SelecMix/util.py at main · iwhwang/SelecMix · 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
import torch
import numpy as np
import torch.nn as nn

'''Modified from https://github.com/alinlab/LfF and https://github.com/kakaoenterprise/Learning-Debiased-Disentangled'''
class EMA:
    def __init__(self, label, num_classes=None, alpha=0.9):
        # self.label = label.cuda()
        self.label = label
        self.alpha = alpha
        self.parameter = torch.zeros(label.size(0))
        self.updated = torch.zeros(label.size(0))
        self.num_classes = num_classes
        # self.max = torch.zeros(self.num_classes).cuda()
        self.max = torch.zeros(self.num_classes)

    def update(self, data, index, curve=None, iter_range=None, step=None):
        self.parameter = self.parameter.to(data.device)
        self.updated = self.updated.to(data.device)
        index = index.to(data.device)

        if curve is None:
            self.parameter[index] = self.alpha * self.parameter[index] + (1 - self.alpha * self.updated[index]) * data
        else:
            alpha = curve ** -(step / iter_range)
            self.parameter[index] = alpha * self.parameter[index] + (1 - alpha * self.updated[index]) * data
        self.updated[index] = 1

    def max_loss(self, label):
        label_index = torch.where(self.label == label)[0]
        return self.parameter[label_index].max()


def setup_step_num(args, epch):
    total_epoch = 0
    if args.dataset == 'cifar10c': total_epoch = 300
    elif args.dataset == 'bffhq': total_epoch = 200
    elif args.dataset == 'cmnist': total_epoch = 200
    args.num_steps = epch * total_epoch
    args.valid_freq = epch
    args.log_freq = epch

def mixup_naive(batch_data, batch_label):
    alpha = 1.0
    beta = torch.distributions.beta.Beta(alpha, alpha)
    mix_idx = torch.randperm(batch_data.shape[0], device=batch_data.device)
    lam = beta.sample([batch_data.shape[0]]).to(device=batch_data.device)
    lam_expanded = lam.view([-1] + [1]*(batch_data.dim()-1))
    mixed_batch_data = lam_expanded * batch_data + (1. - lam_expanded) * batch_data[mix_idx]
    mix_batch_label = batch_label[mix_idx]
    return mixed_batch_data, mix_batch_label, lam, (1. - lam)

def inter_mask_with_sim_matrix(sim, indexes):
    diff_label = (indexes.t() != indexes)
    inter_mask = diff_label.to(int)
    masked_sim = sim * inter_mask
    mix_idx = torch.argmax(masked_sim, 1, keepdim=True).view(-1)
    return mix_idx

def intra_mask_with_sim_matrix(sim, indexes):
    same_label = (indexes.t() == indexes)
    intra_mask = same_label.to(int)
    masked_sim = sim * intra_mask + torch.ones_like(sim) * (1 - intra_mask)
    mix_idx = torch.argmin(masked_sim, 1, keepdim=True).view(-1)
    return mix_idx

def get_mixup_index_with_similarity_matrix(args, sim, batch_label):
    indexes = batch_label.unsqueeze(0)
    if args.intra: mix_idx = intra_mask_with_sim_matrix(sim, indexes)
    elif args.inter: mix_idx = inter_mask_with_sim_matrix(sim, indexes)
    elif args.rand:
        a = torch.rand(1)
        if a>0.5: mix_idx = intra_mask_with_sim_matrix(sim, indexes)
        else: mix_idx = inter_mask_with_sim_matrix(sim, indexes)
    will_be_mixed = torch.zeros_like(mix_idx).to(float)
    return mix_idx, will_be_mixed


def mixup_with_similarity_matrix(args, batch_data, batch_label, sim2):
    sim = sim2.clone().detach()
    mix_idx, will_be_mixed = get_mixup_index_with_similarity_matrix(args, sim, batch_label)

    return mixup_with_index(args, batch_data, batch_label, mix_idx, will_be_mixed)

def mixup_with_index(args, batch_data, batch_label, mix_idx, will_be_mixed):
    alpha = 1.0
    beta = torch.distributions.beta.Beta(alpha, alpha)
    lam = beta.sample([batch_data.shape[0]]).to(float).to(device=batch_data.device)
    if args.ours:
        lam = lam * 0.5
    lam_adj = torch.where(lam>will_be_mixed, lam, will_be_mixed)
    lam_expanded = lam_adj.view([-1] + [1]*(batch_data.dim()-1))
    mixed_batch_data = lam_expanded * batch_data + (1. - lam_expanded) * batch_data[mix_idx]
    mix_batch_label = batch_label[mix_idx].clone().detach()
    lam_adj_b = 1. - lam_adj.clone().detach()
    lam_adj = torch.where(lam>will_be_mixed, lam, will_be_mixed)
    if args.ours:
        lam_adj = torch.where(lam>will_be_mixed, torch.zeros_like(will_be_mixed).to(float), will_be_mixed)
        lam_adj_b = 1. - lam_adj.clone().detach()
    return mixed_batch_data.to(torch.float32), mix_batch_label, lam_adj.to(torch.float32), lam_adj_b.to(torch.float32)

def get_mixup_index_with_gt_bias_label(args, batch_attr):
    indexes = batch_attr[:,args.target_attr_idx].unsqueeze(0)
    indexes_bias = batch_attr[:,args.bias_attr_idx].unsqueeze(0)
    if args.intra: mask = intra_mask_with_gt_bias_label(indexes, indexes_bias)
    elif args.inter: mask = inter_mask_with_gt_bias_label(indexes, indexes_bias)
    elif args.rand:
        a = torch.rand(1)
        if a>0.5: mask = intra_mask_with_gt_bias_label(indexes, indexes_bias)
        else: mask = inter_mask_with_gt_bias_label(indexes, indexes_bias)
    mask = mask.clone().detach()
    d = mask.sum(1)
    will_be_mixed = (d==0).to(float)
    idx = torch.arange(mask.shape[1], 0, -1, device=batch_attr.device)
    mix_idx = torch.argmax(mask * idx, 1, keepdim=True).view(-1)
    return mix_idx, will_be_mixed

def mixup_with_gt_bias_label(args, batch_data, batch_attr, batch_label):
    mix_idx, will_be_mixed = get_mixup_index_with_gt_bias_label(args, batch_attr)
    return mixup_with_index(args, batch_data, batch_label, mix_idx, will_be_mixed)

def inter_mask_with_gt_bias_label(indexes, indexes_bias):
    diff_label = (indexes.t() != indexes)
    same_bias = (indexes_bias.t() == indexes_bias)
    inter_mask = (diff_label * same_bias).to(int)
    return inter_mask

def intra_mask_with_gt_bias_label(indexes, indexes_bias):
    same_label = (indexes.t() == indexes)
    diff_bias = (indexes_bias.t() != indexes_bias)
    intra_mask = (same_label * diff_bias).to(int)
    return intra_mask