def go(self):
plotter = Plotter("outdir", "sim", clean=False)
grid = Grid(plotter)
for num_of_agents in poisson.rvs(5, size=42):
agents = [Agent.build() for i in xrange(num_of_agents)]
# print agents
print "\nStep {0}\n".format(self.step)
grid.push(agents)
grid.tick()
self.step += 1
def data_func(net, device, train_queue):
# Function to run in sub processes
envs = [make_env() for _ in range(ENV_COUNT)]
agent = Agent(net=net, device=device)
exp_source = ExperienceSourceFirstLast(envs, agent, gamma=GAMMA, steps_count=STEPS_COUNT)
print(f'{mp.current_process().name} Started')
for exp in exp_source:
new_rewards = exp_source.pop_rewards_steps()
if new_rewards:
train_queue.put(RewardSteps(reward=new_rewards))
train_queue.put(exp)
def data_func(net, device, train_queue):
envs = [make_env() for _ in range(ENV_COUNT)]
agent = Agent(lambda x: net(x)[0], device=device, apply_softmax=True)
exp_source = ExperienceSourceFirstLast(envs,
agent,
gamma=GAMMA,
steps_count=STEPS_COUNT)
print(f'{mp.current_process().name} Started')
for exp in exp_source:
new_rewards = exp_source.pop_total_rewards()
if new_rewards:
train_queue.put(TotalReward(reward=np.mean(new_rewards)))
train_queue.put(exp)
def data_func(net, device, train_queue):
envs = [make_env() for _ in range(NUM_ENVS)]
agent = Agent(net, device=device)
# TODO compare training with rgb to semantic and other variations
exp_source = ExperienceSourceFirstLast(envs,
agent,
gamma=GAMMA,
steps_count=STEPS_COUNT)
print(f'{mp.current_process().name} Started')
for exp in exp_source:
new_rewards = exp_source.pop_total_rewards()
# print('New rewards', new_rewards)
if new_rewards:
train_queue.put(TotalReward(reward=np.mean(new_rewards)))
# print('Pop goes the weasel!')
train_queue.put(exp)
def main():
"""
After training is done we test and watch our model perform
"""
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
net = torch.load(LOAD_MODEL)
net = net.to(device)
net.eval() # set model into evaluation mode
agent = Agent(net, device=device)
print('Model loaded from:', LOAD_MODEL)
spec = gym.envs.registry.spec(ENV_NAME)
spec._kwargs['render'] = True
env = make_env()
state = [env.reset()]
rewards_history, current_rewards = [], []
episode_counter = 0
try:
while True:
env.render()
time.sleep(1 / FRAME_RATE)
if state is not None:
action = agent(state)
else:
action = env.action_space.sample(
) # if obs is none, sample random action
next_state, reward, done, _ = env.step(action[0])
state = [next_state]
current_rewards.append(reward)
if done:
episode_counter += 1
print('Episode', episode_counter,
'Done.') # Mean Reward:', np.mean(current_rewards))
rewards_history.append(current_rewards)
current_rewards.clear()
print('Starting Next Episode...')
state = [env.reset()]
except KeyboardInterrupt:
print('Stopped By The User')
print('Exiting...')
def main():
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
net = torch.load(LOAD_MODEL)
net = net.to(device)
net.eval() # set model into evaluation mode
agent = Agent(lambda x: net(x)[0], device=device, apply_softmax=True)
print('Model loaded from:', LOAD_MODEL)
if FRAME_RATE > 0:
print(f'Playing at {FRAME_RATE} FPS')
else:
print('Playing with no framerate limitations')
env = make_env()
state = [env.reset()]
rewards_history, current_rewards = [], []
game_counter = 0
try:
while True:
env.render()
if FRAME_RATE > 0:
time.sleep(1 / FRAME_RATE)
if state is not None:
action = agent(state)
else:
action = env.action_space.sample(
) # if obs is none, sample random action
next_state, reward, done, _ = env.step(action)
state = [next_state]
current_rewards.append(reward)
if done:
game_counter += 1
print('Game', game_counter,
'Done.') # Mean Reward:', np.mean(current_rewards))
rewards_history.append(current_rewards)
current_rewards.clear()
print('Starting Next Game...')
state = [env.reset()]
except KeyboardInterrupt:
print('Stopped By The User')
print('Exiting...')
def main():
"""
After training is done we test and watch our model perform
"""
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
net = torch.load(LOAD_MODEL)
net = net.to(device)
net.eval() # set model into evaluation mode
agent = Agent(net, device=device)
print('Model loaded from:', LOAD_MODEL)
env = make_env()
state = [env.reset()]
rewards_history, current_rewards = [], []
episode_counter = 0
try:
while True:
time.sleep(1/FRAME_RATE)
if state is not None:
action = agent(state)
else:
action = env.action_space.sample() # if obs is none, sample random action
next_state, reward, done, _ = env.step(action[0])
state = [next_state]
current_rewards.append(reward)
if done:
episode_counter += 1
rewards_history.append(np.mean(current_rewards))
print('Episode', episode_counter, 'Done, mean reward', np.mean(rewards_history[-100:])) # Mean Reward:', np.mean(current_rewards))
current_rewards.clear()
print('Starting Next Episode...')
state = [env.reset()]
except KeyboardInterrupt:
env.close()
print('Stopped By The User')
print('Exiting...')
client = carla.Client('localhost', 2000)
client.reload_world()
def test_would_to_walking_with_zero_twait(self):
fakeAgent = Agent()
#t(0)
self.assertEqual(fakeAgent.switching_walking, True)
self.assertEqual(fakeAgent.switching_waiting, False)
fakeAgent.pstop = 1
self.assertEqual(fakeAgent.would_to_walking(), True)
#t(1)
self.assertEqual(fakeAgent.switching_walking, False)
self.assertEqual(fakeAgent.switching_waiting, False)
fakeAgent.pstop = -1
fakeAgent._twait = 0
self.assertEqual(fakeAgent.would_to_walking(), False)
#t(2) sleep = 0
self.assertEqual(fakeAgent.switching_walking, True)
self.assertEqual(fakeAgent.switching_waiting, True)
fakeAgent.pstop = 1
self.assertEqual(fakeAgent.would_to_walking(), True)
#t(3) sleep = 0
self.assertEqual(fakeAgent.switching_walking, False)
self.assertEqual(fakeAgent.switching_waiting, False)
self.assertEqual(fakeAgent.would_to_walking(), True)
#t(4)
self.assertEqual(fakeAgent.switching_walking, False)
self.assertEqual(fakeAgent.switching_waiting, False)
fakeAgent.pstop = -1
self.assertEqual(fakeAgent.would_to_walking(), False)
#t(5)
self.assertEqual(fakeAgent.switching_walking, True)
self.assertEqual(fakeAgent.switching_waiting, True)
self.assertEqual(fakeAgent.would_to_walking(), False)
#t(6)
self.assertEqual(fakeAgent.switching_walking, True)
self.assertEqual(fakeAgent.switching_waiting, True)
fakeAgent.pstop = 1
self.assertEqual(fakeAgent.would_to_walking(), True)
self.assertEqual(fakeAgent.switching_walking, False)
self.assertEqual(fakeAgent.switching_waiting, False)
BATCH_SIZE = 32
TARGET_UPDATE_FREQ = 1000
DELAY_LEARNING = 50000
GAMMA = 0.99
model = 'DDQN'
# Environment and neural networks
env = make_env("PongNoFrameskip-v4")
env_test = gym.make("PongNoFrameskip-v4")
net = DQN(env.observation_space.shape, env.action_space.n, learning_rate).to(device)
target_net = DQN(env.observation_space.shape, env.action_space.n, learning_rate).to(device)
# Agent and memory handling
memory = Memory(REPLAY_SIZE)
agent = Agent(env, memory)
initial_observation = env.reset()
if 'cuda' in str(device):
print('The GPU is being used')
else:
print('The CPU is being used')
if option_dict['random']:
play_random(env, UP_ACTION, DOWN_ACTION, seconds=5)
if option_dict['train']:
print("Training")
from lib.Arena import Arena from lib.Agent import Agent from random import randint arena = Arena(5) hero = Agent(arena) arena.printEverything() points = hero.play()
