def main():
log_dir = '../../drl_for_surveillance_runs/flood/indiv/drqn/pretrained/'
pretrained_model_name = 'drqn-20_model.pkl'
env = SurveillanceEnv(nr_agents=2,
obs_mode='normal',
obs_type='flood',
obs_radius=500,
world_size=1000,
grid_size=100,
range_cutpoints=30,
angular_cutpoints=40,
torus=False,
dynamics='aircraft',
shared_reward=False,
render_dir=log_dir + 'video/')
policy = StatePlusImageLstmPolicy
model = DRQN.load('../../' + 'pretrained_models/' + pretrained_model_name,
env=env)
model.tensorboard_log = log_dir
model.checkpoint_path = log_dir + 'models/'
model.initial_exploration = 0.5
train(model, callback, num_timesteps=int(2e6), log_dir=log_dir)
def main():
log_dir = '../../../../Results/models/wildfire/indiv/drqn/obs_100/'
saved_model_name = 'checkpoint_model.pkl'
config = tf.ConfigProto(log_device_placement=True)
env = SurveillanceEnv(
nr_agents=2,
obs_mode='normal',
obs_type='wildfire',
obs_radius=100, # CHECK
world_size=1000,
grid_size=100,
range_cutpoints=30,
angular_cutpoints=40,
torus=False,
dynamics='aircraft',
shared_reward=False,
render_dir=log_dir + 'video/')
model = DRQN.load(log_dir + 'models/' + saved_model_name, env=env)
params = find_trainable_variables("main")
params_val = model.sess.run(
model.graph._collections['trainable_variables'][6])
import matplotlib.pyplot as plt
import numpy as np
plt.hist(params_val.flatten(), 40, alpha=0.3)
plt.show()
print('')
def main():
log_dir = '../../drl_for_surveillance_runs/flood/indiv/drqn/v1/'
env = SurveillanceEnv(nr_agents=2,
obs_mode='normal',
obs_type='flood',
obs_radius=500,
world_size=1000,
grid_size=100,
range_cutpoints=30,
angular_cutpoints=40,
torus=False,
dynamics='aircraft',
shared_reward=False,
render_dir=log_dir + 'video/')
policy = StatePlusImageLstmPolicy
model = DRQN(policy,
env,
prioritized_replay=False,
verbose=1,
tensorboard_log=log_dir,
batch_size=100,
target_network_update_freq=100,
learning_starts=15000,
train_freq=10,
trace_length=20,
buffer_size=200000,
exploration_fraction=0.7,
exploration_final_eps=0.1,
checkpoint_path=log_dir + 'models/',
checkpoint_freq=10000)
train(model, callback, num_timesteps=int(3e6), log_dir=log_dir)
def main():
log_dir = '../../drl_for_surveillance_runs/wildfire/indiv/drqn/v4/'
saved_model_name = 'checkpoint_model.pkl'
env = SurveillanceEnv(nr_agents=2,
obs_mode='normal',
obs_type='wildfire',
obs_radius=500,
world_size=1000,
grid_size=100,
range_cutpoints=30,
angular_cutpoints=40,
torus=False,
dynamics='aircraft',
shared_reward=False,
render_dir=log_dir + 'video/')
model = DRQN.load(log_dir + 'models/' + saved_model_name, env=env)
obs = env.reset()
states = None
while True:
actions, states = model.predict(obs, states=states)
obs, rewards, dones, info = env.step(actions)
if env.timestep % 100 == 0:
env.render()
if dones:
obs = env.reset()
states = None
def main():
log_dir = '../../drl_for_surveillance_runs/flood/terrain/'
env = SurveillanceEnv(nr_agents=2,
obs_mode='normal',
obs_type='flood',
obs_radius=500,
world_size=1000,
grid_size=100,
range_cutpoints=30,
angular_cutpoints=40,
torus=False,
dynamics='aircraft',
shared_reward=False,
render_dir=log_dir + 'video/')
env.reset()
while True:
env.world.reset()
def main():
log_dir = '../../drl_for_surveillance_runs/wildfire/indiv/dqn/report/'
n_agents = 2
env = SurveillanceEnv(nr_agents=n_agents,
obs_mode='normal',
obs_type='wildfire',
obs_radius=500,
world_size=1000,
grid_size=100,
range_cutpoints=30,
angular_cutpoints=40,
torus=False,
dynamics='aircraft',
shared_reward=False,
render_dir=log_dir + 'video/')
while True:
obs, done = env.reset(), False
episode_rew = 0
timestep = 0
while True:
if timestep % 500 == 0:
time = datetime.datetime.now().strftime("%Y_%m_%d-%H_%M_%S-")
env.render(render_type='wildfire_state',
output_dir=log_dir + 'video/',
name_append=time + format(timestep, '04d') + '_')
action = np.zeros(n_agents)
obs, rew, done, _ = env.step(action)
episode_rew += rew
timestep += 1
print("Episode reward", episode_rew)
def main():
log_dir = '../../drl_for_surveillance_runs/wildfire/indiv/drqn/test/'
env = SurveillanceEnv(nr_agents=2,
obs_mode='normal',
obs_type='wildfire',
obs_radius=500,
world_size=1000,
grid_size=100,
range_cutpoints=30,
angular_cutpoints=40,
torus=False,
dynamics='aircraft',
shared_reward=False,
render_dir=log_dir + 'video/')
policy = StatePlusImageLstmPolicy
model = DRQN(policy, env, prioritized_replay=False,
verbose=1,
tensorboard_log=log_dir,
batch_size=100,
target_network_update_freq=100,
learning_starts=10000,
train_freq=10,
trace_length=20,
h_size=100,
learning_rate=5e-4,
buffer_size=200000,
exploration_fraction=0.7,
exploration_final_eps=0.1,
checkpoint_path=log_dir + 'models/',
checkpoint_freq=10000)
params_val = model.sess.run(model.graph._collections['trainable_variables'][22])
import matplotlib.pyplot as plt
import numpy as np
plt.hist(params_val.flatten(), 40, alpha=0.3)
plt.show()
print('')
def main():
log_dir = '../../../models/wildfire/indiv/dqn/'
saved_model_name = 'checkpoint_model.pkl'
n_agents = 2
env = SurveillanceEnv(nr_agents=n_agents,
obs_mode='normal',
obs_type='wildfire',
obs_radius=500,
world_size=1000,
grid_size=100,
range_cutpoints=30,
angular_cutpoints=40,
torus=False,
dynamics='aircraft',
shared_reward=False,
render_dir=log_dir + 'video/')
# All tests share the same seed
env.seed(seed=1)
np.random.seed(seed=1)
model = DQN.load(log_dir + 'models/' + saved_model_name, env=env)
total_episodes = 100
episode_rewards = []
for ep in range(total_episodes):
obs = env.reset()
states = None
episode_rewards.append(0.0)
while True:
actions, states = model.predict(obs, states=states)
obs, rewards, dones, info = env.step(actions)
for i in range(n_agents):
# Sum of rewards for logging
episode_rewards[-1] += rewards[i]
if env.timestep % 500 == 0:
print("Timestep " + str(env.timestep) + "/" +
str(env.timestep_limit))
algo = model
time = datetime.datetime.now().strftime("%Y_%m_%d-%H_%M_%S")
# Render map
if (algo.env.is_terminal or algo.env.timestep % 500 == 0) \
and algo.env.timestep > 0:
n = 1 # agent to plot
dim_image = algo.observation_space.dim_image
obs_array = np.array(obs[n])
obs_image = np.reshape(obs_array[:dim_image[0] * dim_image[1]],
dim_image,
order='C')
if algo.env.ax is not None:
x = algo.env.agents[n].obs_angles
y = algo.env.agents[n].obs_ranges
ax1 = algo.env.ax[1]
ax1.set_theta_zero_location("S")
x_mod = np.append(x, 2 * np.pi)
obs_image_append = np.expand_dims(obs_image[:, 0], axis=1)
obs_image_mod = np.append(obs_image,
obs_image_append,
axis=1)
ax1.pcolormesh(x_mod, y, obs_image_mod)
render_dir = algo.env.render_dir
algo.env.render(mode='animate',
output_dir=render_dir + time + '-',
name_append=format(algo.num_timesteps, '04d') +
'_')
print(time)
print("Instantaneous rewards:", rewards)
if dones:
print("Episode", ep + 1, "reward:", episode_rewards[-1])
break
np.save(log_dir + 'test_' + str(total_episodes) + '_episode_rewards',
episode_rewards)