File size: 7,386 Bytes
c91d7b1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch.optim as optim
import torch.nn as nn
import torch.nn.functional as F
from copy import deepcopy
from deeprobust.graph import utils
import torch


class BaseModel(nn.Module):
    def __init__(self):
        super(BaseModel, self).__init__()
        pass

    def fit(self, pyg_data, train_iters=1000, initialize=True, verbose=False, patience=100, **kwargs):
        if initialize:
            self.initialize()

        # self.data = pyg_data[0].to(self.device)
        self.data = pyg_data.to(self.device)
        # By default, it is trained with early stopping on validation
        self.train_with_early_stopping(train_iters, patience, verbose)

    def finetune(self, edge_index, edge_weight, feat=None, train_iters=10, verbose=True):
        if verbose:
            print(f'=== finetuning {self.name} model ===')
        optimizer = optim.Adam(self.parameters(), lr=self.lr, weight_decay=self.weight_decay)
        labels = self.data.y
        if feat is None:
            x = self.data.x
        else:
            x = feat
        train_mask, val_mask = self.data.train_mask, self.data.val_mask
        best_loss_val = 100
        best_acc_val = 0
        for i in range(train_iters):
            self.train()
            optimizer.zero_grad()
            output = self.forward(x, edge_index, edge_weight)
            loss_train = F.nll_loss(output[train_mask], labels[train_mask])
            loss_train.backward()
            optimizer.step()

            if verbose and i % 50 == 0:
                print('Epoch {}, training loss: {}'.format(i, loss_train.item()))

            self.eval()
            with torch.no_grad():
                output = self.forward(x, edge_index)
            loss_val = F.nll_loss(output[val_mask], labels[val_mask])
            acc_val = utils.accuracy(output[val_mask], labels[val_mask])

            # if best_loss_val > loss_val:
            #     best_loss_val = loss_val
            #     best_output = output
            #     weights = deepcopy(self.state_dict())

            if best_acc_val < acc_val:
                best_acc_val = acc_val
                best_output = output
                weights = deepcopy(self.state_dict())

        print('best_acc_val:', best_acc_val.item())
        self.load_state_dict(weights)
        return best_output


    def _fit_with_val(self, pyg_data, train_iters=1000, initialize=True, verbose=False, **kwargs):
        if initialize:
            self.initialize()

        # self.data = pyg_data[0].to(self.device)
        self.data = pyg_data.to(self.device)
        if verbose:
            print(f'=== training {self.name} model ===')
        optimizer = optim.Adam(self.parameters(), lr=self.lr, weight_decay=self.weight_decay)

        labels = self.data.y
        train_mask, val_mask = self.data.train_mask, self.data.val_mask

        x, edge_index = self.data.x, self.data.edge_index
        for i in range(train_iters):
            self.train()
            optimizer.zero_grad()
            output = self.forward(x, edge_index)
            loss_train = F.nll_loss(output[train_mask+val_mask], labels[train_mask+val_mask])
            loss_train.backward()
            optimizer.step()

            if verbose and i % 50 == 0:
                print('Epoch {}, training loss: {}'.format(i, loss_train.item()))

    def fit_with_val(self, pyg_data, train_iters=1000, initialize=True, patience=100, verbose=False, **kwargs):
        if initialize:
            self.initialize()

        self.data = pyg_data.to(self.device)
        self.data.train_mask = self.data.train_mask + self.data.val1_mask
        self.data.val_mask = self.data.val2_mask
        self.train_with_early_stopping(train_iters, patience, verbose)

    def train_with_early_stopping(self, train_iters, patience, verbose):
        """early stopping based on the validation loss
        """
        if verbose:
            print(f'=== training {self.name} model ===')
        optimizer = optim.Adam(self.parameters(), lr=self.lr, weight_decay=self.weight_decay)

        labels = self.data.y
        train_mask, val_mask = self.data.train_mask, self.data.val_mask

        early_stopping = patience
        best_loss_val = 100
        best_acc_val = 0
        best_epoch = 0

        x, edge_index = self.data.x, self.data.edge_index
        for i in range(train_iters):
            self.train()
            optimizer.zero_grad()

            output = self.forward(x, edge_index)

            loss_train = F.nll_loss(output[train_mask], labels[train_mask])
            loss_train.backward()
            optimizer.step()

            if verbose and i % 50 == 0:
                print('Epoch {}, training loss: {}'.format(i, loss_train.item()))

            self.eval()
            output = self.forward(x, edge_index)
            loss_val = F.nll_loss(output[val_mask], labels[val_mask])
            acc_val = utils.accuracy(output[val_mask], labels[val_mask])
            # print(acc)

            # if best_loss_val > loss_val:
            #     best_loss_val = loss_val
            #     self.output = output
            #     weights = deepcopy(self.state_dict())
            #     patience = early_stopping
            #     best_epoch = i
            # else:
            #     patience -= 1

            if best_acc_val < acc_val:
                best_acc_val = acc_val
                self.output = output
                weights = deepcopy(self.state_dict())
                patience = early_stopping
                best_epoch = i
            else:
                patience -= 1

            if i > early_stopping and patience <= 0:
                break

        if verbose:
             # print('=== early stopping at {0}, loss_val = {1} ==='.format(best_epoch, best_loss_val) )
             print('=== early stopping at {0}, acc_val = {1} ==='.format(best_epoch, best_acc_val) )
        self.load_state_dict(weights)

    def test(self):
        """Evaluate model performance on test set.
        Parameters
        ----------
        idx_test :
            node testing indices
        """
        self.eval()
        test_mask = self.data.test_mask
        labels = self.data.y
        output = self.forward(self.data.x, self.data.edge_index)
        # output = self.output
        loss_test = F.nll_loss(output[test_mask], labels[test_mask])
        acc_test = utils.accuracy(output[test_mask], labels[test_mask])
        print("Test set results:",
              "loss= {:.4f}".format(loss_test.item()),
              "accuracy= {:.4f}".format(acc_test.item()))
        return acc_test.item()

    def predict(self, x=None, edge_index=None, edge_weight=None):
        """
        Returns
        -------
        torch.FloatTensor
            output (log probabilities)
        """
        self.eval()
        if x is None or edge_index is None:
            x, edge_index = self.data.x, self.data.edge_index
        return self.forward(x, edge_index, edge_weight)

    def _ensure_contiguousness(self,
                               x,
                               edge_idx,
                               edge_weight):
        if not x.is_sparse:
            x = x.contiguous()
        if hasattr(edge_idx, 'contiguous'):
            edge_idx = edge_idx.contiguous()
        if edge_weight is not None:
            edge_weight = edge_weight.contiguous()
        return x, edge_idx, edge_weight