def __init__(self, domain: Domain = None, subgraph: dict = None,
logger : DiasysLogger = DiasysLogger()):
Module.__init__(self, domain, subgraph, logger = logger)
self.initial_turn = True
self.description = ''
init_gui()
self.dialogs = 0
def __init__(self,
domain: Domain = None,
subgraph: dict = None,
logger: DiasysLogger = DiasysLogger(),
language: Language = None):
Module.__init__(self, domain, subgraph,
logger=logger) # independent of the domain
self.language = language
def __init__(self,
*modules,
domain=None,
logger: DiasysLogger = DiasysLogger()):
self.domain = domain
self.logger = logger
self.modules = modules
self.is_training = False
self.num_dialogs = 0
self.num_turns = 0
def train(domain_name: str, log_to_file: bool, seed: int, train_epochs: int, train_dialogs: int,
eval_dialogs: int, max_turns: int, train_error_rate: float, test_error_rate: float,
lr: float, eps_start: float, grad_clipping: float, buffer_classname: str,
buffer_size: int, use_tensorboard: bool):
common.init_random(seed=seed)
file_log_lvl = LogLevel.DIALOGS if log_to_file else LogLevel.NONE
logger = DiasysLogger(console_log_lvl=LogLevel.RESULTS, file_log_lvl=file_log_lvl)
if buffer_classname == "prioritized":
buffer_cls = NaivePrioritizedBuffer
elif buffer_classname == "uniform":
buffer_cls = UniformBuffer
domain = JSONLookupDomain(name=domain_name)
bst = HandcraftedBST(domain=domain, logger=logger)
user = HandcraftedUserSimulator(domain, logger=logger)
noise = SimpleNoise(domain=domain, train_error_rate=train_error_rate,
test_error_rate=test_error_rate, logger=logger)
policy = DQNPolicy(domain=domain, lr=lr, eps_start=eps_start,
gradient_clipping=grad_clipping, buffer_cls=buffer_cls,
replay_buffer_size=buffer_size, train_dialogs=train_dialogs,
logger=logger)
evaluator = PolicyEvaluator(domain=domain, use_tensorboard=use_tensorboard,
experiment_name=domain_name, logger=logger)
ds = DialogSystem(policy,
user,
noise,
bst,
evaluator)
for epoch in range(train_epochs):
# train one epoch
ds.train()
evaluator.start_epoch()
for episode in range(train_dialogs):
ds.run_dialog(max_length=max_turns)
evaluator.end_epoch() # important for statistics!
ds.num_dialogs = 0 # IMPORTANT for epsilon scheduler in dqnpolicy
policy.save() # save model
# evaluate one epoch
ds.eval()
evaluator.start_epoch()
for episode in range(eval_dialogs):
ds.run_dialog(max_length=max_turns)
evaluator.end_epoch() # important for statistics!
ds.num_dialogs = 0 # IMPORTANT for epsilon scheduler in dqnpolicy
def __init__(self, domain=None, subgraph: dict = None,
logger : DiasysLogger = DiasysLogger()):
Module.__init__(self, domain, subgraph, logger = logger)
self.initial_turn = True
init_gui() # initialise the GUI
def __init__(self, domain: Domain = None, subgraph: dict = None,
logger : DiasysLogger = DiasysLogger()):
Module.__init__(self, domain, subgraph, logger = logger)
init_gui()
self.dialogs = 0
def train(domain_name: str, log_to_file: bool, seed: int, train_epochs: int, train_dialogs: int,
eval_dialogs: int, max_turns: int, train_error_rate: float, test_error_rate: float,
lr: float, eps_start: float, grad_clipping: float, buffer_classname: str,
buffer_size: int, use_tensorboard: bool):
"""
Training loop for the RL policy, for information on the parameters, look at the descriptions
of commandline arguments in the "if main" below
"""
seed = seed if seed != -1 else None
common.init_random(seed=seed)
file_log_lvl = LogLevel.DIALOGS if log_to_file else LogLevel.NONE
logger = DiasysLogger(console_log_lvl=LogLevel.RESULTS, file_log_lvl=file_log_lvl)
summary_writer = SummaryWriter(log_dir='logs') if use_tensorboard else None
if buffer_classname == "prioritized":
buffer_cls = NaivePrioritizedBuffer
elif buffer_classname == "uniform":
buffer_cls = UniformBuffer
domain = JSONLookupDomain(name=domain_name)
bst = HandcraftedBST(domain=domain, logger=logger)
user = HandcraftedUserSimulator(domain, logger=logger)
# noise = SimpleNoise(domain=domain, train_error_rate=train_error_rate,
# test_error_rate=test_error_rate, logger=logger)
policy = DQNPolicy(domain=domain, lr=lr, eps_start=eps_start,
gradient_clipping=grad_clipping, buffer_cls=buffer_cls,
replay_buffer_size=buffer_size, train_dialogs=train_dialogs,
logger=logger, summary_writer=summary_writer)
evaluator = PolicyEvaluator(domain=domain, use_tensorboard=use_tensorboard,
experiment_name=domain_name, logger=logger,
summary_writer=summary_writer)
ds = DialogSystem(services=[user, bst, policy, evaluator], protocol='tcp')
# ds.draw_system_graph()
error_free = ds.is_error_free_messaging_pipeline()
if not error_free:
ds.print_inconsistencies()
for j in range(train_epochs):
# START TRAIN EPOCH
evaluator.train()
policy.train()
evaluator.start_epoch()
for episode in range(train_dialogs):
if episode % 100 == 0:
print("DIALOG", episode)
logger.dialog_turn("\n\n!!!!!!!!!!!!!!!! NEW DIALOG !!!!!!!!!!!!!!!!!!!!!!!!!!!!\n\n")
ds.run_dialog(start_signals={f'user_acts/{domain.get_domain_name()}': []})
evaluator.end_epoch()
policy.save()
# START EVAL EPOCH
evaluator.eval()
policy.eval()
evaluator.start_epoch()
for episode in range(eval_dialogs):
logger.dialog_turn("\n\n!!!!!!!!!!!!!!!! NEW DIALOG !!!!!!!!!!!!!!!!!!!!!!!!!!!!\n\n")
ds.run_dialog(start_signals={f'user_acts/{domain.get_domain_name()}': []})
evaluator.end_epoch()
ds.shutdown()
def __init__(self, domain, logger=DiasysLogger()):
# only calls super class' constructor
super(WeatherNLU, self).__init__(domain, debug_logger=logger)
class TurnDiff(object):
def __init__(self, module_idx, module_name, diff_dict):
self.module_idx = module_idx
self.module_name = module_name
self.diff_dict = diff_dict
def to_dict(self):
return {
"index": self.module_idx,
"name": self.module_name,
"diff": self.diff_dict
}
ACTIVE_LANGUAGE = Language.ENGLISH
logger = DiasysLogger()
# TODO make domain configurable via command line parameters
domain = JSONLookupDomain('ImsCourses')
nlu = HandcraftedNLU(domain=domain, logger=logger, language=ACTIVE_LANGUAGE)
bst = HandcraftedBST(domain=domain, logger=logger)
# policy = DQNPolicy(domain=domain, logger=self.logger)
policy = HandcraftedPolicy(domain=domain, logger=logger)
# policy.load()
policy.eval()
nlg = HandcraftedNLG(domain=domain, logger=logger, language=ACTIVE_LANGUAGE)
class DialogSystemWrapper(DialogSystem):
SESSIONS = {}
def __init__(self, uid):
help=
"maximum turns per dialog (dialogs with more turns will be terminated and counting as failed"
)
parser.add_argument("-eer",
"--evalerror",
type=float,
default=0.0,
help="simulation error rate while evaluating")
args = parser.parse_args()
# init random generator and logger
common.init_random(args.randomseed)
file_log_lvl = LogLevel.DIALOGS if args.logtofile else LogLevel.NONE
logger = DiasysLogger(console_log_lvl=LogLevel.RESULTS,
file_log_lvl=file_log_lvl)
# choose 'real' domain name from shorthand
if args.domain == 'courses':
domain_name = 'ImsCourses'
elif args.domain == 'lecturers':
domain_name = 'ImsLecturers'
test_hdc_usersim(domain_name=domain_name,
logger=logger,
eval_epochs=args.epochs,
eval_dialogs=args.evaldialogs,
max_turns=args.maxturns,
test_error=args.evalerror,
use_tensorboard=args.logtensorboard)