def generate_dagger_demos(env_name, seeds, fail_obss, fail_actions, mean_steps):
env = gym.make(env_name)
agent = BotAgent(env)
demos = []
for i in range(len(fail_obss)):
# Run the expert for one episode
env.seed(int(seeds[i]))
new_obs = env.reset()
agent.on_reset()
env0_str = env.__str__()
actions = []
images = []
directions = []
debug_info = {'seed': [int(seeds[i])], 'actions': []}
try:
for j in range(min(int(args.dagger_trim_coef * mean_steps), len(fail_obss[i]) - 1)):
obs = fail_obss[i][j]
assert check_obss_equality(obs, new_obs), "Observations {} of seed {} don't match".format(j, seeds[i])
mission = obs['mission']
action = agent.act(update_internal_state=False)['action']
_ = agent.bot.take_action(fail_actions[i][j])
debug_info['actions'].append(fail_actions[i][j])
new_obs, reward, done, _ = env.step(fail_actions[i][j])
if done and reward > 0:
raise ValueError(
"The baby's actions shouldn't solve the task. Env0 {}, Env9{}, Seed {}, actions {}.".format(
env0_str, env.__str__(), int(seeds[i]), fail_actions[i]
))
actions.append(action)
images.append(obs['image'])
directions.append(obs['direction'])
if args.continue_dagger:
obs = new_obs
while not done:
action = agent.act(obs)['action']
debug_info['actions'].append(action)
new_obs, reward, done, _ = env.step(action)
agent.analyze_feedback(reward, done)
actions.append(action)
images.append(obs['image'])
directions.append(obs['direction'])
print(debug_info, actions)
demos.append((mission, blosc.pack_array(np.array(images)), directions, actions))
except Exception as e:
logger.exception("error while generating demo #{}: {}. Env0 {}, Env9{}, Seed {}, actions {}.".format(
len(demos), e, env0_str, env.__str__(), int(seeds[i]), fail_actions[i]))
continue
return demos
def generate_demos(env_name, seeds):
env = gym.make(env_name)
agent = BotAgent(env)
demos = []
for seed in seeds:
# Run the expert for one episode
done = False
env.seed(int(seed))
obs = env.reset()
agent.on_reset()
actions = []
mission = obs["mission"]
images = []
directions = []
try:
while not done:
action = agent.act(obs)['action']
new_obs, reward, done, _ = env.step(action)
agent.analyze_feedback(reward, done)
actions.append(action)
images.append(obs['image'])
directions.append(obs['direction'])
obs = new_obs
if reward > 0:
demos.append((mission, blosc.pack_array(np.array(images)),
directions, actions))
if reward == 0:
logger.info("failed to accomplish the mission")
except Exception:
logger.exception("error while generating demo #{}".format(
len(demos)))
continue
# logger.info("demo #{}".format(len(demos)))
return demos