print("Loaded weight from {}\n".format(weight_path))
start = time.time()
env.reset()
reward_ll_sum = 0
done_sum = 0
average_dones = 0.
n_steps = math.floor(cfg['environment']['max_time'] /
cfg['environment']['control_dt'])
total_steps = n_steps * 1
start_step_id = 0
print("Visualizing and evaluating the policy", weight_path + ".pt")
loaded_graph = torch.jit.load(weight_path + '.pt')
env.load_scaling(
weight_path.rsplit('/', 1)[0],
int(weight_path.rsplit('/', 1)[1].split('_', 1)[1]))
print(
"Load observation scaling in",
weight_path.rsplit('/', 1)[0] + ":", "mean" +
str(int(weight_path.rsplit('/', 1)[1].split('_', 1)[1])) + ".csv",
"and", "var" +
str(int(weight_path.rsplit('/', 1)[1].split('_', 1)[1])) + ".csv")
env.turn_on_visualization()
# max_steps = 1000000
max_steps = 1000 ## 10 secs
for step in range(max_steps):
time.sleep(0.01)
obs = env.observe(False)
action_ll = loaded_graph(torch.from_numpy(obs).cpu())
'{:6.4f}'.format(end - start)))
print('{:<40} {:>6}'.format(
"fps: ", '{:6.0f}'.format(total_steps / (end - start))))
print('std: ')
print(np.exp(actor.distribution.std.cpu().detach().numpy()))
print('----------------------------------------------------\n')
if test_mode:
curriculum_setting = True
save_dir = os.environ['WORKSPACE'] + "/ME491TermProject/data/~~~~~~~~"
test_policy = 1000
env.turn_on_visualization()
env.start_video_recording(
datetime.datetime.now().strftime("%Y-%m-%d-%H-%M-%S") + "policy_" +
str(test_policy) + '.mp4')
env.load_scaling(save_dir, test_policy)
if curriculum_setting:
if test_policy > curriculum_start:
for i in range(test_policy - curriculum_start):
env.curriculum_callback()
loaded_graph = torch.jit.load(save_dir + "/policy_" + str(test_policy) +
'.pt')
print("load_graph")
dones = False
steps = 0
env.reset()
time.sleep(1)
while (not dones) and steps < 2 * n_steps:
obs = env.observe(False)
action_ll = loaded_graph(torch.from_numpy(obs).cpu())
env.reset()
reward_ll_sum = 0
done_sum = 0
average_dones = 0.
n_steps = math.floor(cfg['environment']['max_time'] /
cfg['environment']['control_dt'])
total_steps = n_steps * 1
start_step_id = 0
print("Visualizing and evaluating the policy: ", weight_path)
loaded_graph = ppo_module.MLP(cfg['architecture']['policy_net'],
torch.nn.LeakyReLU, ob_dim, act_dim)
loaded_graph.load_state_dict(
torch.load(weight_path)['actor_architecture_state_dict'])
env.load_scaling(weight_dir, int(iteration_number))
env.turn_on_visualization()
# max_steps = 1000000
max_steps = 1000 ## 10 secs
for step in range(max_steps):
time.sleep(0.01)
obs = env.observe(False)
action_ll = loaded_graph.architecture(torch.from_numpy(obs).cpu())
reward_ll, dones = env.step(action_ll.cpu().detach().numpy())
reward_ll_sum = reward_ll_sum + reward_ll[0]
if dones or step == max_steps - 1:
print('----------------------------------------------------')
print('{:<40} {:>6}'.format(
"average ll reward: ",