def train_val(args, agent, train_env, val_envs):
''' Train on the training set, and validate on seen and unseen splits. '''
m_dict = {
'follower': [agent.encoder,agent.decoder],
'pm': [agent.prog_monitor],
'follower+pm': [agent.encoder, agent.decoder, agent.prog_monitor],
'all': agent.modules()
}
if agent.scorer:
m_dict['scorer_all'] = agent.scorer.modules()
m_dict['scorer_scorer'] = [agent.scorer.scorer]
if agent.bt_button:
m_dict['bt_button'] = [agent.bt_button]
optimizers = [optim.Adam(filter_param(m), lr=learning_rate,
weight_decay=weight_decay) for m in m_dict[args.grad] if len(filter_param(m))]
if args.use_pretraining:
_train(args, pretrain_env, agent, optimizers,
args.n_pretrain_iters, val_envs=val_envs)
_train(args, train_env, agent, optimizers,
args.n_iters, val_envs=val_envs)
x_1.append(i)
x_2.append(j)
valid_points += 1
batch_labels.append((x_1, x_2))
print('valid points', valid_points)
###
from utils import filter_param
m_dict = {
'follower': [agent.encoder, agent.decoder],
}
optimizers = [
optim.Adam(filter_param(m), lr=0.0001, weight_decay=0.0005)
for m in m_dict['follower'] if len(filter_param(m))
]
###
def eval(test_envs, agent):
for env_name, (val_env, evaluator) in test_envs.items():
agent.env = val_env
if hasattr(agent, 'speaker') and agent.speaker:
agent.speaker.env = val_env
agent.search = True
agent.search_logit = True
agent.search_mean = False
agent.search_early_stop = True