def train_model(num_steps=-1,
disp_interval=1000,
save_interval=10000,
network_name='cluster_net',
max_n=500,
batch_size=32,
delta_diff=1,
delta_max=5):
if num_steps == -1:
num_steps = np.inf
i = 1
while i < num_steps:
s1, s2, steps = bh.get_batch(batch_size, max_n)
[_, loss] = network.sess.run(
[network.train, network.loss],
feed_dict={
network.inp_state1: s1,
network.inp_state2: s2,
network.inp_step: steps,
network.inp_delta_diff: delta_diff,
network.inp_delta_max: delta_max
})
print(i, loss)
if i % disp_interval == 0:
vh.build_plot('recent', 10, batch_size=batch_size, max_n=max_n)
if i % save_interval == 0:
network.saver.save(network.sess, './' + network_name + '.ckpt')
i += 1
def build_plot(save_name, num_samples, batch_size=32, max_n=5, store_points=False):
f, ax = plt.subplots()
all_points = []
for i in tqdm.tqdm(list(range(num_samples))):
s1, s2, n = bh.get_batch(batch_size, max_n)
[e1] = network.sess.run([network.e1], feed_dict={network.inp_state1: s1})
all_points.append(e1)
ax.scatter(e1[:, 0], e1[:, 1])
if store_points:
all_points = np.concatenate(all_points, axis=0)
np.save('./' + save_name + '.npy', all_points)
f.savefig('./' + save_name+ '.png')
def train_vae(num_steps=-1,
batch_size=32,
disp_interval=100,
save_interval=10000,
network_name='vae_net'):
import network as network
if num_steps == -1:
num_steps = np.inf
i = 1
while i < num_steps:
images = np.reshape(bh.get_batch(batch_size), [-1, 84, 84, 1])
[_, loss, recon
] = network.sess.run([network.train_op, network.loss, network.mu_x],
feed_dict={network.inp_image: images / 255.})
print i, loss
if i % disp_interval == 0:
vh.display_side_by_side('recent', images[0], recon[0])
if i % save_interval == 0:
network.saver.save(network.sess, './' + network_name + '.ckpt')
i += 1