PosteL/relabel.py at main · ml-postech/PosteL · 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
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
212
213
214
215
import torch
from torch_geometric.utils import one_hot, to_torch_sparse_tensor, spmm, remove_self_loops, to_dense_adj, degree, dense_to_sparse, k_hop_subgraph

def neighborhood_label_distribution(data, train_mask, val_mask, test_mask, num_classes, num_hop=1, for_train=True):
    # Get neighborhood label distribution from train data
    assert data.y.dim() == 1
    y = data.y.clone().detach()
    if for_train:
        y[val_mask] = num_classes
        y[test_mask] = num_classes
    one_hot_y = one_hot(y, num_classes+1)
    one_hot_y = one_hot_y[:,:-1]
    sparse_edge_index = to_torch_sparse_tensor(data.edge_index)

    label_dist = spmm(sparse_edge_index, one_hot_y)

    label_distribution = []
    for i in range(num_classes):
        if for_train:
            label_histogram = label_dist[train_mask][data.y[train_mask]==i].sum(dim=0)
        else:
            label_histogram = label_dist[data.y==i].sum(dim=0)
        label_histogram = label_histogram / label_histogram.sum()
        if label_histogram.isnan().max():
            label_distribution.append(torch.ones_like(label_histogram).view(1,-1)/label_histogram.shape[0])
        else:
            label_distribution.append(label_histogram.view(1,-1))
    label_distribution = torch.cat(label_distribution, dim=0)

    return label_distribution.to(torch.float64), label_dist.to(torch.int)

def nodewise_neighborhood_label_distribution(data, train_mask, val_mask, test_mask, num_classes, for_train=True):
    # Get neighborhood label distribution from train data
    assert data.y.dim() == 1
    y = data.y.clone().detach()
    if for_train:
        y[val_mask] = num_classes
        y[test_mask] = num_classes
    one_hot_y = one_hot(y, num_classes+1)
    one_hot_y = one_hot_y[:,:-1]
    dense_adj = to_dense_adj(data.edge_index).squeeze()
    normalized_dense_adj = (1/degree(data.edge_index[0], data.x.shape[0])).unsqueeze(1) * dense_adj
    normalized_sparse_edge_index = dense_to_sparse(normalized_dense_adj)
    normalized_sparse_edge_index = to_torch_sparse_tensor(normalized_sparse_edge_index[0], normalized_sparse_edge_index[1])
    label_dist = spmm(normalized_sparse_edge_index, one_hot_y)

    # re-normazlie to exclude valid, test nodes
    label_dist = label_dist / label_dist.sum(dim=1, keepdim=True)
    label_dist[label_dist.isnan().max(dim=1)[0]] = 0

    label_distribution = []
    for i in range(num_classes):
        if for_train:
            label_histogram = label_dist[train_mask][data.y[train_mask]==i].sum(dim=0)
        else:
            label_histogram = label_dist[data.y==i].sum(dim=0)
        label_histogram = label_histogram / label_histogram.sum()
        label_distribution.append(label_histogram.view(1,-1))
    label_distribution = torch.cat(label_distribution, dim=0)

    sparse_edge_index = to_torch_sparse_tensor(data.edge_index)
    unnormalized_label_dist = spmm(sparse_edge_index, one_hot_y)

    return label_distribution.to(torch.float64), unnormalized_label_dist.to(torch.int)

def nodewise_neighborhood_label_distribution_with_pseudo_label(data, train_mask, val_mask, test_mask, num_classes, pseudo_label):
    # Get neighborhood label distribution from train data
    assert data.y.dim() == 1
    y = data.y.clone().detach()
    test_val_mask = torch.logical_or(val_mask, test_mask)
    y[test_val_mask] = pseudo_label
    one_hot_y = one_hot(y, num_classes)

    dense_adj = to_dense_adj(data.edge_index).squeeze()
    normalized_dense_adj = (1/degree(data.edge_index[0], data.x.shape[0])).unsqueeze(1) * dense_adj
    normalized_sparse_edge_index = dense_to_sparse(normalized_dense_adj)
    normalized_sparse_edge_index = to_torch_sparse_tensor(normalized_sparse_edge_index[0], normalized_sparse_edge_index[1])
    label_dist = spmm(normalized_sparse_edge_index, one_hot_y)

    # re-normazlie to exclude valid, test nodes
    label_dist = label_dist / label_dist.sum(dim=1, keepdim=True)
    label_dist[label_dist.isnan().max(dim=1)[0]] = 0

    label_distribution = []
    for i in range(num_classes):
        label_histogram = label_dist[train_mask][data.y[train_mask]==i].sum(dim=0)
        label_histogram = label_histogram / label_histogram.sum()
        label_distribution.append(label_histogram.view(1,-1))
    label_distribution = torch.cat(label_distribution, dim=0)

    sparse_edge_index = to_torch_sparse_tensor(data.edge_index)
    unnormalized_label_dist = spmm(sparse_edge_index, one_hot_y)

    return label_distribution.to(torch.float64), unnormalized_label_dist.to(torch.int)

def neighborhood_label_distribution_with_pseudo_label(data, train_mask, val_mask, test_mask, num_classes, pseudo_label):
    # Get neighborhood label distribution from train data and valid, test data with pseudo label
    assert data.y.dim() == 1
    y = data.y.clone().detach()
    test_val_mask = torch.logical_or(val_mask, test_mask)
    y[test_val_mask] = pseudo_label

    one_hot_y = one_hot(y, num_classes+1)
    one_hot_y = one_hot_y[:,:-1]
    sparse_edge_index = to_torch_sparse_tensor(data.edge_index)
    label_dist = spmm(sparse_edge_index, one_hot_y)

    label_distribution = []
    for i in range(num_classes):
        label_histogram = label_dist[train_mask][data.y[train_mask]==i].sum(dim=0)
        label_histogram = label_histogram / label_histogram.sum()
        label_distribution.append(label_histogram.view(1,-1))
    label_distribution = torch.cat(label_distribution, dim=0)

    return label_distribution.to(torch.float64), label_dist.to(torch.int)

def neighborhood_histogram_likelihood(label_histogram, label_distribution):
    likelihood = []
    for label_hist in label_histogram:
        likeli = (label_distribution**label_hist).prod(dim=1)
        likelihood.append(likeli)
    likelihood = torch.stack(likelihood)

    return likelihood

def label_smoothing(data, num_classes, smoothing_ratio):
    one_hot_y = one_hot(data.y[data.train_mask])
    new_y = (1-smoothing_ratio) * one_hot_y + smoothing_ratio/(num_classes-1) * torch.logical_not(one_hot_y)

    return new_y

def postel(data, num_classes, args):
    label_marginal_prob = one_hot(data.y[data.train_mask], num_classes).mean(dim=0)
    label_distribution, label_histogram = neighborhood_label_distribution(data, data.train_mask, data.val_mask, data.test_mask, num_classes, 1)
    hist_log_likelihood, degree_mask = neighborhood_histogram_log_likelihood(label_histogram[data.train_mask], label_distribution, args.degree_cutoff)

    new_label, mask = prob_label_from_log_likelihood(hist_log_likelihood, label_marginal_prob, args.temperature)

    mask = torch.logical_or(mask, degree_mask)
    new_label[mask] = one_hot(data.y[data.train_mask], num_classes)[mask]

    one_hot_y = one_hot(data.y[data.train_mask], num_classes)
    new_label = (1-args.smoothing_ratio) * new_label + args.smoothing_ratio/(num_classes-1) * torch.logical_not(one_hot_y)
    new_label = args.soft_label_ratio * new_label + (1-args.soft_label_ratio) * one_hot(data.y[data.train_mask], num_classes)

    return new_label

def postel_with_pseudo_label(data, num_classes, args, pseudo_label):
    label_marginal_prob = one_hot(torch.cat((data.y[data.train_mask], pseudo_label[(pseudo_label!=num_classes).nonzero().squeeze()])),num_classes=num_classes).mean(dim=0)
    label_distribution, label_histogram = neighborhood_label_distribution_with_pseudo_label(data, data.train_mask, data.val_mask, data.test_mask, num_classes, pseudo_label)
    hist_log_likelihood, degree_mask = neighborhood_histogram_log_likelihood(label_histogram[data.train_mask], label_distribution, args.degree_cutoff)

    new_label, mask = prob_label_from_log_likelihood(hist_log_likelihood, label_marginal_prob, args.temperature)

    mask = torch.logical_or(mask, degree_mask)
    new_label[mask] = one_hot(data.y[data.train_mask], num_classes)[mask]

    one_hot_y = one_hot(data.y[data.train_mask])
    new_label = (1-args.smoothing_ratio) * new_label + args.smoothing_ratio/(num_classes-1) * torch.logical_not(one_hot_y)
    new_label = args.soft_label_ratio * new_label + (1-args.soft_label_ratio) * one_hot(data.y[data.train_mask])

    return new_label

def neighborhood_histogram_log_likelihood(label_histogram, label_distribution, degree_cutoff):
    likelihood = []
    mask = []
    for label_hist in label_histogram:
        likeli = (label_distribution**label_hist).prod(dim=1)
        likelihood.append(torch.log(likeli))
        mask.append(label_hist.sum()<degree_cutoff)
    likelihood = torch.stack(likelihood)
    mask = torch.tensor(mask).to(label_histogram.device)

    return likelihood, mask

def prob_label_from_log_likelihood(log_likelihood, prior, temperature=1):
    log_prob = log_likelihood + torch.log(prior)
    log_prob = log_prob - log_prob.max()
    prob = torch.exp(log_prob)
    prob = torch.pow(prob, temperature)
    prob = prob / prob.sum(dim=-1, keepdim=True)
    return prob.to(torch.float32), prob.isnan().max(dim=1).values

def postel_nodewise(data, num_classes, args):
    one_hot_y = one_hot(data.y[data.train_mask])
    label_marginal_prob = one_hot_y.mean(dim=0)
    label_distribution, label_histogram = nodewise_neighborhood_label_distribution(data, data.train_mask, data.val_mask, data.test_mask, num_classes)
    hist_log_likelihood, degree_mask = neighborhood_histogram_log_likelihood(label_histogram[data.train_mask], label_distribution, args.degree_cutoff)
    new_label, mask = prob_label_from_log_likelihood(hist_log_likelihood, label_marginal_prob, args.temperature)

    mask = torch.logical_or(mask, degree_mask)

    new_label[mask] = one_hot(data.y[data.train_mask], num_classes)[mask]

    one_hot_y = one_hot(data.y[data.train_mask], num_classes)
    new_label = (1-args.smoothing_ratio) * new_label + args.smoothing_ratio/(num_classes-1) * torch.logical_not(one_hot_y)
    new_label = args.soft_label_ratio * new_label + (1-args.soft_label_ratio) * one_hot(data.y[data.train_mask], num_classes)

    return new_label

def postel_nodewise_with_pseudo_label(data, num_classes, args, pseudo_label):
    label_marginal_prob = one_hot(torch.cat((data.y[data.train_mask], pseudo_label))).mean(dim=0)
    label_distribution, label_histogram = nodewise_neighborhood_label_distribution_with_pseudo_label(data, data.train_mask, data.val_mask, data.test_mask, num_classes, pseudo_label)
    hist_log_likelihood, degree_mask = neighborhood_histogram_log_likelihood(label_histogram[data.train_mask], label_distribution, args.degree_cutoff)
    new_label, mask = prob_label_from_log_likelihood(hist_log_likelihood, label_marginal_prob, args.temperature)

    mask = torch.logical_or(mask, degree_mask)

    new_label[mask] = one_hot(data.y[data.train_mask], num_classes)[mask]

    one_hot_y = one_hot(data.y[data.train_mask], num_classes)
    new_label = (1-args.smoothing_ratio) * new_label + args.smoothing_ratio/(num_classes-1) * torch.logical_not(one_hot_y)
    new_label = args.soft_label_ratio * new_label + (1-args.soft_label_ratio) * one_hot(data.y[data.train_mask], num_classes)

    return new_label