def test_render():
"""
Launch interactive environment where you can move the pusher around with
the arrow keys
:return:
:rtype:
"""
env = PusherSlider()
env.config = PusherSlider.make_default_config()
env.reset()
action = np.zeros(2)
env.step(action) # this is necessary, doesn't work otherwise . . .
env.render(mode='human')
print(env._slider_body.position)
time.sleep(3.0)
print("moving the block")
env._slider_body.position = [400, 300]
env._space.step(1e-4)
# env.step(action)
env.render(mode='human')
time.sleep(3.0)
def run_interactive_circle_slider():
"""
Launch interactive environment where you can move the pusher around with
the arrow keys
:return:
:rtype:
"""
config_file = os.path.join(get_project_root(),
'experiments/03/config.yaml')
config = load_yaml(config_file)
env = PusherSlider(config=config)
env.reset()
while env._running:
action = env.process_events()
env.step(action)
obs, reward, done, info = env.step(action)
env.render(mode='human')
if True:
print("\n\n\n")
print("slider position", obs['slider']['position'])
print("pusher position", obs['pusher']['position'])
def test_gym_API():
"""
Collects two episodes and saves them to disk
Currently no images are being saved
:return:
:rtype:
"""
config_file = os.path.join(get_project_root(),
'experiments/01/config.yaml')
config = load_yaml(config_file)
env = PusherSlider(config=config)
env.setup_environment()
episode_container = None # for namespacing purposes
multi_episode_container = MultiEpisodeContainer()
print("fps", env._fps)
num_episodes = 1
# num_timesteps_per_episode = 40
for episode_idx in range(num_episodes):
env.reset()
obs_prev = env.get_observation()
episode_container = EpisodeContainer()
episode_container.set_config(env.config)
for i in range(env.config['dataset']['num_timesteps']):
env.render(mode='human')
if i == 50:
# save this image out in local directory
img = env.render(mode='rgb_array')
if DEBUG_PRINTS:
print("saving figure to file")
print("type(img)", type(img))
print(img.shape)
print(img.dtype)
pil_img = numpy_to_PIL(img)
pil_img.save('test_PIL.png')
action = np.array([100, 0])
obs, reward, done, info = env.step(action)
episode_container.add_obs_action(obs_prev, action)
obs_prev = obs
if DEBUG_PRINTS:
print("\n\nsim_time = ", env._sim_time)
print("obs sim_time = ", obs['sim_time'])
multi_episode_container.add_episode(episode_container)
if True:
save_file = os.path.join(os.getcwd(), "single_episode_log.p")
episode_container.save_to_file(save_file)
multi_episode_save_file = os.path.join(os.getcwd(),
"multi_episode_log.p")
multi_episode_container.save_to_file(multi_episode_save_file)
def mpc_episode_keypoint_observation(
config, # the global config
mpc_idx, # int: index of this trial for logging purposes
model_dy, # the dynamics model
mpc_dir, # str: directory to store results
planner,
obs_goal,
action_function,
observation_function,
video=True,
image=True,
use_gpu=True):
'''
setup
'''
# optionally add noise to observation as specified in config
add_noise_to_observation = False
if "add_noise_to_observation" in config["mpc"]:
# raise ValueError("testing")
add_noise_to_observation = config["mpc"]["add_noise_to_observation"]
# basic info
n_his = config['train']['n_history']
num_steps = config['mpc']['num_timesteps'] + n_his
# create PusherSlider env
env = PusherSlider(config)
env.setup_environment()
env.reset()
env.step_space_to_initialize_render()
# set up episode container
action_zero = np.zeros(2)
# states used to roll out the MPC model
states_roll_mpc = np.zeros((num_steps, config['dataset']['state_dim']))
# gt states that actually occurred
states_roll_gt = np.zeros(states_roll_mpc.shape)
actions_roll = np.zeros((num_steps, config['dataset']['action_dim']))
gt_obs_list = []
img_gt_PIL_list = []
actions_roll = np.zeros((num_steps, config['dataset']['action_dim']))
for i in range(n_his, num_steps):
actions_roll[i] = sample_pusher_velocity()
action_lower_lim = np.array(config['mpc']['action_lower_lim'])
action_upper_lim = np.array(config['mpc']['action_upper_lim'])
'''
model predictive control
'''
for step in range(num_steps):
print('roll_step %d/%d' % (step, num_steps))
states_roll_mpc[step] = observation_function(
env.get_observation(), data_augmentation=add_noise_to_observation)
img_gt_PIL_list.append(numpy_to_PIL(env.render(mode='rgb_array')))
# stand still for n_his timesteps
if step < n_his:
action = action_zero
else:
actions_roll[step:] = planner.trajectory_optimization(
states_roll_mpc[step - n_his + 1:step + 1],
obs_goal,
model_dy,
actions_roll[step - n_his + 1:],
n_sample=config['mpc']['n_sample'],
n_look_ahead=min(num_steps - step,
config['mpc']['n_look_ahead']),
n_update_iter=config['mpc']['n_update_iter_init']
if step == n_his else config['mpc']['n_update_iter'],
action_lower_lim=action_lower_lim,
action_upper_lim=action_upper_lim,
use_gpu=use_gpu)
action = actions_roll[step]
obs, reward, done, info = env.step(action)
gt_obs_list.append(obs)
'''
render the prediction
'''
# reset PusherSlider env
env.reset()
# setup rendering data
width = config['env']['display_size'][0]
height = config['env']['display_size'][1]
n_split = 3
split = 4
if image:
eval_img_dir = os.path.join(mpc_dir, '%d_img' % mpc_idx)
os.system('mkdir -p ' + eval_img_dir)
print('Save images to %s' % eval_img_dir)
if video:
eval_vid_path = os.path.join(mpc_dir, '%d_vid.avi' % mpc_idx)
fourcc = cv2.VideoWriter_fourcc('M', 'J', 'P', 'G')
print('Save video as %s' % eval_vid_path)
fps = int(env.fps)
print("fps:", fps)
out = cv2.VideoWriter(eval_vid_path, fourcc, fps,
(width * n_split + split *
(n_split - 1), height))
# gt represents goal
for i, start_idx in enumerate(range(num_steps)):
# get ground truth positions at goal
goal_gt_state_dict = pusher_pose_slider_keypoints_from_tensor(obs_goal)
# now render them
img_goal_gt_PIL = pusher_slider_keypoints_image(
config, goal_gt_state_dict['pusher']['position'],
goal_gt_state_dict['keypoint_positions'])
img_goal_gt = np.array(img_goal_gt_PIL)
#
img_gt_PIL = img_gt_PIL_list[i]
# # actual GT positions that happened
state_roll = states_roll_mpc[i, :]
state_roll_dict = pusher_pose_slider_keypoints_from_tensor(state_roll)
img_gt_PIL = pusher_slider_keypoints_image(
config,
state_roll_dict['pusher']['position'],
state_roll_dict['keypoint_positions'],
img=img_gt_PIL)
img_gt = np.array(img_gt_PIL)
# blend the images
img_merge_PIL = Image.blend(img_goal_gt_PIL, img_gt_PIL, 0.5)
# get numpy image
img_merge = np.array(img_merge_PIL)
img_output = np.zeros(
(img_goal_gt.shape[0],
img_goal_gt.shape[1] * n_split + split * (n_split - 1),
3)).astype(np.uint8)
img_output[:, :img_goal_gt.
shape[1]] = img_goal_gt # goal keypoints image
img_output[:, img_goal_gt.shape[1] + split:img_goal_gt.shape[1] * 2 +
split] = img_gt
img_output[:, (img_goal_gt.shape[1] + split) * 2:] = img_merge
if image:
# convert to PIL then save
img_output_PIL = numpy_to_PIL(img_output)
img_output_PIL.save(os.path.join(eval_img_dir, 'fig_%d.png' % i))
if video:
# convert from RGB to BGR since video writer wants BGR
img_output_bgr = cv2.cvtColor(img_output, cv2.COLOR_RGB2BGR)
out.write(img_output_bgr)
if video:
out.release()
def eval_episode(
config, # the global config
dataset, # MultiEpisodeDataset
episode_name, # str,
roll_idx, # int: index of this rollout for logging purposes
model_dy, # dynamics model
eval_dir, # str: directory to store results
video=True, # save a video of rollout
image=True, # save images of rollout
use_gpu=True, # bool: whether or not to use the gpu
start_idx=None, # int: starting index for rollout
n_prediction=5, # prediction horizon,
render_human=False, # whether to render in human visible mode
):
episode = dataset._episodes[episode_name]
n_his = config["train"]["n_history"]
n_roll = config["train"]["n_rollout"]
if start_idx is None:
low, high = dataset.get_valid_idx_range_for_episode(episode)
start_idx = int((low + high) / 2.0)
data = dataset.__getitem__(0,
episode_name=episode_name,
episode_idx=start_idx,
rollout_length=0)
# need to do the unsqueeze to make it look like they
# are in batch form for doing forward pass through the
# network
states_unnorm = data['observations'].unsqueeze(0)
actions_unnorm = data['actions'].unsqueeze(0)
if use_gpu:
device = 'cuda'
states_unnorm = states_unnorm.cuda()
actions_unnorm = actions_unnorm.cuda()
else:
device = 'cpu'
# normalize states for the entire rollout
states = model_dy.state_normalizer.normalize(states_unnorm)
# state_cur: B x n_his x state_dim
state_cur = states[:, :n_his]
# states.shape[2] should be state_dim
states_predicted = torch.zeros((n_prediction, states.shape[2]),
device=device)
states_predicted_idx = [None] * n_prediction
print("states_predicted.shape", states_predicted.shape)
# this code is exactly copied from the training setup
# ropagate dynamics forward through time using the learned model
with torch.no_grad():
for j in range(n_prediction):
# action_cur: B x n_his x action_dim
# get actions from the dataloader
action_cur = None
data = dataset.__getitem__(0,
episode_name=episode_name,
episode_idx=(start_idx + j),
rollout_length=0)
actions_unnorm = data['actions'].unsqueeze(0).cuda()
actions = model_dy.action_normalizer.normalize(actions_unnorm)
action_cur = actions[:, :n_his]
# state_pred: B x state_dim
# state_cur: B x n_his x state_dim
state_pred = model_dy(state_cur, action_cur)
states_predicted[j] = state_pred.squeeze()
states_predicted_idx[j] = start_idx + j + 1
# update state_cur
state_cur = torch.cat(
[state_cur[:, 1:], state_pred.unsqueeze(1)], 1)
states_predicted_unnorm = model_dy.state_normalizer.denormalize(
states_predicted)
# create PusherSlider env
env = PusherSlider(config)
env.setup_environment()
env.reset()
# setup rendering data
'''
render the prediction
'''
width = config['env']['display_size'][0]
height = config['env']['display_size'][1]
n_split = 3
split = 4
if image:
eval_img_dir = os.path.join(eval_dir, '%d_img' % roll_idx)
os.system('mkdir -p ' + eval_img_dir)
print('Save images to %s' % eval_img_dir)
if video:
eval_vid_path = os.path.join(eval_dir, '%d_vid.avi' % roll_idx)
fourcc = cv2.VideoWriter_fourcc('M', 'J', 'P', 'G')
print('Save video as %s' % eval_vid_path)
fps = int(env.fps)
print("fps:", fps)
out = cv2.VideoWriter(eval_vid_path, fourcc, fps,
(width * n_split + split *
(n_split - 1), height))
# can easily read the ground truth states directly from the episode . . .
for i, start_idx in enumerate(states_predicted_idx):
# get predicted positions
pred_state_np = states_predicted_unnorm[i, :].cpu().numpy()
predicted_state_dict = pusher_slider_pose_from_tensor(pred_state_np)
# get ground truth positions
gt_state_dict = episode.get_observation(start_idx)
# now render them
env.set_object_positions_for_rendering(gt_state_dict)
img_gt = env.render(mode='rgb_array')
img_gt_PIL = numpy_to_PIL(img_gt)
if render_human:
print("rendering human")
env.render(mode='human')
# predicted image
env.set_object_positions_for_rendering(predicted_state_dict)
img_pred = env.render(mode='rgb_array')
img_pred_PIL = numpy_to_PIL(img_pred)
if render_human:
env.render(mode='human')
# blend the images
img_merge_PIL = Image.blend(img_gt_PIL, img_pred_PIL, 0.5)
# # show the predictions
# if SHOW_BLENDED_IMAGE:
# print("showing PIL image")
# img_merge_PIL.show()
# time.sleep(1.0)
# get numpy image
img_merge = np.array(img_merge_PIL)
img_output = np.zeros((img_gt.shape[0], img_gt.shape[1] * n_split +
split * (n_split - 1), 3)).astype(np.uint8)
img_output[:, :img_gt.shape[1]] = img_gt
img_output[:, img_gt.shape[1] + split:img_gt.shape[1] * 2 +
split] = img_pred
img_output[:, (img_gt.shape[1] + split) * 2:] = img_merge
if image:
# convert to PIL then save
img_output_PIL = numpy_to_PIL(img_output)
img_output_PIL.save(os.path.join(eval_img_dir, 'fig_%d.png' % i))
if video:
# convert from RGB to BGR since video writer wants BGR
img_output_bgr = cv2.cvtColor(img_output, cv2.COLOR_RGB2BGR)
out.write(img_output_bgr)
if video:
out.release()