def __getitem__(self, idx):
env = SimpleSimEnv(draw_curve=False)
observs = torch.zeros(self.batch_size, 3, 120, 160)
for i in range(self.batch_size):
obs = env.reset()
observs[i, :, :, :] = torch.FloatTensor(obs.transpose((2, 0, 1)))
return observs
def write_imgs_from_map(map_name: str, save_dir: pathlib.Path, test_percentage=0.3):
env = SimpleSimEnv(map_name=map_name)
file_path = pathlib.Path('experiments/demos_{}.json'.format(map_name))
if not file_path.is_file():
raise ValueError("Could not find the file containing the generated trajectories: {}".format(file_path))
data = json.loads(file_path.read_text())
demos = data['demos']
positions = map(lambda d: d['positions'], demos)
actions = map(lambda d: d['actions'], demos)
positions = sum(positions, [])
actions = sum(actions, [])
test_dir = save_dir / "test"
train_dir = save_dir / "train"
test_dir.mkdir(parents=True, exist_ok=True)
train_dir.mkdir(parents=True, exist_ok=True)
print("Found {} positions to be converted to images...".format(len(positions)))
for idx, position in enumerate(positions):
cur_pos = np.array(position[0])
cur_angle = position[1]
vels = actions[idx]
env.cur_pos = cur_pos
env.cur_angle = cur_angle
obs = env.render_obs().copy()
obs = obs[..., ::-1]
if random.random() < test_percentage:
img_path = test_dir / "{0:06d}.jpg".format(idx)
lbl_path = test_dir / "{0:06d}.txt".format(idx)
else:
img_path = train_dir / "{0:06d}.jpg".format(idx)
lbl_path = train_dir / "{0:06d}.txt".format(idx)
cv2.imwrite(img_path.as_posix(), obs)
lbl_path.write_text(" ".join(map(str, vels)))
def create_environment(args, with_heading=True):
if args.env_name == 'SimpleSim-v0':
environment = SimpleSimEnv(
max_steps=math.inf,
domain_rand=False,
draw_curve=True
)
else:
environment = gym.make(args.env_name)
if with_heading:
environment = HeadingWrapper(environment)
return environment
def main():
""" Main launcher that starts the gym thread when the command 'duckietown-start-gym' is invoked
"""
# get parameters from environment (set during docker launch) otherwise take default
map = os.getenv('DUCKIETOWN_MAP', DEFAULTS["map"])
domain_rand = bool(
os.getenv('DUCKIETOWN_DOMAIN_RAND', DEFAULTS["domain_rand"]))
env = SimpleSimEnv(
map_name=map,
# draw_curve = args.draw_curve,
# draw_bbox = args.draw_bbox,
domain_rand=domain_rand)
obs = env.reset()
# env.render("rgb_array") # TODO: do we need this? does this initialize anything?
publisher_socket = None
command_socket, command_poll = make_pull_socket()
print("Simulator listening to incoming connections...")
while True:
if has_pull_message(command_socket, command_poll):
success, data = receive_data(command_socket)
if not success:
print(data) # in error case, this will contain the err msg
continue
if data["topic"] == 0:
obs, reward, done, misc = env.step(data["msg"])
print('step_count = %s, reward=%.3f' %
(env.step_count, reward))
if done:
env.reset()
if data["topic"] == 1:
print("received ping:", data)
# can only initialize socket after first listener is connected - weird ZMQ bug
if publisher_socket is None:
publisher_socket = make_pub_socket(for_images=True)
if data["topic"] in [0, 1]:
send_array(publisher_socket, obs)
optimizer.zero_grad()
# forward + backward + optimize
output = model(image)
loss = (output - target).norm(2).mean()
loss.backward()
optimizer.step()
error = (output - target).abs().mean()
return loss.data[0], error.data[0]
if __name__ == "__main__":
env = SimpleSimEnv()
env.reset()
model = Model()
model.printInfo()
model.cuda()
optimizer = optim.Adam(model.parameters(), lr=0.001)
avg_error = 0
for epoch in range(1, 1000000):
startTime = time.time()
images, targets = genBatch()
images = Variable(torch.from_numpy(images).float()).cuda()
targets = Variable(torch.from_numpy(targets).float()).cuda()
parser.add_argument('--map-file', default='gym_duckietown/maps/udem1.yaml')
parser.add_argument('--draw-curve',
action='store_true',
help='draw the lane following curve')
parser.add_argument('--no-random',
action='store_true',
help='disable domain randomization')
parser.add_argument('--full-res',
action='store_true',
help='render at full window resolution')
args = parser.parse_args()
if args.env_name == 'SimpleSim-v0':
env = SimpleSimEnv(map_file=args.map_file,
draw_curve=args.draw_curve,
domain_rand=not args.no_random,
full_res=args.full_res)
else:
env = gym.make(args.env_name)
env.reset()
env.render()
def save_numpy_img(file_name, img):
img = np.ascontiguousarray(img)
img = (img * 255).astype(np.uint8)
img = np.flip(img, 0)
from skimage import io
io.imsave(file_name, img)
#!/usr/bin/env python3
import time
import numpy as np
import gym
import gym_duckietown
from gym_duckietown.envs import SimpleSimEnv
# Benchmark loading time
st = time.time()
env = SimpleSimEnv(max_steps=20000)
env.seed(0)
env.reset()
load_time = 1000 * (time.time() - st)
# Benchmark the reset time
st = time.time()
for i in range(100):
env.reset()
reset_time = 1000 * (time.time() - st) / 100
# Benchmark the rendering/update speed
num_frames = 0
st = time.time()
while True:
dt = time.time() - st
if dt > 5:
import argparse
import pyglet
import numpy as np
import gym
import gym_duckietown
from gym_duckietown.envs import SimpleSimEnv
parser = argparse.ArgumentParser()
parser.add_argument('--env-name', default='SimpleSim-v0')
parser.add_argument('--draw-curve', action='store_true')
args = parser.parse_args()
if args.env_name == 'SimpleSim-v0':
env = SimpleSimEnv(draw_curve=args.draw_curve)
else:
env = gym.make(args.env_name)
env.reset()
env.render()
def save_numpy_img(file_name, img):
img = np.ascontiguousarray(img)
img = (img * 255).astype(np.uint8)
img = np.flip(img, 0)
from skimage import io
io.imsave(file_name, img)
env.unwrapped.cur_pos = cur_pos
env.unwrapped.cur_angle = cur_angle
obs = env.unwrapped.render_obs().copy()
obs = obs.transpose(2, 0, 1)
return obs, vels
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument('--map-name', required=True)
args = parser.parse_args()
load_data(args.map_name)
env = SimpleSimEnv(map_name=args.map_name)
env = HeadingWrapper(env)
model = Model()
model.train()
if torch.cuda.is_available():
model.cuda()
print_model_info(model)
# weight_decay is L2 regularization, helps avoid overfitting
optimizer = optim.SGD(
model.parameters(),
lr=0.0004,
weight_decay=1e-3
)
parser.add_argument('--env-name', default='SimpleSim-v0')
parser.add_argument('--map-name', default='udem1')
parser.add_argument('--draw-curve',
action='store_true',
help='draw the lane following curve')
parser.add_argument('--draw-bbox',
action='store_true',
help='draw collision detection bounding boxes')
parser.add_argument('--domain-rand',
action='store_true',
help='enable domain randomization')
args = parser.parse_args()
if args.env_name == 'SimpleSim-v0':
env = SimpleSimEnv(map_name=args.map_name,
draw_curve=args.draw_curve,
draw_bbox=args.draw_bbox,
domain_rand=args.domain_rand)
else:
env = gym.make(args.env_name)
env.reset()
env.render()
def save_numpy_img(file_name, img):
img = np.ascontiguousarray(img)
img = (img * 255).astype(np.uint8)
from skimage import io
io.imsave(file_name, img)
from gym_duckietown.envs import SimpleSimEnv
from PIL import Image
import numpy as np
env = SimpleSimEnv(draw_curve=False)
num_samples = 10
for i in range(num_samples):
obs = env.reset()
img = Image.fromarray(np.flipud((obs * 255)).astype('uint8'))
img.save(fp='representation_analysis/trajectories/{}.jpg'.format(i))
if i + 1 % 1000 == 0:
print('got to sample {}'.format(i))
env.close()
#!/usr/bin/env python3
import time
import numpy as np
import gym
import gym_duckietown
from gym_duckietown.envs import SimpleSimEnv
# Benchmark loading time
st = time.time()
env = SimpleSimEnv(max_steps=20000, map_name='loop_obstacles')
env.seed(0)
env.reset()
load_time = 1000 * (time.time() - st)
# Benchmark the reset time
st = time.time()
for i in range(100):
env.reset()
reset_time = 1000 * (time.time() - st) / 100
# Benchmark the rendering/update speed
num_frames = 0
st = time.time()
while True:
dt = time.time() - st
if dt > 5:
env = gym.make('SimpleSim-v0')
first_obs = env.reset()
# Check that the human rendering resembles the agent's view
first_render = env.render('rgb_array')
m0 = first_obs.mean()
m1 = first_render.mean()
assert m0 > 0 and m0 < 255
assert abs(m0 - m1) < 5
# Try stepping a few times
for i in range(0, 10):
obs, _, _, _ = env.step(np.array([0.1, 0.099]))
# Try loading each of the available map files
for map_file in os.listdir('gym_duckietown/maps'):
map_name = map_file.split('.')[0]
env = SimpleSimEnv(map_name=map_name)
# Test the multi-map environment
env = MultiMapEnv()
# Check that we do not spawn too close to obstacles
env = SimpleSimEnv(map_name='loop_obstacles')
for i in range(0, 75):
obs = env.reset()
assert not env._collision(), "collision on spawn"
env.step(np.array([0.1, 0.1]))
assert not env._collision(), "collision after one step"
import argparse
import numpy as np
import tqdm
from PIL import Image
import os
import torch
from torch.autograd import Variable
from gym_duckietown.envs import SimpleSimEnv
from representation_analysis.models import VAE
env = SimpleSimEnv(draw_curve=False)
parser = argparse.ArgumentParser(description='VAE')
parser.add_argument('--num_samples_train',
type=int,
default=5000,
help='num samples for training')
parser.add_argument('--num_samples_test',
type=int,
default=500,
help='num samples per testing')
parser.add_argument('--no-cuda',
action='store_true',
default=False,
help='Enables CUDA training')
parser.add_argument('--saved_model', type=str, help='Save file to use')
parser.add_argument('--state_size',
type=int,
default=50,
help='Size of latent code (default: 100)')
default=30,
help='number of values to compute average over')
parser.add_argument(
'--heading-speed',
type=float,
default=0.7,
help='We\'re assuming the heading speed to be constant, default is 0.7')
args = parser.parse_args()
logger_dir = os.path.join(
"logs", "{}_{}_{}".format(args.env_name, args.algo, args.suffix))
if not os.path.exists(logger_dir):
os.makedirs(logger_dir)
logger = Logger(logger_dir)
env = HeadingWrapper(SimpleSimEnv(map_name="small_loop"))
env.seed(args.seed)
if torch.cuda.is_available():
device = torch.device("cuda:0")
torch.cuda.manual_seed(args.seed)
else:
device = torch.device("cpu")
torch.manual_seed(args.seed)
np.random.seed(args.seed)
obs_shape = env.observation_space.shape
obs_shape = (obs_shape[0] * args.num_stack, *obs_shape[1:])
if len(env.observation_space.shape) == 3:
parser = argparse.ArgumentParser()
parser.add_argument('--env-name', default='SimpleSim-v0')
parser.add_argument('--map-name', default='udem1')
parser.add_argument('--draw-curve',
action='store_true',
help='draw the lane following curve')
parser.add_argument('--no-random',
action='store_true',
help='disable domain randomization')
args = parser.parse_args()
if args.env_name == 'SimpleSim-v0':
env = SimpleSimEnv(map_name=args.map_name,
draw_curve=args.draw_curve,
domain_rand=not args.no_random)
else:
env = gym.make(args.env_name)
env.reset()
env.render()
def save_numpy_img(file_name, img):
img = np.ascontiguousarray(img)
img = (img * 255).astype(np.uint8)
img = np.flip(img, 0)
from skimage import io
io.imsave(file_name, img)
from gym_duckietown.envs import SimpleSimEnv
from PIL import Image
import numpy as np
import argparse
import os
env = SimpleSimEnv(draw_curve=False)
parser = argparse.ArgumentParser(description='VAE')
parser.add_argument('--num_samples_train',
type=int,
default=10000,
help='num samples per factor')
parser.add_argument('--num_samples_test',
type=int,
default=1000,
help='num samples per factor')
args = parser.parse_args()
factors = [
'ground_color', 'road_color', 'wheelDist', 'camHeight', 'camAngle',
'camFovY'
]
for factor in factors:
if not os.path.exists('representation_analysis/train_data/{}/trajectories/'
.format(factor)):
os.makedirs(
'representation_analysis/train_data/{}/trajectories/'.format(
factor))
if not os.path.exists('representation_analysis/test_data/{}/trajectories/'.
format(factor)):
os.makedirs(
# Check that the human rendering resembles the agent's view
first_obs = env.reset()
first_render = env.render('rgb_array')
m0 = first_obs.mean()
m1 = first_render.mean()
assert m0 > 0 and m0 < 255
assert abs(m0 - m1) < 5
# Try stepping a few times
for i in range(0, 10):
obs, _, _, _ = env.step(np.array([0.1, 0.099]))
# Try loading each of the available map files
for map_file in os.listdir('gym_duckietown/maps'):
map_name = map_file.split('.')[0]
env = SimpleSimEnv(map_name=map_name)
env.reset()
# Test the multi-map environment
env = MultiMapEnv()
for i in range(0, 50):
env.reset()
# Check that we do not spawn too close to obstacles
env = SimpleSimEnv(map_name='loop_obstacles')
for i in range(0, 75):
obs = env.reset()
assert not env._collision(), "collision on spawn"
env.step(np.array([0.1, 0.1]))
assert not env._collision(), "collision after one step"
for itr in range(1, num_itrs + 1):
new_actions = mutate_actions(best_actions)
positions, r = eval_actions(env, seed, new_actions)
if r > best_r:
best_r = r
best_actions = new_actions
best_positions = positions
print('iteration #%d, r=%f' % (itr, r))
print('r=%f' % best_r)
return best_positions, best_actions
env = SimpleSimEnv(map_name=args.map_name)
demos = []
total_steps = 0
for _ in range(1000):
seed = random.randint(0, 0xFFFFFFFF)
p, a = gen_trajectory(env, seed, args.demo_len + args.tail_len,
args.num_itrs)
print('trajectory length: %d' % len(p))
# If the agent did not fall off the road
if len(p) == args.demo_len + args.tail_len:
# Drop the last few actions, because the agent behaves more
# greedily in the last steps (doesn't maximize future reward)
from gym_duckietown.wrappers import HeadingWrapper
parser = argparse.ArgumentParser()
parser.add_argument('--env-name', default='SimpleSim-v0')
parser.add_argument('--map-name', default='udem1')
parser.add_argument('--draw-curve',
action='store_true',
help='draw the lane following curve')
parser.add_argument('--domain-rand',
action='store_true',
help='enable domain randomization')
args = parser.parse_args()
if args.env_name == 'SimpleSim-v0':
env = SimpleSimEnv(map_name=args.map_name,
domain_rand=args.domain_rand,
max_steps=math.inf)
else:
env = gym.make(args.env_name)
env = HeadingWrapper(env)
env.reset()
env.render()
# global variables for demo recording
positions = []
actions = []
demos = []
recording = False
from train_imitation import Model
from utils import make_var
parser = argparse.ArgumentParser()
parser.add_argument('--env-name', default='SimpleSim-v0')
parser.add_argument('--map-name', required=True)
parser.add_argument('--no-random',
action='store_true',
help='disable domain randomization')
parser.add_argument('--no-pause',
action='store_true',
help="don't pause on failure")
args = parser.parse_args()
if args.env_name == 'SimpleSim-v0':
env = SimpleSimEnv(map_name=args.map_name, domain_rand=not args.no_random)
#env.max_steps = math.inf
env.max_steps = 500
else:
env = gym.make(args.env_name)
env = HeadingWrapper(env)
obs = env.reset()
env.render()
avg_frame_time = 0
max_frame_time = 0
def load_model():
global model
#!/usr/bin/env python3
import os
import numpy as np
import gym
import gym_duckietown
from gym_duckietown.envs import SimpleSimEnv, MultiMapEnv
env = gym.make('SimpleSim-v0')
first_obs = env.reset()
# Check that the human rendering resembles the agent's view
first_render = env.render('rgb_array')
m0 = first_obs.mean()
m1 = first_render.mean()
assert m0 > 0 and m0 < 255
assert abs(m0 - m1) < 5
# Try stepping a few times
for i in range(0, 10):
obs, _, _, _ = env.step(np.array([0.1, 0.099]))
# Try loading each of the available map files
for map_file in os.listdir('gym_duckietown/maps'):
map_name = map_file.split('.')[0]
env = SimpleSimEnv(map_name=map_name)
# Test the multi-map environment
env = MultiMapEnv()