment runs too long then it
# will exceed its allocated time and stop
if db_expr is not None:
if db_expr.status == 'stopped':
kill_subprocess()
return
if experiment.max_duration is not None and \
time.time() > experiment.time_started + \
int(str2duration(experiment.max_duration)
.total_seconds()):
kill_subprocess()
return
# If our tasks queue is signalled inactive
# during work process execution, that means we need to drop
# current execution and exit
if not self.task_queue.is_active():
kill_subprocess()
sched.add_job(kill_if_stopped, 'interval', seconds=10)
while True:
output = p.stdout.readline()
if output == '' and p.poll() is not None:
break
if output:
line_out = output.strip()
print(line_out)
output_file.write(line_out)
try:
p.wait()
finally:
save_metrics(metrics_path)
sched.shutdown()
db.checkpoint_experiment(experiment)
db.finish_experiment(experiment)
return p.returncode
def get_db_provider(self):
config = model.get_config(self.client_config_file)
config['database']['serverUrl'] = 'http://localhost:' + str(self.port)
return model.get_db_provider(config)
def get_db_provider(self, config_name):
config_file = os.path.join(
os.path.dirname(
os.path.realpath(__file__)),
config_name)
return model.get_db_provider(model.get_config(config_file))
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 studio_run(self, line, cell=None):
script_text = []
pickleable_ns = {}
for varname, var in six.iteritems(self.shell.user_ns):
if not varname.startswith('__'):
if isinstance(var, ModuleType) and \
var.__name__ != 'studio.magics':
script_text.append(
'import {} as {}'.format(var.__name__, varname)
)
else:
try:
pickle.dumps(var)
pickleable_ns[varname] = var
except BaseException:
pass
script_text.append(cell)
script_text = '\n'.join(script_text)
stub_path = os.path.join(
os.path.dirname(os.path.realpath(__file__)),
'run_magic.py.stub')
with open(stub_path) as f:
script_stub = f.read()
script = script_stub.format(script=script_text)
experiment_key = str(int(time.time())) + \
"_jupyter_" + str(uuid.uuid4())
print('Running studio with experiment key ' + experiment_key)
config = model.get_config()
if config['database']['type'] == 'http':
print("Experiment progress can be viewed/shared at:")
print("{}/experiment/{}".format(
config['database']['serverUrl'],
experiment_key))
workspace_new = fs_tracker.get_artifact_cache(
'workspace', experiment_key)
rsync_cp('.', workspace_new)
with open(os.path.join(workspace_new, '_script.py'), 'w') as f:
f.write(script)
ns_path = fs_tracker.get_artifact_cache('_ns', experiment_key)
with gzip.open(ns_path, 'wb') as f:
f.write(pickle.dumps(pickleable_ns))
if any(line):
runner_args = line.strip().split(' ')
else:
runner_args = []
runner_args.append('--capture={}:_ns'.format(ns_path))
runner_args.append('--capture-once=.:workspace')
runner_args.append('--force-git')
runner_args.append('--experiment=' + experiment_key)
notebook_cwd = os.getcwd()
os.chdir(workspace_new)
print(runner_args + ['_script.py'])
runner_main(runner_args + ['_script.py'])
os.chdir(notebook_cwd)
with model.get_db_provider() as db:
while True:
experiment = db.get_experiment(experiment_key)
if experiment and experiment.status == 'finished':
break
time.sleep(10)
new_ns_path = db.get_artifact(experiment.artifacts['_ns'])
with open(new_ns_path) as f:
new_ns = pickle.loads(f.read())
self.shell.user_ns.update(new_ns)