def save_videos(self, step, policy, save_name='sample_video', num_rollouts=2, teacher_dict={}, save_video=False,
log_prefix=None, stochastic=True):
policy.eval()
self.env.set_level_distribution(self.curriculum_step)
save_wandb = (save_video and not self.is_debug)
paths, accuracy = rollout(self.env, policy,
max_path_length=200,
reset_every=self.rollouts_per_meta_task,
stochastic=stochastic,
batch_size=1, record_teacher=True, teacher_dict=teacher_dict,
video_directory=self.exp_name, video_name=save_name + str(self.curriculum_step),
num_rollouts=num_rollouts, save_wandb=save_wandb, save_locally=False,
obs_preprocessor=self.obs_preprocessor,)
if log_prefix is not None:
logger.logkv(log_prefix + "Acc", accuracy)
logger.logkv(log_prefix + "Reward", np.mean([sum(path['rewards']) for path in paths]))
logger.logkv(log_prefix + "PathLength",
np.mean([path['env_infos'][-1]['episode_length'] for path in paths]))
logger.logkv(log_prefix + "Success", np.mean([path['env_infos'][-1]['success'] for path in paths]))
return paths, accuracy
# with tf.Session():
# [rest of the code]
with tf.Session() as sess:
pkl_path = args.param
print("Testing policy %s" % pkl_path)
data = joblib.load(pkl_path)
policy = data['policy']
env = normalize(ArmReacherEnv(side='right'))
real_rewards = np.array([])
act_rewards = np.array([])
pos_rewards = np.array([])
for _ in range(args.num_rollouts):
path = rollout(env,
policy,
max_path_length=args.max_path_length,
animated=False,
speedup=args.speedup,
video_filename=args.video_filename,
save_video=False,
ignore_done=args.ignore_done,
stochastic=args.stochastic)
real_rewards = np.append(real_rewards, np.sum(path[0]['rewards']))
print("Real Reward Sum", np.sum(path[0]['rewards']))
#print(np.mean(path[0]['rewards']))
#print(len(path_act[0]['rewards']))
#print(len(path_pos[0]['rewards']))
print("Real Reward Avg")
print(np.mean(real_rewards))
experimet_paths = load_exps_data(args.path,
gap=args.gap_pkl,
max=args.max_pkl)
for exp_path in experimet_paths:
max_path_length = exp_path['json'][
'max_path_length'] if args.max_path_length is None else args.max_path_length
if valid_experiment(exp_path['json']):
for pkl_path in exp_path['pkl']:
with tf.Session() as sess:
print("\n Testing policy %s \n" % pkl_path)
data = joblib.load(pkl_path)
policy = data['policy']
if hasattr(policy, 'switch_to_pre_update'):
policy.switch_to_pre_update()
env = data['env']
video_filename = pkl_path.split('.')[0] + '.mp4'
paths = rollout(env,
policy,
max_path_length=max_path_length,
animated=False,
speedup=args.speedup,
video_filename=video_filename,
save_video=True,
ignore_done=args.ignore_done,
stochastic=args.stochastic,
num_rollouts=args.num_rollouts)
print('Average Returns: ',
np.mean([sum(path['rewards']) for path in paths]))
tf.reset_default_graph()
env_class = getattr(iclr19_levels, args.class_name)
# env_args = {
# 'start_loc': args.start_loc,
# 'include_holdout_obj': args.holdout_obj,
# }
# if args.grid_size is not None:
# env_args['room_size'] = args.grid_size
# if args.num_dists is not None:
# env_args['num_dists'] = args.num_dists
# e_new = env_class(**env_args)
e_new = env_class(**arguments)
e_new.use_teacher = args.use_teacher
if args.use_teacher:
teacher = PostActionAdvice(Bot, e_new)
e_new.teacher = teacher
e_new.teacher.set_feedback_type(args.feedback_type)
env = rl2env(normalize(e_new))
video_filename = os.path.join(args.path, 'saved_video.mp4')
paths, accuracy = rollout(env, policy, max_path_length=max_path_length, animated=args.animated, speedup=args.speedup,
video_filename=video_filename, save_video=True, ignore_done=args.ignore_done, batch_size=1,
stochastic=args.stochastic, num_rollouts=args.num_rollouts, reset_every=args.reset_every,
record_teacher=True, reward_predictor=reward_predictor, dense_rewards=args.dense_rewards)
print('Average Returns: ', np.mean([sum(path['rewards']) for path in paths]))
print('Average Path Length: ', np.mean([path['env_infos'][-1]['episode_length'] for path in paths]))
print('Average Success Rate: ', np.mean([path['env_infos'][-1]['success'] for path in paths]))
tf.reset_default_graph()
print("Testing policy %s" % pkl_path)
data = joblib.load(pkl_path)
policy = data['policy']
env = normalize(ArmReacherEnv(side='right'))
goal = data['env'].goal
real_rewards = np.array([])
act_rewards = np.array([])
pos_rewards = np.array([])
for i in range(args.num_rollouts):
path = rollout(env,
policy,
max_path_length=args.max_path_length,
animated=False,
speedup=args.speedup,
video_filename=args.video_filename,
save_video=False,
ignore_done=args.ignore_done,
stochastic=args.stochastic)
mujoco_env_mimic_act = normalize(BlueEnv(actions=env.actions))
pickle.dump(env.actions, open("actions_ppo_0.pkl", "wb"))
mujoco_env_mimic_act.goal = env.goal
act_filename = "local_act_maml_" + str(i) + ".mp4"
path_act = rollout(mujoco_env_mimic_act,
policy,
max_path_length=args.max_path_length,
animated=True,