proxynca_pp/loss.py at master · counterfake/proxynca_pp · 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
from similarity import pairwise_distance
import torch
import torch.nn.functional as F
import numpy as np
import torch.nn as nn
import copy
import math

def binarize_and_smooth_labels(T, nb_classes, smoothing_const = 0):
    import sklearn.preprocessing
    T = T.cpu().numpy()
    T = sklearn.preprocessing.label_binarize(
        T, classes = range(0, nb_classes)
    )
    T = T * (1 - smoothing_const)
    T[T == 0] = smoothing_const / (nb_classes - 1)
    T = torch.FloatTensor(T).cuda()

    return T

class ProxyNCA_classic(torch.nn.Module):
    def __init__(self, nb_classes, sz_embed, scale, **kwargs):
        torch.nn.Module.__init__(self)
        self.proxies = torch.nn.Parameter(torch.randn(nb_classes, sz_embed) / 8)
        self.scale = scale

    def forward(self, X, T):
        P = self.proxies

        #note: self.scale is equal to sqrt(1/T)
        # in the paper T = 1/9, therefore, scale = sart(1/(1/9)) = sqrt(9) = 3
        #  we need to apply sqrt because the pairwise distance is calculated as norm^2
        P = self.scale * F.normalize(P, p = 2, dim = -1)
        X = self.scale * F.normalize(X, p = 2, dim = -1)

        D = pairwise_distance(
            torch.cat(
                [X, P]
            ),
            squared = True
        )[:X.size()[0], X.size()[0]:]

        T = binarize_and_smooth_labels(
            T = T, nb_classes = len(P), smoothing_const = 0
        )
        loss1 = torch.sum(T * torch.exp(-D), -1)
        loss2 = torch.sum((1-T) * torch.exp(-D), -1)
        loss = -torch.log(loss1/loss2)
        loss = loss.mean()
        return loss

class ProxyNCA_prob(torch.nn.Module):
    def __init__(self, nb_classes, sz_embed, scale, **kwargs):
        torch.nn.Module.__init__(self)
        self.proxies = torch.nn.Parameter(torch.randn(nb_classes, sz_embed) / 8)
        self.scale = scale

    def forward(self, X, T):
        P = self.proxies
        #note: self.scale is equal to sqrt(1/T)
        # in the paper T = 1/9, therefore, scale = sart(1/(1/9)) = sqrt(9) = 3
        #  we need to apply sqrt because the pairwise distance is calculated as norm^2

        P = self.scale * F.normalize(P, p = 2, dim = -1)
        X = self.scale * F.normalize(X, p = 2, dim = -1)

        D = pairwise_distance(
            torch.cat(
                [X, P]
            ),
            squared = True
        )[:X.size()[0], X.size()[0]:]

        T = binarize_and_smooth_labels(
            T = T, nb_classes = len(P), smoothing_const = 0
        )

        loss = torch.sum(- T * F.log_softmax(-D, -1), -1)
        loss = loss.mean()
        return loss