def init_trainer(self, args):
if args.gpuid:
print('Running with GPU {}.'.format(args.gpuid[0]))
cuda.set_device(args.gpuid[0])
else:
print('Running with CPU.')
if args.random_seed:
random.seed(args.random_seed + os.getpid())
np.random.seed(args.random_seed + os.getpid())
schema = Schema(args.schema_path)
scenario_db = ScenarioDB.from_dict(schema,
read_json(args.scenarios_path),
Scenario)
valid_scenario_db = ScenarioDB.from_dict(
schema, read_json(args.valid_scenarios_path), Scenario)
# if len(args.agent_checkpoints) == 0
# assert len(args.agent_checkpoints) <= len(args.agents)
if len(args.agent_checkpoints) < len(args.agents):
ckpt = [None] * 2
else:
ckpt = args.agent_checkpoints
systems = [
get_system(name, args, schema, False, ckpt[i])
for i, name in enumerate(args.agents)
]
rl_agent = 0
system = systems[rl_agent]
model = system.env.model
loss = None
# optim = build_optim(args, [model, system.env.critic], None)
optim = {
'model': build_optim(args, model, None),
'critic': build_optim(args, system.env.critic, None)
}
optim['critic']._set_rate(0.05)
scenarios = {
'train': scenario_db.scenarios_list,
'dev': valid_scenario_db.scenarios_list
}
from neural.a2c_trainer import RLTrainer as A2CTrainer
trainer = A2CTrainer(systems,
scenarios,
loss,
optim,
rl_agent,
reward_func=args.reward,
cuda=(len(args.gpuid) > 0),
args=args)
self.args = args
self.trainer = trainer
self.systems = systems
schema = Schema(args.schema_path)
scenario_db = ScenarioDB.from_dict(schema, read_json(args.scenarios_path),
Scenario)
valid_scenario_db = ScenarioDB.from_dict(
schema, read_json(args.valid_scenarios_path), Scenario)
assert len(args.agent_checkpoints) <= len(args.agents)
systems = [
get_system(name, args, schema, False, args.agent_checkpoints[i])
for i, name in enumerate(args.agents)
]
rl_agent = 0
system = systems[rl_agent]
model = system.env.model
loss = make_loss(args, model, system.mappings['tgt_vocab'])
optim = build_optim(args, model, None)
scenarios = {
'train': scenario_db.scenarios_list,
'dev': valid_scenario_db.scenarios_list
}
trainer = RLTrainer(systems,
scenarios,
loss,
optim,
rl_agent,
reward_func=args.reward,
reward_beta=args.reward_beta)
trainer.learn(args)
parser.add_argument('--critic-path', default=None, help='Output path for the critic Model')
cocoa.options.add_scenario_arguments(parser)
options.add_system_arguments(parser)
options.add_rl_arguments(parser)
options.add_model_arguments(parser)
args = parser.parse_args()
if args.random_seed:
random.seed(args.random_seed)
np.random.seed(args.random_seed)
schema = Schema(args.schema_path)
scenario_db = ScenarioDB.from_dict(schema, read_json(args.scenarios_path), Scenario)
valid_scenario_db = ScenarioDB.from_dict(schema, read_json(args.valid_scenarios_path), Scenario)
assert len(args.agent_checkpoints) <= len(args.agents)
systems = [get_system(name, args, schema, False, args.agent_checkpoints[i]) for i, name in enumerate(args.agents)]
rl_agent = 0
system = systems[rl_agent]
model = system.env.model
critic_model = system.env.critic_model
loss = make_loss(args, model, system.mappings['tgt_vocab'])
optim = build_optim(args, model, None)
critic_optim = build_optim(args, critic_model, None)
scenarios = {'train': scenario_db.scenarios_list, 'dev': valid_scenario_db.scenarios_list}
trainer = CriticTrainer(systems, scenarios, loss, critic_optim, rl_agent, reward_func=args.reward, cuda=(len(args.gpuid) > 0))
trainer.learn(args)
ckpt = [None] * 2
else:
ckpt = args.agent_checkpoints
systems = [
get_system(name, args, schema, False, ckpt[i])
for i, name in enumerate(args.agents)
]
rl_agent = 0
system = systems[rl_agent]
model = system.env.model
loss = None
# optim = build_optim(args, [model, system.env.critic], None)
optim = {
'model': build_optim(args, model, None),
'critic': build_optim(args, system.env.critic, None)
}
optim['critic']._set_rate(0.05)
scenarios = {
'train': scenario_db.scenarios_list,
'dev': valid_scenario_db.scenarios_list
}
from neural.a2c_trainer import RLTrainer as A2CTrainer
trainer = A2CTrainer(systems,
scenarios,
loss,
optim,
rl_agent,
reward_func=args.reward,