def callback(itr):
def samplefun(num_samps):
import numpy as np
z = np.array(np.random.randn(num_samps, zdim), dtype=np.float32)
return decode(z).eval(session=sess)
viz.plot_samples(itr, samplefun, savedir='avae_mnist_samples')
def sample_normal(mu, log_sigmasq, M=5):
dim = mu.get_shape()[1].value
eps = tf.random_normal((M, dim), dtype=tf.float32)
return mu + tf.exp(0.5 * log_sigmasq) * eps
def recons(num_samps):
# choose one datapoint, encode it
subset = X[np.random.choice(X.shape[0], 1)]
# compute encoder - first aux, then encoder conditioned on aux
amu, alog_sigmasq = aux_encode(subset)
asamps = sample_normal(amu, alog_sigmasq, M=24)
Xa = tf.concat(1, [tf.tile(subset, [24, 1]), asamps])
zmu, zlog_sigmasq = encode(Xa)
imgs = decode(sample_normal(zmu, zlog_sigmasq,
M=24)).eval(session=sess)
return np.row_stack([subset, imgs])
viz.plot_samples(itr, recons, savedir='avae_mnist_samples', stub='recon')
test_lb = test_lb_fun.eval(session=sess) * Ntest
print "test data VLB: ", np.mean(test_lb)
def callback(itr):
# plot samples from the model generative process
viz.plot_samples(itr, sample_fun, savedir=output_dir)
# plot reconstructed samples
viz.plot_samples(itr, recon_fun, savedir=output_dir, stub='recon')
# evaluate test
test_lb = test_vlb_fun.eval(session=sess) * Ntest
print "test data VLB: ", np.mean(test_lb)
# plot z space for each test data item, colored by digit type
def plot_test_latent():
z = latent_space_fun()
fig = plt.figure()
ax = plt.gca()
for i in xrange(10):
idx = np.argmax(Yt, 1) == i
ax.scatter(z[idx, 0], z[idx, 1], c=colors[i], label='%d' % i)
ax.legend()
fig.savefig(os.path.join(output_dir,
"latent_space_%03d.png" % itr))
plot_test_latent()
def callback(itr):
def samplefun(num_samps):
import numpy as np
z = np.array(np.random.randn(num_samps, zdim), dtype=np.float32)
return decode(z).eval(session=sess)
viz.plot_samples(itr, samplefun, savedir='avae_mnist_samples')
def sample_normal(mu, log_sigmasq, M=5):
dim = mu.get_shape()[1].value
eps = tf.random_normal((M, dim), dtype=tf.float32)
return mu + tf.exp(0.5 * log_sigmasq) * eps
def recons(num_samps):
# choose one datapoint, encode it
subset = X[np.random.choice(X.shape[0], 1)]
# compute encoder - first aux, then encoder conditioned on aux
amu, alog_sigmasq = aux_encode(subset)
asamps = sample_normal(amu, alog_sigmasq, M=24)
Xa = tf.concat(1, [tf.tile(subset, [24, 1]), asamps])
zmu, zlog_sigmasq = encode(Xa)
imgs = decode(sample_normal(zmu, zlog_sigmasq, M=24)).eval(session=sess)
return np.row_stack([subset, imgs])
viz.plot_samples(itr, recons, savedir='avae_mnist_samples', stub='recon')
test_lb = test_lb_fun.eval(session=sess) * Ntest
print "test data VLB: ", np.mean(test_lb)
def callback(itr):
def samplefun(num_samps):
import numpy as np
z = np.array(np.random.randn(num_samps, zdim), dtype=np.float32)
return decode(z).eval(session=sess)
viz.plot_samples(itr, samplefun, savedir='avae_mnist_samples')
def sample_normal(mu, log_sigmasq, M=5):
dim = mu.get_shape()[1].value
eps = tf.random_normal((M, dim), dtype=tf.float32)
return mu + tf.exp(0.5 * log_sigmasq) * eps
def recons(num_samps):
# choose one datapoint, encode it
subset = X[np.random.choice(X.shape[0], 1)]
# compute encoder - first aux, then encoder conditioned on aux
amu, alog_sigmasq = aux_encode(subset)
asamps = sample_normal(amu, alog_sigmasq, M=24)
Xa = tf.concat(1, [tf.tile(subset, [24, 1]), asamps])
zmu, zlog_sigmasq = encode(Xa)
imgs = decode(sample_normal(zmu, zlog_sigmasq,
M=24)).eval(session=sess)
return np.row_stack([subset, imgs])
viz.plot_samples(itr,
recons,
savedir='avae_mnist_samples',
stub='recon')
test_lb = test_lb_fun.eval(session=sess) * Ntest
print "test data VLB: ", np.mean(test_lb)
# plot z space for each test data item, colored by digit type
def plot_test_latent():
amu, alog_sigmasq = aux_encode(Xtest)
asamps = sample_normal(amu, alog_sigmasq, M=1)
zmu, zlog_sigmasq = encode(tf.concat(1, [Xtest, asamps]))
z = sample_normal(zmu, zlog_sigmasq, M=1).eval(session=sess)
print z.shape
import matplotlib.pyplot as plt
fig = plt.figure()
ax = plt.gca()
import seaborn as sns
colors = sns.color_palette('muted', n_colors=10)
for i in xrange(10):
idx = np.argmax(Yt, 1) == i
ax.scatter(z[idx, 0], z[idx, 1], c=colors[i], label='%d' % i)
ax.legend()
fig.savefig("avae_mnist_samples/latent_space_%03d.png" % itr)
plot_test_latent()
def callback(itr):
def samplefun(num_samps):
import numpy as np
z = np.array(np.random.randn(num_samps, zdim), dtype=np.float32)
return decode(z).eval(session=sess)
viz.plot_samples(itr, samplefun, savedir='vae_mnist_samples')
def sample_z(mu, log_sigmasq, M=5):
eps = tf.random_normal((M, zdim), dtype=tf.float32)
return mu + tf.exp(0.5 * log_sigmasq) * eps
def recons(num_samps):
# random subset
subset = X[np.random.choice(X.shape[0], 1)]
mu, log_sigmasq = encode(subset)
imgs = decode(sample_z(mu, log_sigmasq, M=24)).eval(session=sess)
return np.row_stack([subset, imgs])
viz.plot_samples(itr, recons, savedir='vae_mnist_samples', stub='recon')
test_lb = test_lb_fun.eval(session=sess) * Ntest
print "test data VLB: ", np.mean(test_lb)
def callback(itr):
def samplefun(num_samps):
import numpy as np
z = np.array(np.random.randn(num_samps, zdim), dtype=np.float32)
return decode(z).eval(session=sess)
viz.plot_samples(itr, samplefun, savedir='vae_mnist_samples')
def sample_z(mu, log_sigmasq, M=5):
eps = tf.random_normal((M, zdim), dtype=tf.float32)
return mu + tf.exp(0.5 * log_sigmasq) * eps
def recons(num_samps):
# random subset
subset = X[np.random.choice(X.shape[0], 1)]
mu, log_sigmasq = encode(subset)
imgs = decode(sample_z(mu, log_sigmasq, M=24)).eval(session=sess)
return np.row_stack([subset, imgs])
viz.plot_samples(itr, recons, savedir='vae_mnist_samples', stub='recon')
test_lb = test_lb_fun.eval(session=sess) * Ntest
print "test data VLB: ", np.mean(test_lb)
def callback(itr):
# plot samples from the model generative process
viz.plot_samples(itr, sample_fun, savedir=output_dir)
# plot reconstructed samples
viz.plot_samples(itr, recon_fun, savedir=output_dir, stub='recon')
# evaluate test
test_lb = test_vlb_fun()
print "test data VLB: ", np.mean(test_lb)
# plot z space for each test data item, colored by digit type
def plot_test_latent():
Z, Y = latent_space_fun()
fig = plt.figure()
ax = plt.gca()
for i in xrange(10):
idx = np.argmax(Y, 1) == i
ax.scatter(Z[idx,0], Z[idx,1], c=colors[i], label='%d'%i)
ax.legend()
fig.savefig(os.path.join(output_dir, "latent_space_%03d.png" % itr))
plot_test_latent()
def make_gmm_sample_fun():
z_gmm = tf.placeholder(tf.float32, shape=(None, zdim))
sfun_gmm = decode(z_gmm)
def gmm_sample_fun(num_samps):
return sfun_gmm.eval(feed_dict={z_gmm: gmm.sample(num_samps)},
session=sess)
return gmm_sample_fun
gmm_sample_fun = make_gmm_sample_fun()
viz.plot_samples(0,
gmm_sample_fun,
savedir=OUTPUT_DIR,
stub='gmm_',
sidelen=10)
viz.plot_samples(0,
sample_fun,
savedir=OUTPUT_DIR,
stub='big_',
sidelen=10)
import matplotlib.pyplot as plt
xg = yg = np.linspace(-3, 3, 100)
xx, yy = np.meshgrid(xg, yg)
fig = plt.figure(figsize=(8, 8))
z = gmm.score(np.column_stack([xx.flatten(), yy.flatten()]))
plt.contourf(xx, yy, np.exp(z.reshape(xx.shape)))
import os
zt = ztrain.eval(session=sess, feed_dict={Xtrain:X})
# fit a mixture model to zt, and sample from GMM prior
from sklearn.mixture import GMM
gmm = GMM(n_components=20, covariance_type='full')
gmm.fit(zt)
def make_gmm_sample_fun():
z_gmm = tf.placeholder(tf.float32, shape=(None, zdim))
sfun_gmm = decode(z_gmm)
def gmm_sample_fun(num_samps):
return sfun_gmm.eval(feed_dict={z_gmm: gmm.sample(num_samps)}, session=sess)
return gmm_sample_fun
gmm_sample_fun = make_gmm_sample_fun()
viz.plot_samples(0, gmm_sample_fun, savedir=OUTPUT_DIR, stub='gmm_', sidelen=10)
viz.plot_samples(0, sample_fun, savedir=OUTPUT_DIR, stub='big_', sidelen=10)
import matplotlib.pyplot as plt
xg = yg = np.linspace(-3, 3, 100)
xx, yy = np.meshgrid(xg, yg)
fig = plt.figure(figsize=(8,8))
z = gmm.score(np.column_stack([xx.flatten(), yy.flatten()]))
plt.contourf(xx, yy, np.exp(z.reshape(xx.shape)))
import os
plt.savefig(os.path.join(OUTPUT_DIR, 'prob_z_gmm.png'))
plt.close("all")
else: