def run_episodes(name, new_episode_func, update_func):
global behavior_tracker, helper
for e in range(behavior_tracker.num_episodes):
# gets state and action sequence
s_seq = behavior_tracker.s_s[e]
a_seq = behavior_tracker.s_a[e]
seq_len = len(s_seq)
# signals new episode
new_episode_func(e, seq_len)
old_obs = env.reset()
helper.get_state_from_observation(old_obs, 0, False)
clean_console()
print('Processing {} episode {}...'.format(name, e))
for t in range(seq_len):
# gets state-action pair
old_s = s_seq[t]
a = a_seq[t]
# steps the environment, gets next state
obs, r, done, _ = env.step(a)
s = helper.get_state_from_observation(obs, r, done)
r = helper.get_reward(old_s, a, r, s, done)
# checks possible synchronization errors
if done and t != seq_len - 1:
raise ValueError(
'Environment ended at {}, before tracked behavior which ended at: {}'
.format(t, seq_len - 1))
if t == seq_len - 1 and not done:
raise ValueError(
'Environment did not end at {} like it was supposed to'.
format(t))
if not done and s != s_seq[t + 1]:
raise ValueError(
'Environment state {} does not match tracked state {}'.
format(s, s_seq[t + 1]))
if name == 'analysis' and explanation_t == ReportType.Heatmaps:
helper.update_stats(e, t, old_obs, obs, old_s, a, r, s)
update_func(t, old_obs, old_s, a, r, s)
old_obs = obs
agent.load(agent_dir)
# runs all episodes
advance_time = advance_episode = False
for e in range(behavior_tracker.num_episodes):
# gets state and action sequence
s_seq = behavior_tracker.s_s[e]
a_seq = behavior_tracker.s_a[e]
advance_episode = e < START_EPISODE
old_obs = env.reset()
old_s = helper.get_state_from_observation(old_obs, 0, False)
seq_len = len(s_seq)
clean_console()
print('Replaying episode {} ({} time-steps)...'.format(e, seq_len))
for t in range(seq_len):
# waits until one of the keys is pressed if not advancing episode
while not advance_episode and not advance_time:
process_keys()
pygame.event.pump()
advance_time = False
# hides the display when advancing an episode
if advance_episode:
pygame.display.iconify()
else:
def update_stats_episode(self, e, path=None):
if (e + 1) % (self.config.num_episodes / 100) == 0:
clean_console()
print('Episode {} ({:.0f}%)...'.format(
e + 1, ((e + 1) / self.config.num_episodes) * 100))
self._print_stats(e, PRINT_SCREEN_VAR_NAMES)