state, reward, done, _ = _env.step(action.cpu().numpy()[0])
total_reward += reward
state = agent.Tensor([state])
if done:
break
return total_reward
test_episodes = 100
for env_name in [args.env]:#os.listdir(base_dir):
env = NormalizedActions(gym.make(env_name))
agent = DDPG(beta=0.9, epsilon=0, learning_rate=1e-4, gamma=0.99, tau=0.01, hidden_size_dim0=args.hidden_size, hidden_size_dim1=args.hidden_size, num_inputs=env.observation_space.shape[0], action_space=env.action_space, train_mode=False, alpha=0, replay_size=0, optimizer = 0, two_player=args.two_player, normalize_obs=True)
noise = uniform.Uniform(agent.Tensor([-1.0]), agent.Tensor([1.0]))
basic_bm = copy.deepcopy(env.env.env.model.body_mass.copy())
env_dir = base_dir + env_name + '/'
for optimizer in [args.optimizer]: #['RMSprop', 'SGLD_thermal_0.01', 'SGLD_thermal_0.001', 'SGLD_thermal_0.0001', 'SGLD_thermal_1e-05']:
for noise_type in [args.action_noise]:
noise_dir = env_dir + optimizer + '/' + noise_type + '/nr_mdp_' + str(args.alpha) + '_1/'
if os.path.exists(noise_dir):
for subdir in sorted(os.listdir(noise_dir)):
results = {}
run_number = 0
dir = noise_dir + subdir #+ '/' + str(run_number)
print(dir)
if os.path.exists(noise_dir + subdir)\
def reset_noise(a, a_noise):
if a_noise is not None:
a_noise.reset()
total_steps = 0
print(base_dir)
if args.num_steps is not None:
assert args.num_epochs is None
nb_epochs = int(args.num_steps) // (args.num_epochs_cycles *
args.num_rollout_steps)
else:
nb_epochs = 500
state = agent.Tensor([env.reset()])
eval_state = agent.Tensor([eval_env.reset()])
eval_reward = 0
episode_reward = 0
agent.train()
reset_noise(agent, normalnoise)
if args.visualize:
vis = visdom.Visdom(env=base_dir)
else:
vis = None
train_steps = 0
ratio = args.ratio + 1
