-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathANP.py
executable file
·169 lines (151 loc) · 8.22 KB
/
ANP.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
import logging
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.distributions import Normal
from backbone import xavier, num_flat_features
import math
import numpy as np
from backbone.utils.MLP import MLP, LatentMLP
from backbone.utils.attention_modules import MAB as Attention
from backbone.neural_processes.NP_Head_Base import NP_HEAD
class DetEncoder(nn.Module):
def __init__(self, input_dim, num_classes, latent_dim, num_layers=2, num_attn_heads=1, label_embedder=False, xavier_init=False):
super(DetEncoder, self).__init__()
set_encoding_dim = input_dim + num_classes if not label_embedder else input_dim
self.set_encoder = MLP(set_encoding_dim, latent_dim, latent_dim, num_layers, xavier_init=xavier_init)
self.attentions = nn.ModuleList([Attention(latent_dim, latent_dim, latent_dim, num_attn_heads, xavier_init=xavier_init) for _ in range(num_layers)])
self.context_projection = MLP(input_dim, latent_dim, latent_dim, 1, xavier_init=xavier_init)
self.target_projection = MLP(input_dim, latent_dim, latent_dim, 1, xavier_init=xavier_init)
self.cross_attentions = nn.ModuleList([Attention(latent_dim, latent_dim, latent_dim, num_attn_heads, xavier_init=xavier_init) for _ in range(num_layers)])
self.label_embedder = label_embedder
if label_embedder:
self.label_emb = nn.Embedding(num_classes, input_dim)
def forward(self, x, y, x_target):
if self.label_embedder:
d = x + self.label_emb(y)
else:
d = torch.cat((x, y), -1)
s = self.set_encoder(d)
# add attention here optionally
for attention in self.attentions:
s = attention(s, s, s)
x = self.context_projection(x)
x_target = self.target_projection(x_target)
for attention in self.cross_attentions:
x_target = attention(x, s, x_target)
return x_target
class LatentEncoder(nn.Module):
def __init__(self, input_dim, num_classes, latent_dim, num_layers=2, num_attn_heads=1, label_embedder=False, xavier_init=False):
super(LatentEncoder, self).__init__()
set_encoding_dim = input_dim + num_classes if not label_embedder else input_dim
self.set_encoder = MLP(set_encoding_dim, latent_dim, latent_dim, num_layers, xavier_init=xavier_init)
self.attentions = nn.ModuleList([Attention(latent_dim, latent_dim, latent_dim, num_attn_heads, xavier_init=xavier_init) for _ in range(num_layers)])
self.label_embedder = label_embedder
if label_embedder:
self.label_emb = nn.Embedding(num_classes, input_dim)
self.global_amortizer = LatentMLP(latent_dim, latent_dim, latent_dim,
num_layers, xavier_init=xavier_init)
def forward(self, x, y):
if self.label_embedder:
d = x + self.label_emb(y)
else:
d = torch.cat((x, y), -1)
s = self.set_encoder(d)
for attention in self.attentions:
s = attention(s, s, s)
q =self.global_amortizer(s.mean(0))
return q
class Decoder(nn.Module):
def __init__(self, decoder_input_dim, latent_dim, num_layers=2, xavier_init=False):
super(Decoder, self).__init__()
self.decoder = MLP(decoder_input_dim, latent_dim, latent_dim, num_layers, xavier_init=xavier_init)
def forward(self, x_in, r_det=None, z_lat=None):
decoder_in = torch.cat((x_in, r_det), -1) if r_det is not None else x_in
if z_lat is not None:
decoder_in = torch.cat((decoder_in, z_lat), -1)
decoder_out = self.decoder(decoder_in)
return decoder_out
class ANP_HEAD(NP_HEAD):
def __init__(self, input_dim,
latent_dim,
num_classes,
n_tasks,
cls_per_task,
num_layers=2,
xavier_init=False,
num_attn_heads=4,
label_embedder=False,
context_task_labels=None,
target_task_labels=None,
test_oracle=False,
use_deterministic=True,
hierarchy=False
):
super().__init__(input_dim, latent_dim, num_classes, n_tasks, cls_per_task, num_layers, xavier_init)
self.det_encoder = DetEncoder(input_dim, num_classes, latent_dim, num_layers=num_layers,
num_attn_heads=num_attn_heads, label_embedder=label_embedder,
xavier_init=xavier_init) if use_deterministic else None
self.latent_encoder = LatentEncoder(input_dim, num_classes, latent_dim, num_layers=num_layers,
num_attn_heads=num_attn_heads, label_embedder=label_embedder,
xavier_init=xavier_init)
self.fc_decoder = Decoder(self.decoder_input_dim, input_dim, num_layers=num_layers, xavier_init=xavier_init)
self.classifier = nn.Linear(input_dim, num_classes, bias=True)
self.use_deterministic = use_deterministic
if xavier_init:
self.net = nn.Sequential(self.classifier)
self.reset_parameters()
def reset_parameters(self) -> None:
"""
Calls the Xavier parameter initialization function.
"""
self.net.apply(xavier)
def forward(self, x_context_in, labels_context_in, x_target_in, labels_target_in=None,
phase_train=True, forward_times=1, epoch=None, cur_test_task=None,
clnp_stochasticity='all_global', context_task_labels=None, target_task_labels=None,
task_to_labels=None, top_k_decode=1, x_percent=10, prev_task_distr = None) -> tuple:
B = x_target_in.size(0)
context_representation_deterministic_expand = None
if phase_train:
q_target = self.latent_encoder(x_target_in, labels_target_in)
q_context = self.latent_encoder(x_context_in, labels_context_in)
latent_z_target = None
for i in range(0, forward_times):
z =Normal(q_target[0], q_target[1]).rsample()
z = z.unsqueeze(0)
if i == 0:
latent_z_target = z
else:
latent_z_target = torch.cat((latent_z_target, z))
latent_z_target_expand = latent_z_target.unsqueeze(1).expand(-1, B, -1)
x_target_in_expand = x_target_in.unsqueeze(0).expand(forward_times, -1, -1)
if self.use_deterministic:
x_representation_deterministic = self.det_encoder(x_context_in, labels_context_in, x_target_in)
context_representation_deterministic_expand = x_representation_deterministic.unsqueeze(0).expand(
forward_times, -1, -1)
################## decoder ####
##############
output_function = self.fc_decoder(x_target_in_expand, context_representation_deterministic_expand,
latent_z_target_expand)
output = self.classifier(output_function)
return output, ((q_context, None), (q_target, None)), None
else:
q = self.latent_encoder(x_context_in, labels_context_in)
latent_z_target = None
for i in range(0, forward_times):
z = Normal(q[0], q[1]).rsample()
z = z.unsqueeze(0)
if i == 0:
latent_z_target = z
else:
latent_z_target = torch.cat((latent_z_target, z))
latent_z_target_expand = latent_z_target.unsqueeze(1).expand(-1, B, -1)
x_target_in_expand = x_target_in.unsqueeze(0).expand(forward_times, -1, -1)
if self.use_deterministic:
x_representation_deterministic = self.det_encoder(x_context_in, labels_context_in, x_target_in)
context_representation_deterministic_expand = x_representation_deterministic.unsqueeze(0).expand(
forward_times, -1, -1)
output_function = self.fc_decoder(x_target_in_expand, context_representation_deterministic_expand,
latent_z_target_expand)
output = self.classifier(output_function)
return output, None