def plot_variance_vs_position(result_dir='variance_vs_position',
network_name='vae_net',
var_max=5.0):
import network as network
if not os.path.isdir(result_dir):
os.mkdir(result_dir)
game = wind_tunnel.WindTunnel()
encodings = []
recons = []
images = []
losses = []
network.saver.restore(network.sess, './' + network_name + '.ckpt')
for i in tqdm.tqdm(range(100)):
game.agent = i
game.generate_new_state()
image = np.reshape(np.array(game.get_current_state()), [-1, 84, 84, 1])
images.append(image)
[encoding_mean, recon,
loss] = network.sess.run([network.mu_z, network.mu_x, network.loss],
feed_dict={network.inp_image: image / 255.})
encodings.append(encoding_mean)
losses.append(loss)
recons.append(recon)
losses = np.array(losses)
losses = losses - np.min(losses)
losses = losses / np.max(losses)
encodings = np.concatenate(encodings, 0)
variances = np.sqrt(np.sum((encodings**2), axis=1))
novelty = var_max * losses + variances * (1. - losses)
f, ax = plt.subplots()
ax.plot(range(100), variances, 'blue')
ax.plot(range(100), losses, 'red')
ax.plot(range(100), novelty, 'green')
f.savefig(os.path.join(result_dir, 'variance_plot.png'))
side_by_side_dir = os.path.join(result_dir, 'reconstructions')
if not os.path.isdir(side_by_side_dir):
os.mkdir(side_by_side_dir)
for i in tqdm.tqdm(range(100)):
vh.display_side_by_side(os.path.join(side_by_side_dir, '%5d' % i),
images[i][0], recons[i][0])
def setup_wind_tunnel_env():
env = wind_tunnel.WindTunnel()
num_actions = len(env.get_actions_for_state(None))
return env, num_actions
import numpy as np
import wind_tunnel
import toy_mr
game = wind_tunnel.WindTunnel()
class RandomAgent(object):
def get_action(self, game):
actions = game.get_actions_for_state(None)
return np.random.choice(actions)
def get_batch_func(batch_size, game, agent):
states = []
for i in range(batch_size):
if game.is_current_state_terminal():
game.reset_environment()
action = agent.get_action(game)
game.perform_action(action)
states.append(game.get_current_state()[0])
return np.array(states)
def setup_toy_mr_env():
env = toy_mr.ToyMR('../mr_maps/full_mr_map.txt')
num_actions = len(env.get_actions_for_state(None))
return env, num_actions
toy_mr_env, toy_mr_num_actions = setup_toy_mr_env()
