def worker_policy(args, manager, config):
init_logging_handler(args.log_dir, '_policy')
agent = GAN(None, args, manager, config, 0, pre=True)
best = float('inf')
for e in range(args.epoch):
agent.imitating(e)
best = agent.imit_test(e, best)
def worker_user(args, manager, config):
init_logging_handler(args.log_dir, '_user')
env = UserNeural(args, manager, config, True)
best = float('inf')
for e in range(args.epoch):
env.imitating(e)
best = env.imit_test(e, best)
def worker_estimator(args, manager, config, make_env):
init_logging_handler(args.log_dir, '_estimator')
agent = DiaSeq(make_env, args, manager, config, args.process, pre_irl=True)
agent.load(args.save_dir+'/best')
best0, best1 = float('inf'), float('inf')
for e in range(args.epoch):
agent.train_disc(e, args.batchsz_traj)
best0 = agent.test_disc(e, args.batchsz, best0)
def worker_policy_usr(args, manager, config):
init_logging_handler(args.log_dir, '_policy_usr')
if args.config == 'multiwoz':
print("MultiWoz Agent Usr")
agent = Policy(None, args, manager, config, 0, 'usr', True)
elif args.config == 'dstcsgds':
print("DSTC Agent Usr")
agent = DSTCPolicy(None, args, manager, config, 0, 'usr', True)
else:
raise NotImplementedError(
'Policy usr of the dataset {} not implemented'.format(args.config))
best = float('inf')
for e in range(2):
agent.imitating(e)
best = agent.imit_test(e, best)
def main(config):
parser = get_parser()
argv = sys.argv[1:]
args, _ = parser.parse_known_args(argv)
init_logging_handler(config.log_dir)
logging.info(args)
config = update_cfg(config, args)
logging.info("Start initializing")
irl_model = RewardModule(config).to(device=device) # this is the reward model only, which will be fed to RewardEstimator.
reward_agent = RewardEstimator(config=config, irl_model=irl_model)
user_policy = ActorCriticDiscrete(config).to(device=device)
user_policy = init_net(user_policy)
user_ppo = PPO(config, user_policy)
system_policy = ActorCriticContinuous(config).to(device=device)
system_policy = init_net(system_policy)
init_system_policy = ActorCriticContinuous(config).to(device=device)
init_system_policy.load_state_dict(system_policy.state_dict())
system_ppo = PPO(config, system_policy, init_policy=init_system_policy)
# reward_true = RewardTruth(config).to(device=device) # this is the ground truth which will not be updated once randomly initialized.
reward_true = RewardTruthSampled(config).to(device)
reward_true = init_net(reward_true)
logging.info("Finish building module: reward agent, user ppo, system ppo")
main_agent = InteractAgent(config=config,
user_agent=user_ppo,
user_reward=reward_agent,
system_agent=system_ppo,
reward_groundtruth=reward_true
)
for e_id in range(config.master_epochs):
main_agent.master_train(e_id)
# for _ in range(3):
# main_agent.system_train()
# raise ValueError("stop here")
logging.info("@@@@@@@@@@ Finished @@@@@@@@@@@")
if __name__ == '__main__':
parser = get_parser()
argv = sys.argv[1:]
args, _ = parser.parse_known_args(argv)
if args.config == 'multiwoz':
print("MultiWoz Config")
config = MultiWozConfig()
elif args.config == 'dstcsgds':
print("DSTC Config")
config = DSTCSGDSConfig(args.data_dir)
else:
raise NotImplementedError(
'Config of the dataset {} not implemented'.format(args.config))
init_logging_handler(args.log_dir)
logging.debug(str(args))
try:
mp = mp.get_context('spawn')
except RuntimeError:
pass
if args.pretrain:
logging.debug('pretrain')
if args.config == 'dstcsgds':
print("DSTC Manager")
manager = DSTCDataManager(args.data_dir, config)
elif args.config == 'multiwoz':
print("MultiWoz Manager")
manager = DataManager(args.data_dir, config)
import sys
from fabric.api import *
from fabric.colors import green, red
# append the home path
PROJECT_HOME = os.path.dirname(
os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
sys.path.append(PROJECT_HOME)
from settings import (hosts, user, key_filename, DirConfig as DC)
from utils import init_logging_handler, timestamp
env.hosts, env.user, env.key_filename = hosts, user, key_filename
logger = init_logging_handler()
def prepare_deploy():
with lcd(DC.ANALYSIS_DIR):
local("ipython analyze.py")
def zip_files():
with lcd(DC.LOCAL_RELATIVE_DATA_DIR + 'data_' + timestamp()):
local('tar -czf data.tar.gz ./')
local('mv data.tar.gz ../../automatic_deploy/deploy_files')
def upload():
# upload the zipped files