def setup_train_dir(logdir, overwrite=False, force_train=True):
"""Setups directory for training."""
tf.io.gfile.makedirs(logdir)
config_path = os.path.join(logdir, 'config.json')
if not os.path.exists(config_path) or overwrite:
logging.info(
'Using the existing passed in config as no config.json file exists in '
'%s', logdir)
with tf.io.gfile.GFile(config_path, 'w') as config_file:
config = dict([(k, to_dict(v)) for k, v in CONFIG.items()])
json.dump(config, config_file, sort_keys=True, indent=4)
else:
logging.info('Using config from config.json that exists in %s.',
logdir)
with tf.io.gfile.GFile(config_path, 'r') as config_file:
config_dict = json.load(config_file)
CONFIG.update(config_dict)
train_logs_dir = os.path.join(logdir, 'train.logs')
if os.path.exists(train_logs_dir) and not force_train:
raise ValueError(
'You might be overwriting a directory that already '
'has train_logs. Please provide a new logdir name in '
'config or pass --force_train while launching script.')
tf.io.gfile.makedirs(train_logs_dir)
def main():
print("copying files")
if os.path.isdir(SERVICES_PATH):
shutil.rmtree(SERVICES_PATH)
shutil.copytree('../services', SERVICES_PATH)
os.chdir(SERVICES_PATH)
for service_name, service_settings in CONFIG.items():
print("deploying", service_name)
docker_port, external_port = find_ports(service_name)
os.chdir(os.path.join(SERVICES_PATH, service_name))
os.system('sudo docker-compose build')
config = CONFIG[service_name]
if not config.get('nonginx', False):
with open(
os.path.join(NGINX_CONF_PATH, 'sites-available',
service_name), 'w') as nginx_conf:
nginx_conf.write(
render_nginx_conf(
service_name,
external_port,
docker_port,
config.get('static_dir_path', None),
os.path.join(os.getcwd(), config['static_dir_path'])
if 'static_dir_path' in config else None,
))
os.chdir('..')
os.system('ln -s {} {}'.format(
os.path.join(NGINX_CONF_PATH, 'sites-available', service_name),
os.path.join(NGINX_CONF_PATH, 'sites-enabled', service_name),
))
def main(filename=f"availabilities_{date.today().strftime('%Y-%m-%d')}.json"):
availabilities = {}
for name, info in CONFIG.items():
fn_string = f"{info['function']}(**{info['args']})"
is_any_available = eval(fn_string)
availabilities[name] = is_any_available
with open(filename, "w") as f:
json.dump(availabilities, f)
def setup_train_dir(logdir):
"""Setups directory for training."""
tf.io.gfile.makedirs(logdir)
config_path = os.path.join(logdir, 'config.yml')
if not os.path.exists(config_path):
logging.info(
'Using config from config.py as no config.yml file exists in '
'%s', logdir)
with tf.io.gfile.GFile(config_path, 'w') as config_file:
config = dict([(k, to_dict(v)) for k, v in CONFIG.items()])
yaml.safe_dump(config, config_file, default_flow_style=False)
else:
logging.info('Using config from config.yml that exists in %s.', logdir)
with tf.io.gfile.GFile(config_path, 'r') as config_file:
config_dict = yaml.safe_load(config_file)
CONFIG.update(config_dict)
train_logs_dir = os.path.join(logdir, 'train_logs')
if os.path.exists(train_logs_dir) and not FLAGS.force_train:
raise ValueError(
'You might be overwriting a directory that already '
'has train_logs. Please provide a new logdir name in '
'config or pass --force_train while launching script.')
tf.io.gfile.makedirs(train_logs_dir)
svc.setServiceParent(ret)
# DO IT NOW
ret.setServiceParent(service.IServiceCollection(application))
# Load the config
try:
from config import CONFIG # pylint:disable=no-name-in-module,import-error
except ImportError:
CONFIG = {}
# Merge user config over defaults
DEFAULT_CONFIG = {
'docker_url': None,
'bind_interface': '',
'bind_port': 53,
'bind_protocols': ['tcp', 'udp'],
'no_nxdomain': True,
'authoritive': True,
}
CONFIG = dict(DEFAULT_CONFIG.items() + CONFIG.items())
application = service.Application('dnsserver', 1, 1) # noqa pylint:disable=invalid-name
main()
# Doin' it wrong
if __name__ == '__main__':
import sys
print "Usage: twistd -y %s" % sys.argv[0]
def main():
logger.configure('./{}_logs'.format(C['env_id']))
for k, v in C.items():
logger.record_tabular(k, v)
logger.dump_tabular()
train_tracker = [0.0]
eval_tracker = []
best_reward = 0
sess = tf.InteractiveSession()
train_reward = tf.placeholder(tf.float32, name='train_reward')
eval_reward = tf.placeholder(tf.float32, name='eval_reward')
train_env = make_env(C['env_id'], C['noop_max'])
eval_env = make_env(C['env_id'], C['noop_max'])
agent = Agent(train_env, C)
sess.run(tf.global_variables_initializer())
agent.nn.update_target()
train_summary = tf.summary.scalar('train_rew', train_reward)
eval_summary = tf.summary.scalar('eval_reward', eval_reward)
writer = tf.summary.FileWriter('{}{}_summary'.format('./', C['env_id']),
sess.graph)
train_fs = reset_fs()
train_s = train_env.reset()
for it in range(C['iterations']):
# Training
train_fs.append(train_s)
train_a = agent.act(np.transpose(train_fs, (1, 2, 0)))
ns, train_r, train_d, _ = train_env.step(train_a)
train_tracker[-1] += train_r
agent.perceive(train_s, train_a, train_r, float(train_d), it)
train_s = ns
if train_d:
if train_env.env.env.was_real_done:
if len(train_tracker) % 100 == 0:
summary = sess.run(train_summary,
feed_dict={
train_reward:
np.mean(train_tracker[-100:])
})
writer.add_summary(summary, it)
logger.record_tabular('steps', it)
logger.record_tabular('episode', len(train_tracker))
logger.record_tabular('epsilon', 100 * agent.epsilon)
logger.record_tabular('learning rate', agent.lr)
logger.record_tabular('mean 100 episodes',
np.mean(train_tracker[-100:]))
logger.dump_tabular()
train_tracker.append(0.0)
train_fs = reset_fs()
train_s = train_env.reset()
# Evaluation
if it % C['eval_freq'] == 0:
for _ in range(C['eval_episodes']):
temp_video = []
temp_reward = 0
eval_tracker.append(0.0)
eval_fs = reset_fs()
eval_s = eval_env.reset()
while True:
temp_video.append(eval_s)
eval_fs.append(eval_s)
eval_a = agent.greedy_act(np.transpose(eval_fs, (1, 2, 0)))
eval_s, eval_r, eval_d, _ = eval_env.step(eval_a)
eval_tracker[-1] += eval_r
if eval_env.env.env.was_real_done:
break
if eval_d:
eval_fs = reset_fs()
eval_s = eval_env.reset()
if eval_tracker[-1] > best_reward: # Save best video
best_reward = eval_tracker[-1]
logger.log(
'Dump best video reward: {}'.format(best_reward))
best_video = temp_video
with open('video.pkl', 'wb') as f:
pickle.dump(best_video,
f,
protocol=pickle.HIGHEST_PROTOCOL)
logger.log(
'Evaluate mean reward: {:.2f}, max reward: {:.2f}, std: {:.2f}'
.format(np.mean(eval_tracker[-C['eval_episodes']:]),
np.max(eval_tracker[-C['eval_episodes']:]),
np.std(eval_tracker[-C['eval_episodes']:])))
summary = sess.run(eval_summary,
feed_dict={
eval_reward:
np.mean(eval_tracker[-C['eval_episodes']:])
})
writer.add_summary(summary, it)
agent.nn.save('./{}_model'.format(C['env_id']))
svc.setServiceParent(ret)
# DO IT NOW
ret.setServiceParent(service.IServiceCollection(application))
# Load the config
try:
from config import CONFIG # pylint:disable=no-name-in-module,import-error
except ImportError:
CONFIG = {}
# Merge user config over defaults
DEFAULT_CONFIG = {
'docker_url': None,
'bind_interface': '',
'bind_port': 53,
'bind_protocols': ['tcp', 'udp'],
'no_nxdomain': True,
'authoritive': True,
}
CONFIG = dict(DEFAULT_CONFIG.items() + CONFIG.items())
application = service.Application('dnsserver', 1, 1) # noqa pylint:disable=invalid-name
main()
# Doin' it wrong
if __name__ == '__main__':
import sys
print "Usage: twistd -y %s" % sys.argv[0]
