transp = [0.9, 0.9, 0.5]
fig = plt.figure()
for c in range(0, len(DIG)):
col = colors[c]
plt.scatter(postz[y_train == DIG[c], 0], postz[y_train == DIG[c], 1], color=col,
label=DIG[c], marker=markers[c], alpha=transp[c], s=60)
plt.legend()
plt.show()
# Generate new images
x_gen = m.posterior_predictive('x').sample()
mnist.plot_digits(x_gen)
# Generate new images for a specific number
NUM = 0
postz_0 = postz[y_train == DIG[NUM]]
# Less than plaze size, so we need to tile this up to 1000 instances
postz_0 = np.tile(postz_0, [int(np.ceil(N / postz_0.shape[0])), 1])[:N]
x_gen = m.posterior_predictive('x', data={"z": postz_0}).sample()
mnist.plot_digits(x_gen)
# Show how numbers are codified in the domain of the hidden variable z
# First define the range of the z domain
xaxis_min, yaxis_min = np.min(postz, axis=0)
xaxis_max, yaxis_max = np.max(postz, axis=0)
DIG = [0, 1, 2]
# minimum scale
scale_epsilon = 0.01
# inference parameters
num_epochs = 1000
learning_rate = 0.01
tf.reset_default_graph()
tf.set_random_seed(1234)
#29
from inferpy.data import mnist
# load the data
(x_train, y_train), _ = mnist.load_data(num_instances=N, digits=DIG)
mnist.plot_digits(x_train, grid=[5, 5])
#38
############## Inferpy ##############
# P model and the decoder NN
@inf.probmodel
def vae(k, d0, dx):
with inf.datamodel():
z = inf.Normal(tf.ones(k), 1, name="z")
decoder = inf.layers.Sequential([
tf.keras.layers.Dense(d0, activation=tf.nn.relu),
tf.keras.layers.Dense(dx)
])
p.posterior("z", data={
"x": x_train[i:i + M, :]
}).sample() for i in range(0, N, M)
])
# for each input instance, plot the hidden encoding coloured by the number that it represents
markers = ["x", "+", "o"]
colors = [
plt.get_cmap("gist_rainbow")(0.05),
plt.get_cmap("gnuplot2")(0.08),
plt.get_cmap("gist_rainbow")(0.33)
]
transp = [0.9, 0.9, 0.5]
fig = plt.figure()
for c in range(0, len(DIG)):
col = colors[c]
plt.scatter(postz[y_train == DIG[c], 0],
postz[y_train == DIG[c], 1],
color=col,
label=DIG[c],
marker=markers[c],
alpha=transp[c],
s=60)
plt.legend()
plt.show()
mnist.plot_digits(x_gen, grid=[5, 5])
DIG = [0, 1, 2]
# minimum scale
scale_epsilon = 0.01
# inference parameters
num_epochs = 1000
learning_rate = 0.01
tf.reset_default_graph()
tf.set_random_seed(1234)
#29
from inferpy.data import mnist
# load the data
(x_train, y_train), _ = mnist.load_data(num_instances=N, digits=DIG)
mnist.plot_digits(x_train, grid=[5, 5])
#38
### Model definition
class Decoder(torch.nn.Module):
def __init__(self, k, d0, dx):
super(Decoder, self).__init__()
# setup the two linear transformations used
self.fc1 = torch.nn.Linear(k, d0)
self.fc21 = torch.nn.Linear(d0, dx)
# setup the non-linearities
self.softplus = torch.nn.Softplus()
self.sigmoid = torch.nn.Sigmoid()
DIG = [0, 1, 2]
# minimum scale
scale_epsilon = 0.01
# inference parameters
num_epochs = 1000
learning_rate = 0.01
tf.reset_default_graph()
tf.set_random_seed(1234)
#29
from inferpy.data import mnist
# load the data
(x_train, y_train), _ = mnist.load_data(num_instances=N, digits=DIG)
mnist.plot_digits(x_train, grid=[5, 5])
#38
# Model definition
######################
############## Edward ##############
def vae(k, d0, dx, N):
z = ed.Normal(loc=tf.ones(k), scale=1., sample_shape=N, name="z")
decoder = inf.layers.Sequential([
tf.keras.layers.Dense(d0, activation=tf.nn.relu, name="h0"),
tf.keras.layers.Dense(dx, name="h1")
