1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
|
import torch
from torch import nn
class ConvolutionalVariationalAutoencoder(nn.Module):
def __init__(self, latent_dims, hidden_dims, image_shape, dropout=0.0):
super(ConvolutionalVariationalAutoencoder, self).__init__()
self.latent_dims = latent_dims # Size of the latent space layer
self.hidden_dims = (
hidden_dims # List of hidden layers number of filters/channels
)
self.image_shape = image_shape # Input image shape
self.last_channels = self.hidden_dims[-1]
self.in_channels = self.image_shape[0]
# Simple formula to get the number of neurons after the last convolution layer is flattened
self.flattened_channels = int(
self.last_channels
* (self.image_shape[1] / (2 ** len(self.hidden_dims))) ** 2
)
# For each hidden layer we will create a Convolution Block
modules = []
for h_dim in self.hidden_dims:
modules.append(
nn.Sequential(
nn.Conv2d(
in_channels=self.in_channels,
out_channels=h_dim,
kernel_size=3,
stride=2,
padding=1,
),
nn.BatchNorm2d(h_dim),
nn.LeakyReLU(),
nn.Dropout(p=dropout),
)
)
self.in_channels = h_dim
self.encoder = nn.Sequential(*modules)
# Here are our layers for our latent space distribution
self.fc_mu = nn.Linear(self.flattened_channels, latent_dims)
self.fc_var = nn.Linear(self.flattened_channels, latent_dims)
# Decoder input layer
self.decoder_input = nn.Linear(latent_dims, self.flattened_channels)
# For each Convolution Block created on the Encoder we will do a symmetric Decoder with the same Blocks, but using ConvTranspose
self.hidden_dims.reverse()
modules = []
for h_dim in self.hidden_dims:
modules.append(
nn.Sequential(
nn.ConvTranspose2d(
in_channels=self.in_channels,
out_channels=h_dim,
kernel_size=3,
stride=2,
padding=1,
output_padding=1,
),
nn.BatchNorm2d(h_dim),
nn.LeakyReLU(),
nn.Dropout(p=dropout),
)
)
self.in_channels = h_dim
self.decoder = nn.Sequential(*modules)
# The final layer the reconstructed image have the same dimensions as the input image
self.final_layer = nn.Sequential(
nn.Conv2d(
in_channels=self.in_channels,
out_channels=self.image_shape[0],
kernel_size=3,
padding=1,
),
nn.Sigmoid(),
)
def get_latent_dims(self):
return self.latent_dims
def encode(self, input):
"""
Encodes the input by passing through the encoder network
and returns the latent codes.
"""
result = self.encoder(input)
result = torch.flatten(result, start_dim=1)
# Split the result into mu and var componentsbof the latent Gaussian distribution
mu = self.fc_mu(result)
log_var = self.fc_var(result)
return [mu, log_var]
def decode(self, z):
"""
Maps the given latent codes onto the image space.
"""
result = self.decoder_input(z)
result = result.view(
-1,
self.last_channels,
int(self.image_shape[1] / (2 ** len(self.hidden_dims))),
int(self.image_shape[1] / (2 ** len(self.hidden_dims))),
)
result = self.decoder(result)
result = self.final_layer(result)
return result
def reparameterize(self, mu, log_var):
"""
Reparameterization trick to sample from N(mu, var) from N(0,1).
"""
std = torch.exp(0.5 * log_var)
eps = torch.randn_like(std)
return mu + eps * std
def forward(self, input):
"""
Forward method which will encode and decode our image.
"""
mu, log_var = self.encode(input)
z = self.reparameterize(mu, log_var)
return [self.decode(z), input, mu, log_var, z]
def loss_function(self, recons, input, mu, log_var):
"""
Computes VAE loss function
"""
recons_loss = nn.functional.binary_cross_entropy(
recons.reshape(recons.shape[0], -1),
input.reshape(input.shape[0], -1),
reduction="none",
).sum(dim=-1)
kld_loss = -0.5 * torch.sum(1 + log_var - mu.pow(2) - log_var.exp(), dim=-1)
loss = (recons_loss + kld_loss).mean(dim=0)
return loss
def sample(self, num_samples, device):
"""
Samples from the latent space and return the corresponding
image space map.
"""
z = torch.randn(num_samples, self.latent_dims)
z = z.to(device)
samples = self.decode(z)
return samples
def generate(self, x):
"""
Given an input image x, returns the reconstructed image
"""
return self.forward(x)[0]
def interpolate(self, starting_inputs, ending_inputs, device, granularity=10):
"""This function performs a linear interpolation in the latent space of the autoencoder
from starting inputs to ending inputs. It returns the interpolation trajectories.
"""
mu, log_var = self.encode(starting_inputs.to(device))
starting_z = self.reparameterize(mu, log_var)
mu, log_var = self.encode(ending_inputs.to(device))
ending_z = self.reparameterize(mu, log_var)
t = torch.linspace(0, 1, granularity).to(device)
intep_line = torch.kron(
starting_z.reshape(starting_z.shape[0], -1), (1 - t).unsqueeze(-1)
) + torch.kron(ending_z.reshape(ending_z.shape[0], -1), t.unsqueeze(-1))
decoded_line = self.decode(intep_line).reshape(
(starting_inputs.shape[0], t.shape[0]) + (starting_inputs.shape[1:])
)
return decoded_line
|