actions.append(a)
dones.append(d)
rewards.append(r)
logp_ts.append(logp_t)
o = n_o
if d:
ep += 1
if ep % record_score_size == 0:
if int(ep / record_score_size) < 600:
writer.add_scalar('data/reward', record_score / record_score_size, int(ep / record_score_size))
record_score = 0
writer.add_scalar('data/reward_per_episode', score, ep)
print(score, ep)
score = 0
o, r, d, ep_ret, ep_len = env.reset(), 0, False, 0, 0
a, v_t, logp_t = agent.get_action([o])
values.append(v_t[0])
next_value = values[1:]
value = values[:-1]
adv, target = get_gaes(rewards, dones, value, next_value, agent.gamma, agent.lamda, False)
value_loss, kl, ent = agent.update(states, actions, target, adv, logp_ts)
writer.add_scalar('data/value_loss_per_rollout', value_loss, rollout)
writer.add_scalar('data/kl_per_rollout', kl, rollout)
writer.add_scalar('data/ent_per_rollout', ent, rollout)
writer.add_scalar('data/reward_per_rollout', score_rollout, rollout)
values, states, actions, dones, logp_ts, rewards = [], [], [], [], [], []
print(episode, score)
score = 0
states = next_states
total_state = np.stack(total_state).transpose([1, 0, 2
]).reshape([-1, state_size])
total_next_state = np.stack(total_next_state).transpose([1, 0, 2]).reshape(
[-1, state_size])
total_reward = np.stack(total_reward).transpose().reshape([-1])
total_done = np.stack(total_done).transpose().reshape([-1])
total_action = np.stack(total_action).transpose().reshape([-1])
total_target, total_adv = [], []
for idx in range(num_worker):
value, next_value = agent.get_value(
total_state[idx * num_step:(idx + 1) * num_step],
total_next_state[idx * num_step:(idx + 1) * num_step])
adv, target = get_gaes(
total_reward[idx * num_step:(idx + 1) * num_step],
total_done[idx * num_step:(idx + 1) * num_step], value, next_value,
agent.gamma, agent.lamda, normalize)
total_target.append(target)
total_adv.append(adv)
agent.train_model(total_state, total_action, np.hstack(total_target),
np.hstack(total_adv))
writer.add_scalar('data/reward_per_rollout',
sum(total_reward) / (num_worker), global_update)
saver.save(sess, 'lunarlander_a2c/model')
score += reward
total_state.append(state)
total_next_state.append(next_state)
total_done.append(done)
total_reward.append(reward)
total_action.append(action)
state = next_state
if ep % train_size == 0:
update_step += 1
total_state = np.stack(total_state)
total_next_state = np.stack(total_next_state)
total_reward = np.stack(total_reward)
total_done = np.stack(total_done)
total_action = np.stack(total_action)
value, next_value = agent.get_value(total_state, total_next_state)
adv, target = utils.get_gaes(total_reward, total_done, value,
next_value, agent.gamma, agent.lamda,
False)
agent.train_model(total_state, total_action, target, adv)
print(update_step, score / train_size)
if update_step < 300:
writer.add_scalar('data/reward', score / train_size, update_step)
saver.save(sess, 'pendulum_ppo/model')
total_state, total_reward, total_done, total_next_state, total_action = [], [], [], [], []
score = 0
