vbp/architectures.py at master · manuelhaussmann/vbp · GitHub

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
import math

import numpy as np
import torch as th
import torch.nn as nn
import torch.nn.functional as F
from VBPLayer import VBPLinear, VBPConv


# A net with 2 fully connected VBP layer for regression.
class VBPNetRegression(nn.Module):
    def __init__(self, n_features, beta=1.0, n_hidden=50, C=100):
        super(VBPNetRegression, self).__init__()

        self.beta = beta
        self.C = C

        self.dense1 = VBPLinear(n_features, n_hidden)
        self.dense2 = VBPLinear(n_hidden, 1, isoutput=True)


    def forward(self, inp):
        C = self.C

        h = self.dense1(inp)
        zh = th.sigmoid(C*h) * h
        mu = self.dense2(zh)

        v = self.dense1.var(inp, None, C)
        var = self.dense2.var(zh, v, C)

        return mu, var


    def loss(self, data, target, n_train=None):
        C = self.C

        h = self.dense1(data)
        zh = th.sigmoid(C*h) * h
        mu = self.dense2(zh)

        v = self.dense1.var(data, None, C)
        var = self.dense2.var(zh, v, C)

        KL = sum(l.KL() for l in [self.dense1, self.dense2])

        return 0.5 * self.beta * ((mu - target).pow(2)  + var).mean() + KL/n_train


    def compute_beta(self, loader, iscuda, setbeta=False):
        with th.no_grad():
            coll = 0
            for data, target in loader:
                if iscuda:
                    data, target = data.cuda(), target.cuda()
                mu, var = self.forward(data)
                coll += ((target - mu).pow(2) + var).sum().item()
        if setbeta:
            self.beta = len(loader.dataset)/coll


# A LeNet sized net with convolutional and fully connected layer for classification tasks
class VBPLeNet(nn.Module):
    def __init__(self, n_channels=1, n_classes=10, sample=True):
        super(VBPLeNet, self).__init__()
        self.sample = sample
        self.n_channels = n_channels


        if n_channels == 1:
            self.conv1 = VBPConv(1, 20, 5, stride=2)
            self.conv2 = VBPConv(20, 50, 5, stride=2)
            self.dense1 = VBPLinear(4*4*50, 500)
            self.dense2 = VBPLinear(500, n_classes, isoutput=True)

        elif n_channels == 3:
            self.conv1 = VBPConv(3, 192, 5, stride=2)
            self.conv2 = VBPConv(192, 192, 5, stride=2)
            self.dense1 = VBPLinear(5 * 5 * 192, 1000)
            self.dense2 = VBPLinear(1000, n_classes, isoutput=True)

        self.layers = [self.conv1, self.conv2, self.dense1, self.dense2]

    # Assumg a ReLU forward pass
    def forward(self, data):
        C = 100
        h1 = self.conv1(data); zh1 = th.sigmoid(C*h1) * h1
        h2 = self.conv2(zh1); zh2 = th.sigmoid(C*h2) * h2
        if self.n_channels == 1:
            zh2 = zh2.view(-1, 4*4*50)
        else:
            zh2 = zh2.view(-1, 5*5*192)
        h3 = self.dense1(zh2); zh3 = th.sigmoid(C*h3) * h3
        mu = self.dense2(zh3)
        # variance term
        v = self.conv1.var(data, None, C)
        v = self.conv2.var(zh1, v, C)
        if self.n_channels == 1:
            v = v.view(-1, 4*4*50)
        else:
            v = v.view(-1, 5*5*192)
        v = self.dense1.var(zh2, v, C)
        var = self.dense2.var(zh3, v, C)
        return mu, var

    def loss(self, data, target, n_data):
        mu, var = self.forward(data)

        KL = sum(l.KL() for l in self.layers)

        # Sampling based loss with a single sample
        if self.sample:
            logsoft = F.log_softmax(mu + var.sqrt() * th.randn_like(mu), 1)
            return -th.sum(target * logsoft, 1).mean() + KL/n_data
        else:
            # Delta method
            p = F.softmax(mu, 1)
            lse = th.logsumexp(mu, 1, True)
            snd = (var * p).sum(1, True) - (var * p.pow(2)).sum(1, True)

            logsftmx = mu - lse - 0.5 * snd
            avgnll = - (target * logsftmx).sum(1).mean()
            return avgnll + KL/n_data

    def predict(self, data, map=True):
        mu, var = self.forward(data)
        if map:
            return F.softmax(mu, 1)
        # Sampling based prediction with a single sample
        if self.sample:
            return F.softmax(mu + var.sqrt() * th.randn_like(mu), 1)
        else:
            p = F.softmax(mu, 1)
            snd = 1 + (var * p.pow(2)).sum(1, True) - (var * p).sum(1, True) + 0.5 * var - 0.5 * (var * p).sum(1, True)
            return p * snd