forked from vsimkus/variational-gibbs-inference
-
Notifications
You must be signed in to change notification settings - Fork 0
/
fc_partial_vae.py
211 lines (167 loc) · 9.11 KB
/
fc_partial_vae.py
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
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
import torch
import torch.nn as nn
import torch.nn.functional as F
from cdi.models.fc_vae import FC_VAE
from cdi.util.arg_utils import parse_bool
class FC_PartialVAE(FC_VAE):
"""
Fully-connected PartialVAE with PointNet+
"EDDI: Efficient Dynamic Discovery of High-Value Information with Partial VAE" by Ma et al. (2019)
"""
@staticmethod
def add_model_args(parser):
parser.add_argument('--fc_vae_model.num_z_samples', type=int,
default=1, help='Number of latent samples.')
parser.add_argument('--fc_vae_model.input_dim', type=int,
required=True, help='Image dimension.')
parser.add_argument('--fc_vae_model.decoder_hidden_dims', type=int,
nargs='+', required=True,
help='Decoder layer dimensionalities')
parser.add_argument('--fc_vae_model.activation',
type=str, required=True,
help='Activation: lrelu or sigmoid.',
choices=['lrelu', 'sigmoid'])
parser.add_argument('--fc_vae_model.bound',
type=str, required=True,
choices=['vae', 'miwae'],
help='Which bound to use.')
parser.add_argument('--fc_vae_model.marginalise',
type=parse_bool, required=False,
help=('For VAE bound, marginalise the generator or not.'))
parser.add_argument('--fc_vae_model.marginalise_val',
type=parse_bool, required=False,
help=('For VAE bound, marginalise the generator or not. (val only)'))
parser.add_argument('--fc_vae_model.pos_emb_dim', type=int,
help=('Size of position embedding.'))
parser.add_argument('--fc_vae_model.encoder_shared_net_layers', type=int,
nargs='+', help='Number of hidden layers in the shared net.')
parser.add_argument('--fc_vae_model.encoder_hidden_dims', type=int,
nargs='+', required=True,
help='Encoder layer dimensionalities')
parser.add_argument('--fc_vae_model.encoder_activation',
type=str, required=True,
help='Encoder activation: lrelu or sigmoid.',
choices=['lrelu', 'sigmoid'])
parser.add_argument('--fc_vae_model.encoder_residuals', default=False,
type=parse_bool, help=('Whether to use residual connections in the encoder or not.'))
parser.add_argument('--fc_vae_model.aggregation', default='sum',
choices=['sum', 'avg'], type=str,
help='What type of invariante aggregation should be used.')
parser.add_argument('--fc_vae_model.no_aggregation_activation', default=False,
type=parse_bool, help='If true, then no activation after aggregation.')
return parser
def initialise(self):
assert self.hparams.activation in ('lrelu', 'sigmoid'), \
'Activation not supported!'
activation = self.hparams.activation if self.hparams.activation != 'lrelu' else 'leaky_relu'
pos_emb = torch.empty(self.hparams.input_dim, self.hparams.pos_emb_dim, dtype=torch.float)
nn.init.kaiming_uniform_(pos_emb, a=0.1, nonlinearity=activation)
self.pos_emb = torch.nn.Parameter(pos_emb, requires_grad=True)
b = torch.randn(self.hparams.input_dim, dtype=torch.float)
self.b = torch.nn.Parameter(b, requires_grad=True)
# Add input dimensions to the list
hidden_dims = [self.hparams.pos_emb_dim+2] + self.hparams.encoder_shared_net_layers
shared_encoder = []
for i in range(len(hidden_dims)-1):
if self.hparams.encoder_activation == 'sigmoid':
activation = F.sigmoid
elif self.hparams.encoder_activation == 'lrelu':
activation = F.leaky_relu
shared_encoder.append(FCLayer(hidden_dims[i], hidden_dims[i+1],
residual=self.hparams.encoder_residuals,
activation=activation))
self.shared_encoder = nn.Sequential(*shared_encoder)
# Add input dimensions to the list
hidden_dims = [self.hparams.encoder_shared_net_layers[-1]] + self.hparams.encoder_hidden_dims
encoder = []
for i in range(len(hidden_dims)-2):
if self.hparams.encoder_activation == 'sigmoid':
activation = F.sigmoid
elif self.hparams.encoder_activation == 'lrelu':
activation = F.leaky_relu
encoder.append(FCLayer(hidden_dims[i], hidden_dims[i+1],
residual=self.hparams.encoder_residuals,
activation=activation))
self.encoder = nn.Sequential(*encoder)
self.enc_mean = nn.Linear(hidden_dims[-2], hidden_dims[-1])
self.enc_log_var = nn.Linear(hidden_dims[-2], hidden_dims[-1])
# Add output dimensions to the list
hidden_dims = self.hparams.decoder_hidden_dims + [self.hparams.input_dim]
decoder = []
for i in range(len(hidden_dims)-2):
decoder.append(nn.Linear(hidden_dims[i], hidden_dims[i+1]))
if self.hparams.activation == 'sigmoid':
decoder.append(nn.Sigmoid())
elif self.hparams.activation == 'lrelu':
decoder.append(nn.LeakyReLU())
self.decoder = nn.Sequential(*decoder)
self.dec_mean = nn.Linear(hidden_dims[-2], hidden_dims[-1])
self.dec_log_var = nn.Linear(hidden_dims[-2], hidden_dims[-1])
def encode(self, X, M=None):
X_orig = X
X = X.unsqueeze(-1).expand(-1, -1, self.pos_emb.shape[-1]) * self.pos_emb.unsqueeze(0).expand(X.shape[0], -1, -1)
# Add the original X and a bias as per their implementation (this is not mentioned in their paper)
X = torch.cat([X_orig.unsqueeze(-1), X, self.b.unsqueeze(0).unsqueeze(-1).expand(X_orig.shape[0], -1, -1)], dim=-1)
X_emb = self.shared_encoder(X)
if M is not None:
X_emb = X_emb * M.unsqueeze(-1).expand(-1, -1, X_emb.shape[-1])
if not hasattr(self.hparams, 'aggregation') or self.hparams.aggregation == 'sum':
X_emb = torch.sum(X_emb, axis=-2)
elif self.hparams.aggregation == 'avg':
X_emb = torch.sum(X_emb, axis=-2)
# Compute average
if M is not None:
M_summed = M.sum(axis=-1, keepdim=True)
# Avoid division by zero
M_summed[M_summed == 0] = 1
X_emb = X_emb / M_summed
if not hasattr(self.hparams, 'no_aggregation_activation') or not self.hparams.no_aggregation_activation:
if self.hparams.encoder_activation == 'sigmoid':
X_emb = F.sigmoid(X_emb)
elif self.hparams.encoder_activation == 'lrelu':
X_emb = F.leaky_relu(X_emb)
Z_prime = self.encoder(X_emb)
Z_mean = self.enc_mean(Z_prime)
Z_log_var = self.enc_log_var(Z_prime)
return Z_mean, Z_log_var
def reset_encoder(self):
activation = self.hparams.activation if self.hparams.activation != 'lrelu' else 'leaky_relu'
pos_emb = torch.empty(self.hparams.input_dim, self.hparams.pos_emb_dim, dtype=torch.float)
nn.init.kaiming_uniform_(pos_emb, a=0.1, nonlinearity=activation)
self.pos_emb = torch.nn.Parameter(pos_emb, requires_grad=True)
b = torch.randn(self.hparams.input_dim, dtype=torch.float)
self.b = torch.nn.Parameter(b, requires_grad=True)
for layer in self.shared_encoder:
if not isinstance(layer, (nn.LeakyReLU, nn.Sigmoid)):
layer.reset_parameters()
for layer in self.encoder:
if not isinstance(layer, (nn.LeakyReLU, nn.Sigmoid)):
layer.reset_parameters()
self.enc_mean.reset_parameters()
self.enc_log_var.reset_parameters()
# Hacky way to separate the parameters so I can use separate optimisers
# for both parts of the model
def generator_parameters(self, recurse=True):
for name, param in self.named_parameters(recurse=recurse):
if 'dec' in name:
yield param
def encoder_parameters(self, recurse=True):
for name, param in self.named_parameters(recurse=recurse):
if 'dec' not in name:
yield param
class FCLayer(nn.Linear):
def __init__(self, *args, residual=False, activation=F.relu, **kwargs):
super().__init__(*args, **kwargs)
self.residual = residual
self.activation = activation
if self.in_features != (self.out_features):
self.residual = False
def forward(self, X):
X_in = X # for residual connection
# Apply linear transform
X = super().forward(X)
if self.activation:
X = self.activation(X)
if self.residual:
X = X_in + X # residual connection
return X