def main(args):
# Get experiment data
db = schedy.SchedyDB()
experiment = db.get_experiment(args.experiment)
# Create the directory that will contain the models
try:
os.makedirs(args.models_dir)
except OSError as e:
if e.errno != errno.EEXIST:
raise
# Load the training/test data
(x_train, y_train), (x_test, y_test) = mnist.load_data()
x_train = x_train.reshape(x_train.shape[0], IMG_SIZE, IMG_SIZE, 1).astype(
K.floatx()) / 255.
x_test = x_test.reshape(x_test.shape[0], IMG_SIZE, IMG_SIZE, 1).astype(
K.floatx()) / 255.
y_train = keras.utils.to_categorical(y_train, NUM_CLASSES)
y_test = keras.utils.to_categorical(y_test, NUM_CLASSES)
# Process one job
with experiment.next_job() as job:
# Create and train the model for one epoch
model = make_model(job)
model.fit(x_train, y_train, epochs=5, batch_size=256)
loss, accuracy = model.evaluate(x_test, y_test, batch_size=256)
# Update the results
job.results.setdefault('loss', []).append(loss)
job.results.setdefault('accuracy', []).append(accuracy)
job.results['min_loss'] = min(job.results['loss'])
job.results['max_accuracy'] = max(job.results['accuracy'])
# Do not forget to save the weights, as the next job that will need to
# load these
weights_path = os.path.join(args.models_dir, job.job_id + '.h5')
job.results['weights_path'] = weights_path
model.save_weights(weights_path)
def cmd_add(args):
db = schedy.SchedyDB(config_path=args.config)
if args.scheduler == 'manual':
exp = schedy.ManualSearch(args.experiment, status=args.status)
elif args.scheduler == 'random':
if len(args.hyperparameters) % 3 != 0:
args.parser.error('Invalid hyperparameters (not a list of name/distribution/params).')
hyperparameters = {}
for i in range(0, len(args.hyperparameters), 3):
name = args.hyperparameters[i]
dist_name = args.hyperparameters[i + 1]
params_txt = args.hyperparameters[i + 2]
if name in hyperparameters:
args.parser.error('Duplicate hyperparameter: {}.'.format(name))
try:
dist_type = schedy.random._DISTRIBUTION_TYPES[dist_name]
except KeyError:
args.parser.error('Invalid distribution: {}.'.format(dist_name))
try:
params = json.loads(params_txt)
hyperparameters[name] = dist_type._from_args(params)
except (TypeError, ValueError, KeyError) as e:
args.parser.error('Invalid distribution parameters for {} ({!r}).'.format(name, e))
exp = schedy.RandomSearch(args.experiment, status=args.status, distributions=hyperparameters)
db.add_experiment(exp)
def cmd_show(args):
db = schedy.SchedyDB(config_path=args.config)
exp = db.get_experiment(args.experiment)
if args.job is None:
print_exp(exp)
else:
job = exp.get_job(args.job)
print_job(job)
def cmd_run(args):
db = schedy.SchedyDB(config_path=args.config)
exp = db.get_experiment(args.experiment)
while True:
try:
with exp.next_job() as job:
cmd_args = format_cmd_args(args.cmd, job)
print('Calling {}'.format(cmd_args))
output_block = False
output_content = ''
with subprocess.Popen(cmd_args, stdout=subprocess.PIPE, bufsize=1) as p:
for line in iter(p.stdout.readline, b''):
if PY2:
sys.stdout.write(line)
else:
sys.stdout.buffer.write(line)
try:
line_str = line.decode()
if line_str.rstrip() == '--- RESULTS ---':
output_block = True
elif line_str.rstrip() == '--- END RESULTS ---':
output_block = False
elif output_block:
output_content += line_str
except UnicodeError:
pass
p.communicate()
if p.returncode != 0:
raise SubcommandError('Command {} failed with return code {}'.format(cmd_args, p.returncode))
if not output_content:
raise SubcommandError('No results found in output for command {}'.format(cmd_args))
try:
for key, value in dict(json.loads(output_content)).items():
job.results[key] = value
except (TypeError, ValueError) as e:
raise_from(SubcommandError('Invalid results from command {}'.format(cmd_args)), e)
except (SubcommandError, json.JSONDecodeError):
t, e, tb = sys.exc_info()
if args.ignore_errors:
print(e)
else:
reraise(t, e, tb)
except schedy.errors.NoJobError:
break
if args.once:
break
def cmd_rm(args):
db = schedy.SchedyDB(config_path=args.config)
exp = db.get_experiment(args.experiment)
if args.job is None:
if not args.force:
print_exp(exp)
confirmation = input('Are you sure you want to remove {} and all its jobs? [y/N] '.format(exp.name))
if confirmation.lower() != 'y':
print('{} was not removed.'.format(exp.name))
return
exp.delete()
else:
job = exp.get_job(args.job)
if not args.force:
print_job(job)
confirmation = input('Are you sure you want to remove {}? [y/N] '.format(job.job_id))
if confirmation.lower() != 'y':
print('{} was not removed.'.format(job.job_id))
return
job.delete()
def cmd_gen_token(args):
config = {
'root': args.root,
'token_type': 'password',
}
if args.email is None:
config['email'] = input('Email: ')
else:
config['email'] = args.email
if args.password is None:
config['token'] = getpass.getpass('Password: '******'token'] = args.password
db = schedy.SchedyDB(config_override=config)
url = urljoin(db.root, 'resettoken/')
response = db._authenticated_request('POST', url=url)
schedy.errors._handle_response_errors(response)
new_content = response.json()
if args.config is None:
config_path = schedy.core._default_config_path()
else:
config_path = args.config
config_dir = os.path.dirname(config_path)
if config_dir:
try:
os.makedirs(config_dir)
except OSError:
t, e, tb = sys.exc_info()
if e.errno != errno.EEXIST:
reraise(t, e, tb)
pass
with open(config_path, 'w') as config_file:
config_file.write(json_dumps(new_content, cls=schedy.encoding.SchedyJSONEncoder))
try:
os.chmod(config_path, stat.S_IRUSR | stat.S_IWUSR)
except OSError:
print('Token file permissions could not be set.')
print('Your token has been saved to {}.'.format(config_path))
def cmd_push(args):
db = schedy.SchedyDB(config_path=args.config)
kwargs = dict()
# Status
if args.status is not None:
kwargs['status'] = args.status
# Results
if args.results is not None:
if len(args.results) % 2 != 0:
args.parser.error('Invalid results (not a list of name/value).')
results = dict()
for i in range(0, len(args.results), 2):
name = args.results[i]
value_txt = args.results[i + 1]
if name in results:
args.parser.error('Duplicate result: {}.'.format(name))
try:
results[name] = json.loads(value_txt)
except (TypeError, ValueError) as e:
args.parser.error('Invalid value for result {} ({!r}).'.format(name, e))
kwargs['results'] = results
# Hyperparameters
if len(args.hyperparameters) % 2 != 0:
args.parser.error('Invalid hyperparameters (not a list of name/value).')
hyperparameters = dict()
for i in range(0, len(args.hyperparameters), 2):
name = args.hyperparameters[i]
value_txt = args.hyperparameters[i + 1]
if name in hyperparameters:
args.parser.error('Duplicate hyperparameter: {}.'.format(name))
try:
hyperparameters[name] = json.loads(value_txt)
except (TypeError, ValueError) as e:
args.parser.error('Invalid value for hyperparameter {} ({!r}).'.format(name, e))
kwargs['hyperparameters'] = hyperparameters
exp = db.get_experiment(args.experiment)
job = exp.add_job(**kwargs)
print_job(job)
def cmd_list(args):
db = schedy.SchedyDB(config_path=args.config)
if args.experiment is None:
experiments = db.get_experiments()
table = exp_table(experiments)
else:
exp = db.get_experiment(args.experiment)
jobs = exp.all_jobs()
table = job_table(jobs)
if args.sort is not None:
try:
table.sort(args.sort, reverse=args.decreasing)
except KeyError as e:
args.parser.error(str(e))
if args.field is not None:
table.filter_fields(args.field)
if args.table:
table.print_table()
elif args.paragraph:
table.print_paragraphs()
else:
if args.field is None:
table.filter_categories([DEFAULT_CATEGORY])
table.print_table('plain', include_headers=False)
#!/usr/bin/env python
# -*- coding: utf-8 -*-
from __future__ import absolute_import, division, print_function, unicode_literals
import schedy
import time
db = schedy.SchedyDB()
experiment = db.get_experiment('MinimizeManual')
while True:
try:
# Pull the next job, and start working on it
# The with statement is there so that we always report to Schedy
# whether the job has crashed or succeeded
# The results will only be pushed to Schedy at the end of the with
# statement
with experiment.next_job() as job:
x = job.hyperparameters['x']
y = job.hyperparameters['y']
result = x**2 + y**2
job.results['result'] = result
# Catch any type of exception so that the worker never crashes
# This includes the NoJobError exception thrown by Schedy if there is no
# job queued for this experiment.
except Exception as e:
print(e)
# Wait a minute before issuing the next request
time.sleep(60)
