def run():
env = CarRacing(
allow_reverse=False,
show_info_panel=False,
num_tracks=2,
num_lanes=2,
num_lanes_changes=0,
num_obstacles=100,
random_obstacle_x_position=False,
random_obstacle_shape=False,
)
x_count = 0
t_count = 0
num_tracks = 300
for i in range(num_tracks):
env.reset()
print("track %i / %i" % (i, num_tracks))
if (env.info['x'] == True).sum() > 0: x_count += 1
if (env.info['t'] == True).sum() > 0: t_count += 1
print('')
i += 1
print('num of tracks:', str(num_tracks))
print('time took per track is {0}s'.format(1))
print('{0}% of tracks have x intersections'.format(x_count / num_tracks))
print('{0}% of tracks have t intersections'.format(t_count / num_tracks))
def find_roads():
path = './touching_tracks_tests'
# Check if dir exists TODO
if os.path.isdir(path):
# Remove files TODO
shutil.rmtree(path)
# Create dir TODO
os.mkdir(path)
env = CarRacing(
allow_reverse=False,
show_info_panel=False,
num_tracks=2,
num_lanes=2,
num_lanes_changes=0,
num_obstacles=100,
random_obstacle_x_position=False,
random_obstacle_shape=False,)
env.change_zoom()
for j in range(100):
env.reset()
for i in range(len(env.tracks[0])):
prev_tile = env.tracks[0][i-2]
curr_tile = env.tracks[0][i-1]
next_tile = env.tracks[0][i]
if any(curr_tile[0] != prev_tile[1]):
set_trace()
elif any(curr_tile[1] != next_tile[0]):
set_trace()
env.screenshot(path,name=str(j),quality='high')
np.save(path + "/info_" + str(j) + ".csv", env.info)
np.save(path + "/track0_" + str(j) + ".csv", env.tracks[0])
np.save(path + "/track1_" + str(j) + ".csv", env.tracks[1])
def __init__(self, name="car_racing", renders=False, record_data=False, is_discrete=True, state_dim=-1,
learn_states=False, save_path='srl_zoo/data/', srl_model="raw_pixels", env_rank=0, srl_pipe=None,
lookahead=5, **_):
"""
Gym wrapper for Racing car environment
WARNING: to be compatible with kuka scripts, additional keyword arguments are discarded
:param name: (str) name of the folder where recorded data will be stored
:param renders: (bool) Whether to display the GUI or not
:param record_data: (bool) Set to true, record frames with the rewards.
:param is_discrete: (bool) Whether to use discrete or continuous actions
:param state_dim: (int) When learning states
:param learn_states: (bool)
:param save_path: (str) location where the saved data should go
:param srl_model: (str) The SRL_model used
:param env_rank: (int) the number ID of the environment
:param srl_pipe: (Queue, [Queue]) contains the input and output of the SRL model
:param lookahead: (int) How many segments ahead of the current position of the track should the target be
"""
SRLGymEnv.__init__(self, srl_model=srl_model, relative_pos=RELATIVE_POS, env_rank=env_rank, srl_pipe=srl_pipe)
GymCarRacing.__init__(self)
self._renders = renders
self._width = RENDER_WIDTH
self._height = RENDER_HEIGHT
self._is_discrete = is_discrete
self.lookahead = lookahead
self.relative_pos = RELATIVE_POS
self._env_step_counter = 0
self._observation = None
self.saver = None
if record_data:
self.saver = EpisodeSaver(name, None, state_dim, globals_=getGlobals(), relative_pos=RELATIVE_POS,
learn_states=learn_states, path=save_path)
# Accelerate, brake, stear left, stear right
if self._is_discrete:
self.action_space = spaces.Discrete(N_DISCRETE_ACTIONS)
else:
self.action_space = spaces.Box(low=-1, high=1, shape=(3,), dtype=np.float32)
if self.srl_model == "ground_truth":
self.state_dim = self.getGroundTruthDim()
if self.srl_model == "raw_pixels":
self.observation_space = spaces.Box(low=0, high=255, shape=(self._height, self._width, 3), dtype=np.uint8)
else:
self.observation_space = spaces.Box(low=-np.inf, high=np.inf, shape=(self.state_dim,), dtype=np.float32)
def test_two_lanes_with_no_lane_changes(self):
env = CarRacing(num_lanes=2,num_lanes_changes=0)
env.reset()
assert env._get_extremes_of_position(0,border=1) == (-3-1/3,3+1/3)
assert env._get_extremes_of_position(1,border=0) == (-6-2/3,+6+2/3)
env.close()
del env
def evaluate():
"""
"""
# action variables
a = np.array([0.0, 0.0, 0.0])
# init environement
env = CarRacing()
env.render()
env.reset()
for episode in range(5):
observation = env.reset()
# init modules of the pipeline
LD_module = LaneDetection()
LatC_module = LateralController()
LongC_module = LongitudinalController()
reward_per_episode = 0
for t in range(500):
# perform step
s, r, done, speed, info = env.step(a)
# lane detection
lane1, lane2 = LD_module.lane_detection(s)
# waypoint and target_speed prediction
waypoints = waypoint_prediction(lane1, lane2)
target_speed = target_speed_prediction(waypoints,
max_speed=60,
exp_constant=4.5)
# control
a[0] = LatC_module.stanley(waypoints, speed)
a[1], a[2] = LongC_module.control(speed, target_speed)
# reward
reward_per_episode += r
env.render()
print('episode %d \t reward %f' % (episode, reward_per_episode))
def game_runner():
from pyglet.window import key
a = np.array([0.0, 0.0, 0.0])
def key_press(k, mod):
global restart
if k == 0xff0d: restart = True
if k == key.LEFT: a[0] = -1.0
if k == key.RIGHT: a[0] = +1.0
if k == key.UP: a[1] = +1.0
if k == key.DOWN: a[2] = +0.8
def key_release(k, mod):
if k == key.LEFT and a[0] == -1.0: a[0] = 0
if k == key.RIGHT and a[0] == +1.0: a[0] = 0
if k == key.UP: a[1] = 0
if k == key.DOWN: a[2] = 0
env = CarRacing()
env.render()
env.viewer.window.on_key_press = key_press
env.viewer.window.on_key_release = key_release
while True:
env.reset()
total_reward = 0.0
steps = 0
restart = False
while True:
s, r, done, info = env.step(a)
total_reward += r
if steps == 900:
print("\n")
print("_______________________________")
print("\n")
print("Human Intelligence Result:")
print("Total Steps: {}".format(steps))
print("Total Reward: {:.0f}".format(total_reward))
print("\n")
print("_______________________________")
print("\n")
break
steps += 1
env.render()
if restart: break
env.monitor.close()
def test_two_tracks(self):
env = CarRacing(num_tracks=2)
env.reset()
assert len(env.tracks) == 2
env.close()
def multiple_runs(v, on):
env = CarRacing()
z_set = []
action_set = []
for run in range(MAX_RUNS):
zs = []
actions = []
state = env.reset()
env.render() # must have!
# done = False
counter = 0
for game_time in range(MAX_GAME_TIME):
# env.render()
action = generate_action()
obs = state_to_1_batch_tensor(state)
_, _, _, z = v(obs)
z = z.detach().numpy()
z = z.reshape(32)
# print(z.shape)
# if game_time == 5:
# plt.imshow(state)
# plt.show()
# state = _process_frame(state)
# plt.imshow(state)
# plt.show()
zs.append(z)
actions.append(action)
state, r, done, _ = env.step(action)
# print(r)
print(
'RUN:{},GT:{},DATA:{}'.format(
run, game_time, len(actions)
)
)
# if counter == REST_NUM:
#
# position = np.random.randint(len(env.track))
# env.car = Car(env.world, *env.track[position][1:4])
# counter = 0
# counter += 1
zs = np.array(zs, dtype=np.float16)
# print(zs.shape)
actions = np.array(actions, dtype=np.float16)
# print(actions.shape)
# np.save(dst + '/' + save_name, frame_and_action)
# np.savez_compressed(dst + '/' + save_name, action=actions, z=zs)
z_set.append(zs)
action_set.append(actions)
z_set = np.array(z_set)
# print(z_set.shape)
action_set = np.array(action_set)
# print(action_set.shape)
save_name = name_this + '_{}.npz'.format(on)
np.savez_compressed(dst + '/' + save_name, action=action_set, z=z_set)
def test_one_track(self):
env = CarRacing()
# Tracks should not exist before any reset
with pytest.raises(AttributeError):
env.tracks
env.reset()
assert len(env.tracks) == 1
env.close()
def run_caracing_by_hunman():
a = np.array([0.0, 0.0, 0.0])
def key_press(k, mod):
global restart
if k == 0xff0d: restart = True
if k == key.LEFT: a[0] = -1.0
if k == key.RIGHT: a[0] = +1.0
if k == key.UP: a[1] = +1.0
if k == key.DOWN:
a[2] = +0.8 # set 1.0 for wheels to block to zero rotation
def key_release(k, mod):
if k == key.LEFT and a[0] == -1.0: a[0] = 0
if k == key.RIGHT and a[0] == +1.0: a[0] = 0
if k == key.UP: a[1] = 0
if k == key.DOWN: a[2] = 0
env = CarRacing()
env.render()
env.viewer.window.on_key_press = key_press
env.viewer.window.on_key_release = key_release
while True:
env.reset()
total_reward = 0.0
steps = 0
restart = False
while True:
s, r, done, info = env.step(a)
total_reward += r
if steps % 200 == 0 or done:
print("\naction " + str(["{:+0.2f}".format(x) for x in a]))
print("step {} total_reward {:+0.2f}".format(
steps, total_reward))
steps += 1
env.render()
if done or restart: break
env.monitor.close()
def manual_check_of_not_allowing_touching_tracks():
path = './touching_tracks_tests'
# Check if dir exists TODO
if os.path.isdir(path):
# Remove files TODO
shutil.rmtree(path)
# Create dir TODO
os.mkdir(path)
env = CarRacing(
allow_reverse=False,
show_info_panel=False,
num_tracks=2,
num_lanes=2,
num_lanes_changes=0,
num_obstacles=100,
random_obstacle_x_position=False,
random_obstacle_shape=False,
)
env.change_zoom()
for i in range(100):
env.reset()
env.screenshot(path, name=str(i), quality='high')
from gym.envs.box2d.car_racing import CarRacing
if __name__ == '__main__':
env = CarRacing()
repeat = int(1.5e6)
for i in range(repeat):
print("%i of %i" % (i, repeat))
env.reset()
print("")
print('vae load success')
# from pyglet.window import key
action = np.array( [0.0, 0.0, 0.0] )
# def key_press(k, mod):
# global restart
# if k==0xff0d: restart = True
# if k==key.LEFT: a[0] = -1.0
# if k==key.RIGHT: a[0] = +1.0e
# if k==key.UP: a[1] = +1.0
# if k==key.DOWN: a[2] = +0.8 # set 1.0 for wheels to block to zero rotation
# def key_release(k, mod):
# if k==key.LEFT and a[0]==-1.0: a[0] = 0
# if k==key.RIGHT and a[0]==+1.0: a[0] = 0
# if k==key.UP: a[1] = 0
# if k==key.DOWN: a[2] = 0
env = CarRacing()
env.render()
# env.viewer.window.on_key_press = key_press
# env.viewer.window.on_key_release = key_release
while True:
env.reset()
total_reward = 0.0
steps = 0
restart = False
angle=[]
while True:
obs, reward, done, info = env.step(action)
total_reward += reward
obs=transform(obs).view(1,3,64,64)
recon_c, mu_c, var_c = model(obs)
mu, sigma = mu_c, var_c
def key_press(k, mod):
global restart
if k == 0xff0d: restart = True
if k == key.LEFT: a[0] = -1.0
if k == key.RIGHT: a[0] = +1.0
if k == key.UP: a[1] = +1.0
if k == key.DOWN:
a[2] = +0.8 # set 1.0 for wheels to block to zero rotation
def key_release(k, mod):
if k == key.LEFT and a[0] == -1.0: a[0] = 0
if k == key.RIGHT and a[0] == +1.0: a[0] = 0
if k == key.UP: a[1] = 0
if k == key.DOWN: a[2] = 0
env = CarRacing()
env.render()
env.viewer.window.on_key_press = key_press
env.viewer.window.on_key_release = key_release
while True:
env.reset()
total_reward = 0.0
steps = 0
restart = False
while True:
s, r, done, info = env.step(a)
total_reward += r
if steps % 200 == 0 or done:
print("\naction " + str(["{:+0.2f}".format(x) for x in a]))
print("step {} total_reward {:+0.2f}".format(
steps, total_reward))
self.z_dim * 2]
sigma = torch.exp(sigma)
# N * seq_len, n_mixture
pi = mixture[..., -self.n_mixture:]
pi = F.softmax(pi, -1)
# add temperature
if self.tau > 0:
pi /= self.tau
sigma *= self.tau**0.5
return pi, mean, sigma, hidden_state
# help(cma)
# help(cma.CMAEvolutionStrategy)
env = CarRacing()
path = '/media/ray/SSD/workspace/python/dataset/save_here/model/'
vae_name = 'vae_model.save'
V = torch.load(path + vae_name)
V = V.cpu()
V.train(False)
rnn_name = 'rnn.save'
M = torch.load(path + rnn_name)
M.cpu()
M.train(False)
num_z = 32
num_h = 256
input_size = num_z + num_h
output_size = 3
para = np.random.random(input_size * output_size + output_size)
print(para.shape)
def calculate_score_for_leaderboard():
"""
Evaluate the performance of the network. This is the function to be used for
the final ranking on the course-wide leader-board, only with a different set
of seeds. Better not change it.
"""
# action variables
a = np.array([0.0, 0.0, 0.0])
# init environement
env = CarRacing()
env.render()
env.reset()
seeds = [
22597174, 68545857, 75568192, 91140053, 86018367, 49636746, 66759182,
91294619, 84274995, 31531469
]
total_reward = 0
for episode in range(10):
env.seed(seeds[episode])
observation = env.reset()
# init modules of the pipeline
LD_module = LaneDetection(gradient_threshold=25, spline_smoothness=20)
LatC_module = LateralController(gain_constant=1.8,
damping_constant=0.05)
LongC_module = LongitudinalController(KD=0.001)
reward_per_episode = 0
for t in range(600):
# perform step
s, r, done, speed, info = env.step(a)
# lane detection
lane1, lane2 = LD_module.lane_detection(s)
# waypoint and target_speed prediction
waypoints = waypoint_prediction(lane1, lane2)
target_speed = target_speed_prediction(waypoints,
max_speed=60,
exp_constant=6)
# control
a[0] = LatC_module.stanley(waypoints, speed)
a[1], a[2] = LongC_module.control(speed, target_speed)
# reward
reward_per_episode += r
env.render()
print('episode %d \t reward %f' % (episode, reward_per_episode))
total_reward += np.clip(reward_per_episode, 0, np.infty)
print('---------------------------')
print(' total score: %f' % (total_reward / 10))
print('---------------------------')
def key_press(k, mod):
global restart
if k==0xff0d: restart = True
if k==key.LEFT: a[0] = -1.0
if k==key.RIGHT: a[0] = +1.0
if k==key.UP: a[1] = +1.0
if k==key.DOWN: a[2] = +0.8 # set 1.0 for wheels to block to zero rotation
def key_release(k, mod):
if k==key.LEFT and a[0]==-1.0: a[0] = 0
if k==key.RIGHT and a[0]==+1.0: a[0] = 0
if k==key.UP: a[1] = 0
if k==key.DOWN: a[2] = 0
# init environement
env = CarRacing()
env.render()
env.viewer.window.on_key_press = key_press
env.viewer.window.on_key_release = key_release
env.reset()
# define variables
total_reward = 0.0
steps = 0
restart = False
# init modules of the pipeline
LD_module = LaneDetection()
# init extra plot
fig = plt.figure()
from gym.envs.box2d.car_racing import CarRacing
from tqdm import tqdm
import numpy as np
if __name__ == "__main__":
env = CarRacing()
env.reset()
a = np.zeros(3)
for _ in tqdm(range(1000000)):
env.step(a)
# env.render()
def key_press(k, mod):
global restart
if k==0xff0d: restart = True
if k==key.LEFT: a[0] = -1.0
if k==key.RIGHT: a[0] = +1.0
if k==key.UP: a[1] = +1.0
if k==key.DOWN: a[2] = +0.8 # set 1.0 for wheels to block to zero rotation
def key_release(k, mod):
if k==key.LEFT and a[0]==-1.0: a[0] = 0
if k==key.RIGHT and a[0]==+1.0: a[0] = 0
if k==key.UP: a[1] = 0
if k==key.DOWN: a[2] = 0
# init environement
env = CarRacing()
env.render()
env.viewer.window.on_key_press = key_press
env.viewer.window.on_key_release = key_release
env.reset()
# define variables
total_reward = 0.0
steps = 0
restart = False
# init modules of the pipeline
LD_module = LaneDetection()
# init extra plot
fig = plt.figure()
import numpy as np
from gym.envs.box2d.car_racing import CarRacing
if __name__ == '__main__':
env = CarRacing(load_tracks_from="/hdd/Documents/HRL/tracks")
env.tracks_df = env.tracks_df[(env.tracks_df['x']) | (env.tracks_df['t'])]
repeat = 1000
for i in range(repeat):
print("%i of %i" % (i, repeat))
env.reset()
dictionary = env.understand_intersection(
np.random.choice(np.where(env.info['intersection_id'] != -1)[0]),
1)
# Checking that only one value can be None
print(dictionary)
if list(dictionary.values()).count(None) > 1:
print("breaking, more than 1 None")
break
for val in dictionary.values():
if val is not None and type(val) is not list and len(val) != 2:
print("Breaking, there is some value that \
is not a list or a different kind of list")
break
print("")
