def train(worker_id, agent, hparams, checkpoint):
env = get_env(hparams)
state = env.reset()
while hparams.global_step < hparams.train_steps:
hparams.mode[worker_id] = ModeKeys.TRAIN
last_state = state
action, reward, done, state = step(
hparams, agent, last_state, env, worker_id)
agent.observe(last_state, action, reward, done, state, worker_id)
if done:
hparams.local_episode[worker_id] += 1
log_scalar('episodes/worker_%d' % worker_id,
hparams.local_episode[worker_id])
hparams.global_step += 1
hparams.total_step += 1
hparams.local_step[worker_id] += 1
update_learning_rate(hparams)
if hparams.local_step[worker_id] % hparams.eval_interval == 0:
agent.reset(worker_id)
evaluate(worker_id, agent, env, hparams)
if worker_id == 0:
checkpoint.save()
state = env.reset()
agent.reset(worker_id)
env.close()
def train(worker_id, agent, hparams, checkpoint):
env = get_env(hparams)
while hparams.global_step < hparams.train_steps:
hparams.mode[worker_id] = ModeKeys.TRAIN
state = env.reset()
done = False
while not done:
last_state = state
action, reward, done, state = steps(hparams, agent, last_state, env,
hparams.n_steps, worker_id)
agent.observe(last_state, action, reward, done, state, worker_id)
if done:
hparams.local_episode[worker_id] += 1
log_scalar('episodes/worker_%d' % worker_id,
hparams.local_episode[worker_id])
if hparams.local_step[worker_id] % hparams.test_interval == 0:
test(worker_id, agent, env, hparams)
if worker_id == 0:
checkpoint.save()
done = True
hparams.global_step += 1
hparams.total_step += 1
hparams.local_step[worker_id] += 1
env.close()
def test(worker_id, agent, env, hparams):
hparams.mode[worker_id] = ModeKeys.TEST
rewards = []
for i in range(hparams.test_episodes):
state = env.reset()
done = False
episode_reward = 0
while not done:
last_state = state
action, reward, done, state = steps(
hparams, agent, last_state, env, n=1, worker_id=worker_id)
episode_reward += reward
hparams.total_step += 1
rewards.append(episode_reward)
log_scalar('rewards/worker_%d' % worker_id, np.mean(rewards))
def evaluate(worker_id, agent, env, hparams):
hparams.mode[worker_id] = ModeKeys.EVAL
rewards = []
for i in range(hparams.eval_episodes):
state = env.reset()
recurrent_state = np.zeros((512,))
done = False
episode_reward = 0
while not done:
last_recurrent_state = recurrent_state
last_state = state
action, reward, done, state, recurrent_state = step(
hparams, agent, last_state, env, worker_id, recurrent_state)
episode_reward += reward
hparams.total_step += 1
rewards.append(episode_reward)
log_scalar('rewards/worker_%d' % worker_id, np.mean(rewards))
def train(worker_id, agent, hparams, checkpoint):
env = get_env(hparams)
recurrent_state = None
last_recurrent_state = None
state = env.reset()
if type(env).__name__ == 'NavRLEnv':
recurrent_state = np.zeros((512,))
while hparams.global_step < hparams.train_steps:
hparams.mode[worker_id] = ModeKeys.TRAIN
last_state = state
if type(env).__name__ == 'NavRLEnv':
last_recurrent_state = recurrent_state
action, reward, done, state, recurrent_state = step(hparams, agent,
last_state, env,
worker_id,
recurrent_state)
agent.observe(last_state, action, reward, done, state, worker_id,
last_recurrent_state, recurrent_state)
if done:
hparams.local_episode[worker_id] += 1
log_scalar('episodes/worker_%d' % worker_id,
hparams.local_episode[worker_id])
if hparams.local_step[worker_id] % hparams.eval_interval == 0:
agent.reset(worker_id)
evaluate(worker_id, agent, env, hparams)
if worker_id == 0:
checkpoint.save()
state = env.reset()
agent.reset(worker_id)
hparams.global_step += 1
hparams.total_step += 1
hparams.local_step[worker_id] += 1
update_learning_rate(hparams)
env.close()