diff --git a/CONTRIBUTING.rst b/CONTRIBUTING.rst
index f552f7626cfffca9a4422cb9337c9cdc27d83867..7ae26bcc3d4e0f4a5cbbcfafd055e2c084a42345 100644
--- a/CONTRIBUTING.rst
+++ b/CONTRIBUTING.rst
@@ -79,7 +79,7 @@ Ready to contribute? Here's how to set up `flatland` for local development.
5. When you're done making changes, check that your changes pass flake8 and the
tests, including testing other Python versions with tox::
- $ flake8 flatland tests
+ $ flake8 flatland tests examples
$ python setup.py test or py.test
$ tox
@@ -125,4 +125,4 @@ $ bumpversion patch # possible: major / minor / patch
$ git push
$ git push --tags
-Travis will then deploy to PyPI if tests pass. (To be configured properly by Mohanty)
\ No newline at end of file
+Travis will then deploy to PyPI if tests pass. (To be configured properly by Mohanty)
diff --git a/Makefile b/Makefile
index 691bf84c768aee6794420c467b0387390dec6218..6a655bb9f96397d09ced973bf6f6ccab20d4e42e 100644
--- a/Makefile
+++ b/Makefile
@@ -51,7 +51,7 @@ clean-test: ## remove test and coverage artifacts
rm -fr .pytest_cache
lint: ## check style with flake8
- flake8 flatland tests
+ flake8 flatland tests examples
test: ## run tests quickly with the default Python
py.test
diff --git a/examples/play_model.py b/examples/play_model.py
index 62726c24c96be0e5dae2f4840e18da452163b7ac..174568177a4a886cfe38e53125d0f73f2dae52de 100644
--- a/examples/play_model.py
+++ b/examples/play_model.py
@@ -1,12 +1,14 @@
-from flatland.envs.rail_env import RailEnv
-from flatland.envs.generators import complex_rail_generator
-from flatland.utils.rendertools import RenderTool
-from flatland.baselines.dueling_double_dqn import Agent
-from collections import deque
-import torch
import random
-import numpy as np
import time
+from collections import deque
+
+import numpy as np
+import torch
+
+from flatland.baselines.dueling_double_dqn import Agent
+from flatland.envs.generators import complex_rail_generator
+from flatland.envs.rail_env import RailEnv
+from flatland.utils.rendertools import RenderTool
class Player(object):
@@ -25,7 +27,7 @@ class Player(object):
self.done_window = deque(maxlen=100)
self.scores = []
self.dones_list = []
- self.action_prob = [0]*4
+ self.action_prob = [0] * 4
self.agent = Agent(self.state_size, self.action_size, "FC", 0)
# self.agent.qnetwork_local.load_state_dict(torch.load('../flatland/baselines/Nets/avoid_checkpoint9900.pth'))
self.agent.qnetwork_local.load_state_dict(torch.load(
@@ -33,7 +35,7 @@ class Player(object):
self.iFrame = 0
self.tStart = time.time()
-
+
# Reset environment
# self.obs = self.env.reset()
self.env.obs_builder.reset()
@@ -51,7 +53,6 @@ class Player(object):
env = self.env
# Pass the (stored) observation to the agent network and retrieve the action
- #for handle in env.get_agent_handles():
for handle in env.get_agent_handles():
action = self.agent.act(np.array(self.obs[handle]), eps=self.eps)
self.action_prob[action] += 1
@@ -68,8 +69,8 @@ class Player(object):
# Update replay buffer and train agent
for handle in self.env.get_agent_handles():
self.agent.step(self.obs[handle], self.action_dict[handle],
- all_rewards[handle], next_obs[handle], done[handle],
- train=False)
+ all_rewards[handle], next_obs[handle], done[handle],
+ train=False)
self.score += all_rewards[handle]
self.iFrame += 1
@@ -85,7 +86,7 @@ def max_lt(seq, val):
None is returned if seq was empty or all items in seq were >= val.
"""
- idx = len(seq)-1
+ idx = len(seq) - 1
while idx >= 0:
if seq[idx] < val and seq[idx] >= 0:
return seq[idx]
@@ -94,7 +95,6 @@ def max_lt(seq, val):
def main(render=True, delay=0.0):
-
random.seed(1)
np.random.seed(1)
@@ -118,8 +118,9 @@ def main(render=True, delay=0.0):
done_window = deque(maxlen=100)
scores = []
dones_list = []
- action_prob = [0]*4
+ action_prob = [0] * 4
agent = Agent(state_size, action_size, "FC", 0)
+
# agent.qnetwork_local.load_state_dict(torch.load('../flatland/baselines/Nets/avoid_checkpoint9900.pth'))
def max_lt(seq, val):
@@ -128,7 +129,7 @@ def main(render=True, delay=0.0):
None is returned if seq was empty or all items in seq were >= val.
"""
- idx = len(seq)-1
+ idx = len(seq) - 1
while idx >= 0:
if seq[idx] < val and seq[idx] >= 0:
return seq[idx]
@@ -141,7 +142,8 @@ def main(render=True, delay=0.0):
# Reset environment
obs = env.reset()
- env_renderer.set_new_rail()
+ if render:
+ env_renderer.set_new_rail()
for a in range(env.get_num_agents()):
norm = max(1, max_lt(obs[a], np.inf))
@@ -165,7 +167,6 @@ def main(render=True, delay=0.0):
if render:
env_renderer.renderEnv(show=True, frames=True, iEpisode=trials, iStep=step, action_dict=action_dict)
- #time.sleep(10)
if delay > 0:
time.sleep(delay)
@@ -181,7 +182,6 @@ def main(render=True, delay=0.0):
agent.step(obs[a], action_dict[a], all_rewards[a], next_obs[a], done[a])
score += all_rewards[a]
-
obs = next_obs.copy()
if done['__all__']:
env_done = 1
@@ -196,25 +196,25 @@ def main(render=True, delay=0.0):
print(('\rTraining {} Agents.\tEpisode {}\tAverage Score: {:.0f}\tDones: {:.2f}%' +
'\tEpsilon: {:.2f} \t Action Probabilities: \t {}').format(
- env.get_num_agents(),
- trials,
- np.mean(scores_window),
- 100 * np.mean(done_window),
- eps, action_prob/np.sum(action_prob)),
+ env.get_num_agents(),
+ trials,
+ np.mean(scores_window),
+ 100 * np.mean(done_window),
+ eps, action_prob / np.sum(action_prob)),
end=" ")
if trials % 100 == 0:
tNow = time.time()
rFps = iFrame / (tNow - tStart)
print(('\rTraining {} Agents.\tEpisode {}\tAverage Score: {:.0f}\tDones: {:.2f}%' +
'\tEpsilon: {:.2f} fps: {:.2f} \t Action Probabilities: \t {}').format(
- env.get_num_agents(),
- trials,
- np.mean(scores_window),
- 100 * np.mean(done_window),
- eps, rFps, action_prob / np.sum(action_prob)))
+ env.get_num_agents(),
+ trials,
+ np.mean(scores_window),
+ 100 * np.mean(done_window),
+ eps, rFps, action_prob / np.sum(action_prob)))
torch.save(agent.qnetwork_local.state_dict(),
- '../flatland/baselines/Nets/avoid_checkpoint' + str(trials) + '.pth')
- action_prob = [1]*4
+ '../flatland/baselines/Nets/avoid_checkpoint' + str(trials) + '.pth')
+ action_prob = [1] * 4
if __name__ == "__main__":
diff --git a/examples/qt2.py b/examples/qt2.py
index 6074106523c7fbe503f79b6bc7604f055dc27b35..ee3ea0cd123a3e6d0b1fc970eced219ed8503203 100644
--- a/examples/qt2.py
+++ b/examples/qt2.py
@@ -1,9 +1,8 @@
-
-
import sys
+
from PyQt5 import QtSvg
-from PyQt5.QtWidgets import QApplication, QLabel, QMainWindow, QGridLayout, QWidget
from PyQt5.QtCore import Qt, QByteArray
+from PyQt5.QtWidgets import QApplication, QLabel, QMainWindow, QGridLayout, QWidget
from flatland.utils import svg
@@ -75,4 +74,3 @@ window = MainWindow()
window.show()
app.exec_()
-
diff --git a/examples/temporary_example.py b/examples/temporary_example.py
index 1f3504f221d59d0205974bb135c2237364a22e07..db909e00e1f1a2c1fdef73ba2a35159a698832a3 100644
--- a/examples/temporary_example.py
+++ b/examples/temporary_example.py
@@ -1,11 +1,10 @@
import random
+
import numpy as np
-import matplotlib.pyplot as plt
-from flatland.envs.rail_env import *
-from flatland.envs.generators import *
-from flatland.envs.observations import TreeObsForRailEnv
-from flatland.utils.rendertools import *
+from flatland.envs.generators import random_rail_generator
+from flatland.envs.rail_env import RailEnv
+from flatland.utils.rendertools import RenderTool
random.seed(0)
np.random.seed(0)
@@ -94,7 +93,7 @@ env = RailEnv(width=7,
# print(env.obs_builder.distance_map[0, :, :, i])
# Print the observation vector for agent 0
-obs, all_rewards, done, _ = env.step({0:0})
+obs, all_rewards, done, _ = env.step({0: 0})
for i in range(env.get_num_agents()):
env.obs_builder.util_print_obs_subtree(tree=obs[i], num_features_per_node=5)
@@ -113,6 +112,7 @@ for step in range(100):
while i < len(cmds):
if cmds[i] == 'q':
import sys
+
sys.exit()
elif cmds[i] == 's':
obs, all_rewards, done, _ = env.step(action_dict)
@@ -120,9 +120,9 @@ for step in range(100):
print("Rewards: ", all_rewards, " [done=", done, "]")
else:
agent_id = int(cmds[i])
- action = int(cmds[i+1])
+ action = int(cmds[i + 1])
action_dict[agent_id] = action
- i = i+1
+ i = i + 1
i += 1
env_renderer.renderEnv(show=True)
diff --git a/examples/training_navigation.py b/examples/training_navigation.py
index cabb655e3eb2bc2d12d908559b0102b072163052..85f9531b8820139e5559081feee4a93c4e01ac6c 100644
--- a/examples/training_navigation.py
+++ b/examples/training_navigation.py
@@ -1,11 +1,15 @@
-from flatland.envs.rail_env import *
-from flatland.envs.generators import *
-from flatland.envs.observations import TreeObsForRailEnv
-from flatland.utils.rendertools import *
-from flatland.baselines.dueling_double_dqn import Agent
-from collections import deque
-import torch, random
+import random
import time
+from collections import deque
+
+import numpy as np
+import torch
+
+from flatland.baselines.dueling_double_dqn import Agent
+from flatland.envs.generators import complex_rail_generator
+from flatland.envs.rail_env import RailEnv
+from flatland.utils.rendertools import RenderTool
+
random.seed(1)
np.random.seed(1)
@@ -190,25 +194,34 @@ for trials in range(1, n_trials + 1):
dones_list.append((np.mean(done_window)))
print(
- '\rTraining {} Agents.\tEpisode {}\tAverage Score: {:.0f}\tDones: {:.2f}%\tEpsilon: {:.2f} \t Action Probabilities: \t {}'.format(
+ '\rTraining {} Agents.\t' +
+ 'Episode {}\t' +
+ 'Average Score: {:.0f}\t' +
+ 'Dones: {:.2f}%\t' +
+ 'Epsilon: {:.2f} \t ' +
+ 'Action Probabilities: \t ' +
+ '{}'.format(
env.get_num_agents(),
trials,
- np.mean(
- scores_window),
- 100 * np.mean(
- done_window),
+ np.mean(scores_window),
+ 100 * np.mean(done_window),
eps, action_prob / np.sum(action_prob)),
end=" ")
if trials % 100 == 0:
print(
- '\rTraining {} Agents.\tEpisode {}\tAverage Score: {:.0f}\tDones: {:.2f}%\tEpsilon: {:.2f} \t Action Probabilities: \t {}'.format(
+ '\rTraining {} Agents.\t' +
+ 'Episode {}\t' +
+ 'Average Score: {:.0f}\t' +
+ 'Dones: {:.2f}%\t' +
+ 'Epsilon: {:.2f} \t ' +
+ 'Action Probabilities: \t ' +
+ '{}'.format(
env.get_num_agents(),
trials,
- np.mean(
- scores_window),
- 100 * np.mean(
- done_window),
- eps, action_prob / np.sum(action_prob)))
+ np.mean(scores_window),
+ 100 * np.mean(done_window),
+ eps,
+ action_prob / np.sum(action_prob)))
torch.save(agent.qnetwork_local.state_dict(),
'../flatland/baselines/Nets/avoid_checkpoint' + str(trials) + '.pth')
action_prob = [1] * 4
diff --git a/flatland/envs/observations.py b/flatland/envs/observations.py
index 8cddec0151f691339ace3b367b888878eabe0b07..f90b529d51b66a353ba9ee71d77497b277556494 100644
--- a/flatland/envs/observations.py
+++ b/flatland/envs/observations.py
@@ -203,7 +203,7 @@ class TreeObsForRailEnv(ObservationBuilder):
num_transitions = np.count_nonzero(possible_transitions)
# Root node - current position
# observation = [0, 0, 0, 0, self.distance_map[handle, position[0], position[1], orientation]]
- observation = [0, 0, 0, 0, self.distance_map[(handle, *agent.position, agent.direction)]]
+ observation = [0, 0, 0, 0, self.distance_map[(handle, *agent.position,direc agent.direction)]]
root_observation = observation[:]
visited = set()
# Start from the current orientation, and see which transitions are available;
@@ -478,11 +478,13 @@ class GlobalObsForRailEnv(ObservationBuilder):
- transition map array with dimensions (env.height, env.width, 16),
assuming 16 bits encoding of transitions.
- - Four 2D arrays containing respectively the position of the given agent,
- the position of its target, the positions of the other agents and of
- their target.
+ - Three 2D arrays (map_height, map_width, 3) containing respectively the position of the given agent,
+ the position of its target and the positions of the other agents targets.
- - A 4 elements array with one hot encoding of the direction.
+ - A 3D array (map_height, map_width, 4) containing the one hot encoding of directions
+ of the other agents at their position coordinates.
+
+ - A 4 elements array with one of encoding of the direction of the agent of interest.
"""
def __init__(self):
@@ -492,32 +494,37 @@ class GlobalObsForRailEnv(ObservationBuilder):
self.rail_obs = np.zeros((self.env.height, self.env.width, 16))
for i in range(self.rail_obs.shape[0]):
for j in range(self.rail_obs.shape[1]):
- self.rail_obs[i, j] = np.array(
- list(f'{self.env.rail.get_transitions((i, j)):016b}')).astype(int)
+ bitlist = [int(digit) for digit in bin(self.env.rail.get_transitions((i, j)))[2:]]
+ bitlist = [0] * (16 - len(bitlist)) + bitlist
+ self.rail_obs[i, j] = np.array(bitlist)
+ # self.rail_obs[i, j] = np.array(
+ # list(f'{self.env.rail.get_transitions((i, j)):016b}')).astype(int)
# self.targets = np.zeros(self.env.height, self.env.width)
# for target_pos in self.env.agents_target:
# self.targets[target_pos] += 1
def get(self, handle):
- obs = np.zeros((4, self.env.height, self.env.width))
+ obs_map_state = np.zeros((self.env.height, self.env.width, 3))
+ obs_other_agents_state = np.zeros((self.env.height, self.env.width, 4))
agents = self.env.agents
agent = agents[handle]
agent_pos = agents[handle].position
- obs[0][agent_pos] += 1
- obs[1][agent.target] += 1
+ obs_map_state[agent_pos][0] += 1
+ obs_map_state[agent.target][1] += 1
for i in range(len(agents)):
if i != handle: # TODO: handle used as index...?
agent2 = agents[i]
- obs[3][agent2.position] += 1
- obs[2][agent2.target] += 1
+ obs_other_agents_state[agent2.position][agent2.direction] = 1
+ obs_map_state[agent2.target][2] += 1
direction = np.zeros(4)
direction[agent.direction] = 1
- return self.rail_obs, obs, direction
+ return self.rail_obs, obs_map_state, obs_other_agents_state, direction
+
class LocalObsForRailEnv(ObservationBuilder):