def generate_vae_dataset(variant):
env_class = variant.get('env_class', None)
env_kwargs = variant.get('env_kwargs', None)
env_id = variant.get('env_id', None)
N = variant.get('N', 10000)
test_p = variant.get('test_p', 0.9)
use_cached = variant.get('use_cached', True)
imsize = variant.get('imsize', 84)
num_channels = variant.get('num_channels', 3)
show = variant.get('show', False)
init_camera = variant.get('init_camera', None)
dataset_path = variant.get('dataset_path', None)
oracle_dataset_using_set_to_goal = variant.get(
'oracle_dataset_using_set_to_goal', False)
random_rollout_data = variant.get('random_rollout_data', False)
random_and_oracle_policy_data = variant.get(
'random_and_oracle_policy_data', False)
random_and_oracle_policy_data_split = variant.get(
'random_and_oracle_policy_data_split', 0)
policy_file = variant.get('policy_file', None)
n_random_steps = variant.get('n_random_steps', 100)
vae_dataset_specific_env_kwargs = variant.get(
'vae_dataset_specific_env_kwargs', None)
save_file_prefix = variant.get('save_file_prefix', None)
non_presampled_goal_img_is_garbage = variant.get(
'non_presampled_goal_img_is_garbage', None)
tag = variant.get('tag', '')
from multiworld.core.image_env import ImageEnv, unormalize_image
import rlkit.torch.pytorch_util as ptu
info = {}
if dataset_path is not None:
dataset = load_local_or_remote_file(dataset_path)
N = dataset.shape[0]
else:
if env_kwargs is None:
env_kwargs = {}
if save_file_prefix is None:
save_file_prefix = env_id
if save_file_prefix is None:
save_file_prefix = env_class.__name__
filename = "/tmp/{}_N{}_{}_imsize{}_random_oracle_split_{}{}.npy".format(
save_file_prefix,
str(N),
init_camera.__name__ if init_camera else '',
imsize,
random_and_oracle_policy_data_split,
tag,
)
if use_cached and osp.isfile(filename):
dataset = np.load(filename)
print("loaded data from saved file", filename)
else:
now = time.time()
if env_id is not None:
import gym
import multiworld
multiworld.register_all_envs()
env = gym.make(env_id)
else:
if vae_dataset_specific_env_kwargs is None:
vae_dataset_specific_env_kwargs = {}
for key, val in env_kwargs.items():
if key not in vae_dataset_specific_env_kwargs:
vae_dataset_specific_env_kwargs[key] = val
env = env_class(**vae_dataset_specific_env_kwargs)
if not isinstance(env, ImageEnv):
env = ImageEnv(
env,
imsize,
init_camera=init_camera,
transpose=True,
normalize=True,
non_presampled_goal_img_is_garbage=
non_presampled_goal_img_is_garbage,
)
else:
imsize = env.imsize
env.non_presampled_goal_img_is_garbage = non_presampled_goal_img_is_garbage
env.reset()
info['env'] = env
if random_and_oracle_policy_data:
policy_file = load_local_or_remote_file(policy_file)
policy = policy_file['policy']
policy.to(ptu.device)
if random_rollout_data:
from rlkit.exploration_strategies.ou_strategy import OUStrategy
policy = OUStrategy(env.action_space)
dataset = np.zeros((N, imsize * imsize * num_channels),
dtype=np.uint8)
for i in range(N):
if random_and_oracle_policy_data:
num_random_steps = int(N *
random_and_oracle_policy_data_split)
if i < num_random_steps:
env.reset()
for _ in range(n_random_steps):
obs = env.step(env.action_space.sample())[0]
else:
obs = env.reset()
policy.reset()
for _ in range(n_random_steps):
policy_obs = np.hstack((
obs['state_observation'],
obs['state_desired_goal'],
))
action, _ = policy.get_action(policy_obs)
obs, _, _, _ = env.step(action)
elif oracle_dataset_using_set_to_goal:
print(i)
goal = env.sample_goal()
env.set_to_goal(goal)
obs = env._get_obs()
elif random_rollout_data:
if i % n_random_steps == 0:
g = dict(
state_desired_goal=env.sample_goal_for_rollout())
env.set_to_goal(g)
policy.reset()
# env.reset()
u = policy.get_action_from_raw_action(
env.action_space.sample())
obs = env.step(u)[0]
else:
env.reset()
for _ in range(n_random_steps):
obs = env.step(env.action_space.sample())[0]
img = obs['image_observation']
dataset[i, :] = unormalize_image(img)
if show:
img = img.reshape(3, imsize, imsize).transpose()
img = img[::-1, :, ::-1]
cv2.imshow('img', img)
cv2.waitKey(1)
# radius = input('waiting...')
print("done making training data", filename, time.time() - now)
np.save(filename, dataset)
n = int(N * test_p)
train_dataset = dataset[:n, :]
test_dataset = dataset[n:, :]
return train_dataset, test_dataset, info
self.walls = [
HorizontalWall(
self.ball_radius,
self.inner_wall_max_dist,
-self.inner_wall_max_dist,
self.inner_wall_max_dist,
)
]
if wall_shape == "--":
self.walls = [
HorizontalWall(
self.ball_radius,
0,
-self.inner_wall_max_dist,
self.inner_wall_max_dist,
)
]
if __name__ == "__main__":
# e = Point2DEnv()
import matplotlib.pyplot as plt
# e = Point2DWallEnv("-", render_size=84)
e = ImageEnv(Point2DWallEnv(wall_shape="u", render_size=84))
for i in range(10):
e.reset()
for j in range(50):
e.step(np.random.rand(2))
e.render()
im = e.get_image()
right_x=6,
thickness=0.5
)
]
if __name__ == "__main__":
# e = Point2DEnv()
import matplotlib.pyplot as plt
import time
# e = Point2DWallEnv("-", render_size=84)
# e = ImageEnv(Point2DWallEnv(wall_shape="--", render_size=84))
e = ImageEnv(Point2DWallEnv(wall_shape='maze',
render_size=256,
ball_radius=0.25,
boundary_dist=10,
action_limit=0.25,
fixed_goal=False,
randomize_position_on_reset=False,
target_radius=0,
initial_position=(-8, -8)),
imsize=256)
for i in range(10):
e.reset()
for j in range(50):
# e.step(np.random.rand(2) - 0.5)
e.step(np.array([1, 0]))
e.render()
im = e.get_image()
time.sleep(0.1)
def generate_vae_dataset(
N=10000,
test_p=0.9,
use_cached=True,
imsize=84,
show=False,
dataset_path=None,
action_space_sampling=False,
init_camera=None,
env_class=None,
env_kwargs=None,
):
filename = "/tmp/sawyer_xyz_pos_control_new_zoom_cam" + str(N) + '.npy'
info = {}
if dataset_path is not None:
filename = local_path_from_s3_or_local_path(dataset_path)
dataset = np.load(filename)
elif use_cached and osp.isfile(filename):
dataset = np.load(filename)
print("loaded data from saved file", filename)
else:
now = time.time()
if env_kwargs == None:
env_kwargs = dict()
env = env_class(**env_kwargs)
env = ImageEnv(
env,
imsize,
transpose=True,
init_camera=init_camera,
normalize=True,
)
dataset = np.zeros((N, imsize * imsize * 3), dtype=np.uint8)
if action_space_sampling:
action_space = Box(np.array([-.1, .5, 0]), np.array([.1, .7, .5]))
for i in range(N):
env.set_to_goal(env.sample_goal())
img = env._get_flat_img()
dataset[i, :] = unormalize_image(img)
if show:
cv2.imshow('img', img.reshape(3, 84, 84).transpose())
cv2.waitKey(1)
print(i)
info['env'] = env
else:
policy = RandomPolicy(env.action_space)
es = OUStrategy(action_space=env.action_space, theta=0)
exploration_policy = PolicyWrappedWithExplorationStrategy(
exploration_strategy=es,
policy=policy,
)
for i in range(N):
# Move the goal out of the image
env.wrapped_env.set_goal(np.array([100, 100, 100]))
if i % 50 == 0:
print('Reset')
env.reset()
exploration_policy.reset()
for _ in range(1):
action = exploration_policy.get_action()[0] * 10
env.wrapped_env.step(action)
img = env.step(env.action_space.sample())[0]
dataset[i, :] = img
if show:
cv2.imshow('img', img.reshape(3, 84, 84).transpose())
cv2.waitKey(1)
print(i)
print("done making training data", time.time() - now)
np.save(filename, dataset)
n = int(N * test_p)
train_dataset = dataset[:n, :]
test_dataset = dataset[n:, :]
return train_dataset, test_dataset, info
# env_name = 'SawyerDoorHookResetFreeEnv-v1'
env_name = 'SawyerPushHurdle-v0'
# env_name = 'SawyerPushNIPSFull-v0'
# env_name = 'SawyerPushNIPSEasy-v0'
# env_name = 'SawyerPushHurdleResetFreeEnv-v0'
multiworld.register_all_envs()
imsize = 48
# presampled_goals_path = 'data/local/goals/SawyerDoorHookResetFreeEnv-v1-goal.npy'
# presampled_goals_path = 'data/local/goals/SawyerPickupEnvYZEasy-v0-goal-500.npy'
# presampled_goals = np.load(presampled_goals_path, allow_pickle=True).item()
env = ImageEnv(env,
imsize,
init_camera=sawyer_init_camera_zoomed_in,
transpose=True,
normalize=True,
presampled_goals=None)
print(env.action_space.low)
print(env.action_space.high)
for i in range(50):
o = env.reset()
for t in range(50):
print(t)
action = env.action_space.sample()
s, r, _, _ = env.step(action)
img = s['image_observation']
show_obs(img, imsize=imsize, name=env_name)
def generate_vae_dataset(variant):
print(variant)
from tqdm import tqdm
env_class = variant.get('env_class', None)
env_kwargs = variant.get('env_kwargs', None)
env_id = variant.get('env_id', None)
N = variant.get('N', 10000)
batch_size = variant.get('batch_size', 128)
test_p = variant.get('test_p', 0.9)
use_cached = variant.get('use_cached', True)
imsize = variant.get('imsize', 84)
num_channels = variant.get('num_channels', 3)
show = variant.get('show', False)
init_camera = variant.get('init_camera', None)
dataset_path = variant.get('dataset_path', None)
augment_data = variant.get('augment_data', False)
data_filter_fn = variant.get('data_filter_fn', lambda x: x)
delete_after_loading = variant.get('delete_after_loading', False)
oracle_dataset_using_set_to_goal = variant.get(
'oracle_dataset_using_set_to_goal', False)
random_rollout_data = variant.get('random_rollout_data', False)
random_rollout_data_set_to_goal = variant.get(
'random_rollout_data_set_to_goal', True)
random_and_oracle_policy_data = variant.get(
'random_and_oracle_policy_data', False)
random_and_oracle_policy_data_split = variant.get(
'random_and_oracle_policy_data_split', 0)
policy_file = variant.get('policy_file', None)
n_random_steps = variant.get('n_random_steps', 100)
vae_dataset_specific_env_kwargs = variant.get(
'vae_dataset_specific_env_kwargs', None)
save_file_prefix = variant.get('save_file_prefix', None)
non_presampled_goal_img_is_garbage = variant.get(
'non_presampled_goal_img_is_garbage', None)
conditional_vae_dataset = variant.get('conditional_vae_dataset', False)
use_env_labels = variant.get('use_env_labels', False)
use_linear_dynamics = variant.get('use_linear_dynamics', False)
enviorment_dataset = variant.get('enviorment_dataset', False)
save_trajectories = variant.get('save_trajectories', False)
save_trajectories = save_trajectories or use_linear_dynamics or conditional_vae_dataset
tag = variant.get('tag', '')
assert N % n_random_steps == 0, "Fix N/horizon or dataset generation will fail"
from multiworld.core.image_env import ImageEnv, unormalize_image
import rlkit.torch.pytorch_util as ptu
from rlkit.util.io import load_local_or_remote_file
from rlkit.data_management.dataset import (
TrajectoryDataset, ImageObservationDataset, InitialObservationDataset,
EnvironmentDataset, ConditionalDynamicsDataset,
InitialObservationNumpyDataset, InfiniteBatchLoader,
InitialObservationNumpyJitteringDataset)
info = {}
use_test_dataset = False
if dataset_path is not None:
if type(dataset_path) == str:
dataset = load_local_or_remote_file(
dataset_path, delete_after_loading=delete_after_loading)
dataset = dataset.item()
N = dataset['observations'].shape[0] * dataset[
'observations'].shape[1]
n_random_steps = dataset['observations'].shape[1]
if isinstance(dataset_path, list):
dataset = concatenate_datasets(dataset_path)
N = dataset['observations'].shape[0] * dataset[
'observations'].shape[1]
n_random_steps = dataset['observations'].shape[1]
if isinstance(dataset_path, dict):
if type(dataset_path['train']) == str:
dataset = load_local_or_remote_file(
dataset_path['train'],
delete_after_loading=delete_after_loading)
dataset = dataset.item()
elif isinstance(dataset_path['train'], list):
dataset = concatenate_datasets(dataset_path['train'])
if type(dataset_path['test']) == str:
test_dataset = load_local_or_remote_file(
dataset_path['test'],
delete_after_loading=delete_after_loading)
test_dataset = test_dataset.item()
elif isinstance(dataset_path['test'], list):
test_dataset = concatenate_datasets(dataset_path['test'])
N = dataset['observations'].shape[0] * dataset[
'observations'].shape[1]
n_random_steps = dataset['observations'].shape[1]
use_test_dataset = True
else:
if env_kwargs is None:
env_kwargs = {}
if save_file_prefix is None:
save_file_prefix = env_id
if save_file_prefix is None:
save_file_prefix = env_class.__name__
filename = "/tmp/{}_N{}_{}_imsize{}_random_oracle_split_{}{}.npy".format(
save_file_prefix,
str(N),
init_camera.__name__
if init_camera and hasattr(init_camera, '__name__') else '',
imsize,
random_and_oracle_policy_data_split,
tag,
)
if use_cached and osp.isfile(filename):
dataset = load_local_or_remote_file(
filename, delete_after_loading=delete_after_loading)
if conditional_vae_dataset:
dataset = dataset.item()
print("loaded data from saved file", filename)
else:
now = time.time()
if env_id is not None:
import gym
import multiworld
multiworld.register_all_envs()
env = gym.make(env_id)
else:
if vae_dataset_specific_env_kwargs is None:
vae_dataset_specific_env_kwargs = {}
for key, val in env_kwargs.items():
if key not in vae_dataset_specific_env_kwargs:
vae_dataset_specific_env_kwargs[key] = val
env = env_class(**vae_dataset_specific_env_kwargs)
if not isinstance(env, ImageEnv):
env = ImageEnv(
env,
imsize,
init_camera=init_camera,
transpose=True,
normalize=True,
non_presampled_goal_img_is_garbage=
non_presampled_goal_img_is_garbage,
)
else:
imsize = env.imsize
env.non_presampled_goal_img_is_garbage = non_presampled_goal_img_is_garbage
env.reset()
info['env'] = env
if random_and_oracle_policy_data:
policy_file = load_local_or_remote_file(policy_file)
policy = policy_file['policy']
policy.to(ptu.device)
if random_rollout_data:
from rlkit.exploration_strategies.ou_strategy import OUStrategy
policy = OUStrategy(env.action_space)
if save_trajectories:
dataset = {
'observations':
np.zeros((N // n_random_steps, n_random_steps,
imsize * imsize * num_channels),
dtype=np.uint8),
'actions':
np.zeros((N // n_random_steps, n_random_steps,
env.action_space.shape[0]),
dtype=np.float),
'env':
np.zeros(
(N // n_random_steps, imsize * imsize * num_channels),
dtype=np.uint8),
}
else:
dataset = np.zeros((N, imsize * imsize * num_channels),
dtype=np.uint8)
labels = []
for i in tqdm(range(N)):
if random_and_oracle_policy_data:
num_random_steps = int(N *
random_and_oracle_policy_data_split)
if i < num_random_steps:
env.reset()
for _ in range(n_random_steps):
obs = env.step(env.action_space.sample())[0]
else:
obs = env.reset()
policy.reset()
for _ in range(n_random_steps):
policy_obs = np.hstack((
obs['state_observation'],
obs['state_desired_goal'],
))
action, _ = policy.get_action(policy_obs)
obs, _, _, _ = env.step(action)
elif random_rollout_data: #ADD DATA WHERE JUST PUCK MOVES
if i % n_random_steps == 0:
env.reset()
policy.reset()
env_img = env._get_obs()['image_observation']
if random_rollout_data_set_to_goal:
env.set_to_goal(env.get_goal())
obs = env._get_obs()
u = policy.get_action_from_raw_action(
env.action_space.sample())
env.step(u)
elif oracle_dataset_using_set_to_goal:
print(i)
goal = env.sample_goal()
env.set_to_goal(goal)
obs = env._get_obs()
else:
env.reset()
for _ in range(n_random_steps):
obs = env.step(env.action_space.sample())[0]
img = obs['image_observation']
if use_env_labels:
labels.append(obs['label'])
if save_trajectories:
dataset['observations'][
i // n_random_steps,
i % n_random_steps, :] = unormalize_image(img)
dataset['actions'][i // n_random_steps,
i % n_random_steps, :] = u
dataset['env'][i // n_random_steps, :] = unormalize_image(
env_img)
else:
dataset[i, :] = unormalize_image(img)
if show:
img = img.reshape(3, imsize, imsize).transpose()
img = img[::-1, :, ::-1]
cv2.imshow('img', img)
cv2.waitKey(1)
# radius = input('waiting...')
print("done making training data", filename, time.time() - now)
np.save(filename, dataset)
#np.save(filename[:-4] + 'labels.npy', np.array(labels))
info['train_labels'] = []
info['test_labels'] = []
dataset = data_filter_fn(dataset)
if use_linear_dynamics and conditional_vae_dataset:
num_trajectories = N // n_random_steps
n = int(num_trajectories * test_p)
train_dataset = ConditionalDynamicsDataset({
'observations':
dataset['observations'][:n, :, :],
'actions':
dataset['actions'][:n, :, :],
'env':
dataset['env'][:n, :]
})
test_dataset = ConditionalDynamicsDataset({
'observations':
dataset['observations'][n:, :, :],
'actions':
dataset['actions'][n:, :, :],
'env':
dataset['env'][n:, :]
})
num_trajectories = N // n_random_steps
n = int(num_trajectories * test_p)
indices = np.arange(num_trajectories)
np.random.shuffle(indices)
train_i, test_i = indices[:n], indices[n:]
try:
train_dataset = ConditionalDynamicsDataset({
'observations':
dataset['observations'][train_i, :, :],
'actions':
dataset['actions'][train_i, :, :],
'env':
dataset['env'][train_i, :]
})
test_dataset = ConditionalDynamicsDataset({
'observations':
dataset['observations'][test_i, :, :],
'actions':
dataset['actions'][test_i, :, :],
'env':
dataset['env'][test_i, :]
})
except:
train_dataset = ConditionalDynamicsDataset({
'observations':
dataset['observations'][train_i, :, :],
'actions':
dataset['actions'][train_i, :, :],
})
test_dataset = ConditionalDynamicsDataset({
'observations':
dataset['observations'][test_i, :, :],
'actions':
dataset['actions'][test_i, :, :],
})
elif use_linear_dynamics:
num_trajectories = N // n_random_steps
n = int(num_trajectories * test_p)
train_dataset = TrajectoryDataset({
'observations':
dataset['observations'][:n, :, :],
'actions':
dataset['actions'][:n, :, :]
})
test_dataset = TrajectoryDataset({
'observations':
dataset['observations'][n:, :, :],
'actions':
dataset['actions'][n:, :, :]
})
elif enviorment_dataset:
n = int(n_random_steps * test_p)
train_dataset = EnvironmentDataset({
'observations':
dataset['observations'][:, :n, :],
})
test_dataset = EnvironmentDataset({
'observations':
dataset['observations'][:, n:, :],
})
elif conditional_vae_dataset:
num_trajectories = N // n_random_steps
n = int(num_trajectories * test_p)
indices = np.arange(num_trajectories)
np.random.shuffle(indices)
train_i, test_i = indices[:n], indices[n:]
if augment_data:
dataset_class = InitialObservationNumpyJitteringDataset
else:
dataset_class = InitialObservationNumpyDataset
if 'env' not in dataset:
dataset['env'] = dataset['observations'][:, 0]
if use_test_dataset and ('env' not in test_dataset):
test_dataset['env'] = test_dataset['observations'][:, 0]
if use_test_dataset:
train_dataset = dataset_class({
'observations':
dataset['observations'],
'env':
dataset['env']
})
test_dataset = dataset_class({
'observations':
test_dataset['observations'],
'env':
test_dataset['env']
})
else:
train_dataset = dataset_class({
'observations':
dataset['observations'][train_i, :, :],
'env':
dataset['env'][train_i, :]
})
test_dataset = dataset_class({
'observations':
dataset['observations'][test_i, :, :],
'env':
dataset['env'][test_i, :]
})
train_batch_loader_kwargs = variant.get(
'train_batch_loader_kwargs',
dict(
batch_size=batch_size,
num_workers=0,
))
test_batch_loader_kwargs = variant.get(
'test_batch_loader_kwargs',
dict(
batch_size=batch_size,
num_workers=0,
))
train_data_loader = data.DataLoader(train_dataset,
shuffle=True,
drop_last=True,
**train_batch_loader_kwargs)
test_data_loader = data.DataLoader(test_dataset,
shuffle=True,
drop_last=True,
**test_batch_loader_kwargs)
train_dataset = InfiniteBatchLoader(train_data_loader)
test_dataset = InfiniteBatchLoader(test_data_loader)
else:
n = int(N * test_p)
train_dataset = ImageObservationDataset(dataset[:n, :])
test_dataset = ImageObservationDataset(dataset[n:, :])
return train_dataset, test_dataset, info
multiworld.register_all_envs()
#env = gym.make('SawyerPickupMultiobj-v0')
#env = gym.make('SawyerPickupWideEnv-v0')
env = gym.make('SawyerMultiObj-v0')
#env = gym.make('SawyerPushNIPS-v0'
env = ImageEnv(
env,
imsize=imsize,
init_camera=sawyer_pusher_camera_upright_v3,
transpose=True,
normalize=True,
)
i = 0
env.reset()
for j in range(0, 2000000):
action = np.array(
[random.random() - .5,
random.random() - .5,
random.random() - .5])
obs = env.step(action)[0]['image_observation']
obs_img = 255 * obs.reshape(3, 480, 480).transpose()
cv2.imwrite('/home/lab/imgs/obs' + str(j) + '.png', obs_img[..., ::-1])
#goal = env.sample_goal()
#obs_img = 255*goal['desired_goal'].reshape(3, 480, 480).transpose()
#cv2.imwrite('a.png', obs_img)
#cv2.imshow('window', obs_img)
#cv2.waitKey()
def generate_vae_dataset(
env_class,
N=10000,
test_p=0.9,
use_cached=True,
imsize=84,
show=False,
init_camera=sawyer_init_camera_zoomed_in,
dataset_path=None,
env_kwargs=None,
oracle_dataset=False,
n_random_steps=100,
):
if env_kwargs is None:
env_kwargs = {}
filename = "/tmp/{}_{}_{}_oracle{}.npy".format(
env_class.__name__,
str(N),
init_camera.__name__,
oracle_dataset,
)
info = {}
if dataset_path is not None:
filename = local_path_from_s3_or_local_path(dataset_path)
dataset = np.load(filename)
N = dataset.shape[0]
elif use_cached and osp.isfile(filename):
dataset = np.load(filename)
print("loaded data from saved file", filename)
else:
now = time.time()
env = env_class(**env_kwargs)
env = ImageEnv(
env,
imsize,
init_camera=init_camera,
transpose=True,
normalize=True,
)
env.reset()
info['env'] = env
dataset = np.zeros((N, imsize * imsize * 3))
for i in range(N):
if oracle_dataset:
goal = env.sample_goal()
env.set_to_goal(goal)
else:
env.reset()
for _ in range(n_random_steps):
obs = env.step(env.action_space.sample())[0]
obs = env.step(env.action_space.sample())[0]
img = obs['image_observation']
dataset[i, :] = img
if show:
img = img.reshape(3, 84, 84).transpose()
img = img[::-1, :, ::-1]
cv2.imshow('img', img)
cv2.waitKey(1)
# radius = input('waiting...')
print("done making training data", filename, time.time() - now)
np.save(filename, dataset)
n = int(N * test_p)
train_dataset = dataset[:n, :]
test_dataset = dataset[n:, :]
return train_dataset, test_dataset, info
def generate_vae_dataset(variant):
"""
If not provided a pre-train vae dataset generation function, this function will be used to collect
the dataset for training vae.
"""
import rlkit.torch.pytorch_util as ptu
import gym
import multiworld
multiworld.register_all_envs()
print("generating vae dataset with original images")
env_class = variant.get('env_class', None)
env_kwargs = variant.get('env_kwargs', None)
env_id = variant.get('env_id', None)
N = variant.get('N', 10000)
test_p = variant.get('test_p', 0.9)
use_cached = variant.get('use_cached', True)
imsize = variant.get('imsize', 84)
num_channels = variant.get('num_channels', 3)
show = variant.get('show', False)
init_camera = variant.get('init_camera', None)
dataset_path = variant.get('dataset_path', None)
oracle_dataset_using_set_to_goal = variant.get(
'oracle_dataset_using_set_to_goal', False)
random_rollout_data = variant.get('random_rollout_data', False)
random_and_oracle_policy_data = variant.get(
'random_and_oracle_policy_data', False)
random_and_oracle_policy_data_split = variant.get(
'random_and_oracle_policy_data_split', 0)
policy_file = variant.get('policy_file', None)
n_random_steps = variant.get('n_random_steps', 100)
vae_dataset_specific_env_kwargs = variant.get(
'vae_dataset_specific_env_kwargs', None)
save_file_prefix = variant.get('save_file_prefix', None)
non_presampled_goal_img_is_garbage = variant.get(
'non_presampled_goal_img_is_garbage', None)
tag = variant.get('tag', '')
info = {}
if dataset_path is not None:
print('load vae training dataset from: ', dataset_path)
pjhome = os.environ['PJHOME']
dataset = np.load(osp.join(pjhome, dataset_path),
allow_pickle=True).item()
if isinstance(dataset, dict):
dataset = dataset['image_desired_goal']
dataset = unormalize_image(dataset)
N = dataset.shape[0]
else:
if env_kwargs is None:
env_kwargs = {}
if save_file_prefix is None:
save_file_prefix = env_id
if save_file_prefix is None:
save_file_prefix = env_class.__name__
filename = "/tmp/{}_N{}_{}_imsize{}_random_oracle_split_{}{}.npy".format(
save_file_prefix,
str(N),
init_camera.__name__ if init_camera else '',
imsize,
random_and_oracle_policy_data_split,
tag,
)
if use_cached and osp.isfile(filename):
dataset = np.load(filename)
print("loaded data from saved file", filename)
else:
now = time.time()
if env_id is not None:
import gym
import multiworld
multiworld.register_all_envs()
env = gym.make(env_id)
else:
if vae_dataset_specific_env_kwargs is None:
vae_dataset_specific_env_kwargs = {}
for key, val in env_kwargs.items():
if key not in vae_dataset_specific_env_kwargs:
vae_dataset_specific_env_kwargs[key] = val
env = env_class(**vae_dataset_specific_env_kwargs)
if not isinstance(env, ImageEnv):
env = ImageEnv(
env,
imsize,
init_camera=init_camera,
transpose=True,
normalize=True,
non_presampled_goal_img_is_garbage=
non_presampled_goal_img_is_garbage,
)
else:
imsize = env.imsize
env.non_presampled_goal_img_is_garbage = non_presampled_goal_img_is_garbage
env.reset()
info['env'] = env
if random_and_oracle_policy_data:
policy_file = load_local_or_remote_file(policy_file)
policy = policy_file['policy']
policy.to(ptu.device)
if random_rollout_data:
from rlkit.exploration_strategies.ou_strategy import OUStrategy
policy = OUStrategy(env.action_space)
dataset = np.zeros((N, imsize * imsize * num_channels),
dtype=np.uint8)
for i in range(N):
if random_and_oracle_policy_data:
num_random_steps = int(N *
random_and_oracle_policy_data_split)
if i < num_random_steps:
env.reset()
for _ in range(n_random_steps):
obs = env.step(env.action_space.sample())[0]
else:
obs = env.reset()
policy.reset()
for _ in range(n_random_steps):
policy_obs = np.hstack((
obs['state_observation'],
obs['state_desired_goal'],
))
action, _ = policy.get_action(policy_obs)
obs, _, _, _ = env.step(action)
elif oracle_dataset_using_set_to_goal:
print(i)
goal = env.sample_goal()
env.set_to_goal(goal)
obs = env._get_obs()
elif random_rollout_data:
if i % n_random_steps == 0:
g = dict(
state_desired_goal=env.sample_goal_for_rollout())
env.set_to_goal(g)
policy.reset()
# env.reset()
u = policy.get_action_from_raw_action(
env.action_space.sample())
obs = env.step(u)[0]
else:
print("using totally random rollouts")
for _ in range(n_random_steps):
obs = env.step(env.action_space.sample())[0]
img = obs[
'image_observation'] # NOTE yufei: this is already normalized image, of detype np.float64.
dataset[i, :] = unormalize_image(img)
np.save(filename, dataset)
n = int(N * test_p)
train_dataset = dataset[:n, :]
test_dataset = dataset[n:, :]
return train_dataset, test_dataset, info
def generate_uniform_dataset_door(num_imgs=1000,
use_cached_dataset=False,
init_camera=None,
imsize=48,
policy_file=None,
show=False,
path_length=100,
save_file_prefix=None,
env_id=None,
tag='',
dataset_path=None):
if dataset_path is not None:
dataset = load_local_or_remote_file(dataset_path)
return dataset
import gym
from gym.envs import registration
# trigger registration
import multiworld.envs.pygame
import multiworld.envs.mujoco
env = gym.make(env_id)
env = ImageEnv(
env,
imsize,
init_camera=init_camera,
transpose=True,
normalize=True,
)
env.non_presampled_goal_img_is_garbage = True
if save_file_prefix is None and env_id is not None:
save_file_prefix = env_id
filename = "/tmp/{}_N{}_imsize{}uniform_images_{}.npy".format(
save_file_prefix,
str(num_imgs),
env.imsize,
tag,
)
if use_cached_dataset and osp.isfile(filename):
images = np.load(filename)
print("Loaded data from {}".format(filename))
return images
policy_file = load_local_or_remote_file(policy_file)
policy = policy_file['policy']
policy.to(ptu.device)
print('Sampling Uniform Dataset')
dataset = np.zeros((num_imgs, 3 * env.imsize**2), dtype=np.uint8)
for j in range(num_imgs):
obs = env.reset()
policy.reset()
for i in range(path_length):
policy_obs = np.hstack((
obs['state_observation'],
obs['state_desired_goal'],
))
action, _ = policy.get_action(policy_obs)
obs, _, _, _ = env.step(action)
img_f = obs['image_observation']
if show:
img = obs['image_observation']
img = img.reshape(3, env.imsize, env.imsize).transpose()
img = img[::-1, :, ::-1]
cv2.imshow('img', img)
cv2.waitKey(1)
print(j)
dataset[j, :] = unormalize_image(img_f)
temp = env.reset_free
env.reset_free = True
env.reset()
env.reset_free = temp
np.save(filename, dataset)
print("Saving file to {}".format(filename))
return dataset
def main(training_data_dir, validation_data_dir, test_data_dir, imsize):
if not os.path.exists(training_data_dir): os.makedirs(training_data_dir)
if not os.path.exists(validation_data_dir):
os.makedirs(validation_data_dir)
if not os.path.exists(test_data_dir): os.makedirs(test_data_dir)
backSub = cv2.createBackgroundSubtractorMOG2(history=10)
# Registering required nmultiworld environments.
multiworld.register_all_envs()
base_env_background = gym.make('SawyerPushHurdlePuckAndRobotInvisible-v0')
env_background = ImageEnv(base_env_background,
imsize=imsize,
init_camera=camera,
transpose=True,
normalize=True)
env_background.reset()
base_env_hurdle = gym.make('SawyerPushHurdlePuckInvisible-v0')
env_hurdle = ImageEnv(base_env_hurdle,
imsize=imsize,
init_camera=camera,
transpose=True,
normalize=True)
env_hurdle.reset()
# Generating training, validation and test data
print("Training background subctractor")
for i in range(10):
action = env_background.action_space.sample()
next_obs, reward, done, info = env_background.step(action)
image = next_obs['observation']
image = unnormalize_image(image)
bg_mask = backSub.apply(image, learningRate=-1)
print("Generating training data")
for i in range(1000):
action = env_hurdle.action_space.sample()
next_obs, reward, done, info = env_hurdle.step(action)
image = next_obs['observation']
image = unnormalize_image(image)
fg_mask = backSub.apply(image, learningRate=0)
fg_mask = fg_mask / fg_mask.max()
cv2.imwrite(training_data_dir + "/rgb_" + str(i) + ".jpg", image)
cv2.imwrite(training_data_dir + "/mask_" + str(i) + ".jpg", fg_mask)
print("Generating validation data")
for i in range(200):
action = env_hurdle.action_space.sample()
next_obs, reward, done, info = env_hurdle.step(action)
image = next_obs['observation']
image = unnormalize_image(image)
fg_mask = backSub.apply(image, learningRate=0)
fg_mask = fg_mask / fg_mask.max()
cv2.imwrite(validation_data_dir + "/rgb_" + str(i) + ".jpg", image)
cv2.imwrite(validation_data_dir + "/mask_" + str(i) + ".jpg", fg_mask)
print("Generating testing data")
for i in range(200):
action = env_hurdle.action_space.sample()
next_obs, reward, done, info = env_hurdle.step(action)
image = next_obs['observation']
image = unnormalize_image(image)
fg_mask = backSub.apply(image, learningRate=0)
fg_mask = fg_mask / fg_mask.max()
cv2.imwrite(test_data_dir + "/rgb_" + str(i) + ".jpg", image)
cv2.imwrite(test_data_dir + "/mask_" + str(i) + ".jpg", fg_mask)
print("Completed generating data at:", training_data_dir)
def generate_vae_dataset(variant):
import cv2
env_class = variant.get('env_class', None)
env_kwargs = variant.get('env_kwargs', None)
env_id = variant.get('env_id', None)
N = variant.get('N', 10000)
use_images = variant.get('use_images', True)
imsize = variant.get('imsize', 84)
show = variant.get('show', False)
init_camera = variant.get('init_camera', None)
oracle_dataset = variant.get('oracle_dataset', False)
if 'n_random_steps' in variant:
n_random_steps = variant['n_random_steps']
else:
if oracle_dataset:
n_random_steps = 3
else:
n_random_steps = 100
vae_dataset_specific_env_kwargs = variant.get(
'vae_dataset_specific_env_kwargs', None)
non_presampled_goal_img_is_garbage = variant.get(
'non_presampled_goal_img_is_garbage', None)
from multiworld.core.image_env import ImageEnv, unormalize_image
info = {}
from railrl.core import logger
logdir = logger.get_snapshot_dir()
filename = osp.join(logdir, "vae_dataset.npy")
now = time.time()
if env_id is not None:
import gym
env = gym.make(env_id)
else:
if vae_dataset_specific_env_kwargs is None:
vae_dataset_specific_env_kwargs = {}
for key, val in env_kwargs.items():
if key not in vae_dataset_specific_env_kwargs:
vae_dataset_specific_env_kwargs[key] = val
env = env_class(**vae_dataset_specific_env_kwargs)
if not isinstance(env, ImageEnv):
env = ImageEnv(
env,
imsize,
init_camera=init_camera,
transpose=True,
normalize=True,
non_presampled_goal_img_is_garbage=
non_presampled_goal_img_is_garbage,
)
else:
imsize = env.imsize
env.non_presampled_goal_img_is_garbage = non_presampled_goal_img_is_garbage
env.reset()
info['env'] = env
if use_images:
data_size = len(env.observation_space.spaces['image_observation'].low)
dtype = np.uint8
else:
data_size = len(env.observation_space.spaces['state_observation'].low)
dtype = np.float32
state_size = len(env.observation_space.spaces['state_observation'].low)
dataset = {
'obs': np.zeros((N, data_size), dtype=dtype),
'actions': np.zeros((N, len(env.action_space.low)), dtype=np.float32),
'next_obs': np.zeros((N, data_size), dtype=dtype),
'obs_state': np.zeros((N, state_size), dtype=np.float32),
'next_obs_state': np.zeros((N, state_size), dtype=np.float32),
}
for i in range(N):
if i % (N / 50) == 0:
print(i)
if oracle_dataset:
if i % 100 == 0:
env.reset()
goal = env.sample_goal()
env.set_to_goal(goal)
for _ in range(n_random_steps):
env.step(env.action_space.sample())
else:
env.reset()
for _ in range(n_random_steps):
env.step(env.action_space.sample())
obs = env._get_obs()
if use_images:
dataset['obs'][i, :] = unormalize_image(obs['image_observation'])
else:
dataset['obs'][i, :] = obs['state_observation']
dataset['obs_state'][i, :] = obs['state_observation']
action = env.action_space.sample()
dataset['actions'][i, :] = action
obs = env.step(action)[0]
img = obs['image_observation']
if use_images:
dataset['next_obs'][i, :] = unormalize_image(img)
else:
dataset['next_obs'][i, :] = obs['state_observation']
dataset['next_obs_state'][i, :] = obs['state_observation']
show = True #todo del own
if show:
img = img.reshape(3, imsize, imsize).transpose((1, 2, 0))
img = img[::, :, ::-1]
cv2.imshow('img', img)
cv2.waitKey(1000)
print("keys and shapes:")
for k in dataset.keys():
print(k, dataset[k].shape)
print("done making training data", filename, time.time() - now)
np.save(filename, dataset)
