File size: 10,394 Bytes
d2a8669
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
import torch
import torch.nn.functional as F
from torch.autograd import Variable

import time
import numpy as np
from sklearn import metrics
from sklearn.metrics import f1_score
from tqdm import trange, tqdm
import torch.nn as nn

from CatGCN.layers import StackedGNN
from FairGNN.src.models.GCN import GCN
class ClusterGNNTrainer(object):
    """
    Training a huge graph cluster partition strategy.
    """
    def __init__(self, args, clustering_machine, neptune_run):
        self.args = args
        self.clustering_machine = clustering_machine
        self.neptune_run = neptune_run
        self.device = torch.device("cuda:{}".format(args.gpu) if torch.cuda.is_available() else "cpu")
        print('device:', self.device)
        self.class_weight = clustering_machine.class_weight.to(self.device)
        self.create_model()

        # mew part -- input?
        #self.sens_model = GCN(95, 128, 0.5) # number of feat, number of hidden, dropout percentage
        #self.adv_model = nn.Linear(128, 1)

        # adversary optimizer
        #self.optimizer_A = torch.optim.Adam(self.adv_model.parameters(), lr = args.lr, weight_decay = args.weight_decay) ## can add weight decay
        #self.criterion = nn.BCEWithLogitsLoss()
        #self.A_loss = 0

    def create_model(self):
        """
        Creating a StackedGNN and transferring to CPU/GPU.
        """
        self.model = StackedGNN(self.args, self.clustering_machine.field_count, self.clustering_machine.field_size, self.clustering_machine.class_count)
        self.model = self.model.to(self.device)

    def generate_field_adjs(self, node_count):
        # Normalization by  P'' = Q^{-1/2}*P'*Q^{-1/2}, P' = P+probe*O.
        field_adjs = torch.ones((node_count, self.clustering_machine.field_size, self.clustering_machine.field_size))
        field_adjs += self.args.diag_probe * torch.eye(self.clustering_machine.field_size)
        row_sum = self.clustering_machine.field_size + self.args.diag_probe
        field_adjs = (1. / row_sum) * field_adjs
        return field_adjs

    def do_forward_pass(self, cluster):
        """
        Making a forward pass with data from a given partition.
        :param cluster: Cluster index.
        :return average_loss: Average loss on the cluster.
        :return node_count: Number of nodes.
        """
        edges = self.clustering_machine.sg_edges[cluster].to(self.device)
        macro_nodes = self.clustering_machine.sg_nodes[cluster].to(self.device)
        train_nodes = self.clustering_machine.sg_train_nodes[cluster].to(self.device)
        field_index = self.clustering_machine.sg_field_index[cluster].to(self.device)
        field_adjs = self.generate_field_adjs(field_index.shape[0]).to(self.device)
        target = self.clustering_machine.sg_targets[cluster].to(self.device).squeeze()

        prediction = self.model(edges, field_index, field_adjs)

        # todo add forward pass for estimator and adversary


        average_loss = F.nll_loss(prediction[train_nodes], target[train_nodes], self.class_weight)
        node_count = train_nodes.shape[0]
        return average_loss, node_count

    def do_validation(self, cluster, epoch):
        """
        Making a validation with data from a given partition.
        :param cluster: Cluster index.
        :return average_loss: Average loss on the cluster.
        :return node_count: Number of nodes.
        """
        edges = self.clustering_machine.sg_edges[cluster].to(self.device)
        macro_nodes = self.clustering_machine.sg_nodes[cluster].to(self.device)
        val_nodes = self.clustering_machine.sg_val_nodes[cluster].to(self.device)
        field_index = self.clustering_machine.sg_field_index[cluster].to(self.device)
        field_adjs = self.generate_field_adjs(field_index.shape[0]).to(self.device)
        target = self.clustering_machine.sg_targets[cluster].to(self.device).squeeze()
        prediction = self.model(edges, field_index, field_adjs)
        average_loss = F.nll_loss(prediction[val_nodes], target[val_nodes], self.class_weight)
        node_count = val_nodes.shape[0]
        return average_loss, node_count

    def do_prediction(self, cluster):
        """
        Scoring a cluster.
        :param cluster: Cluster index.
        :return average_loss: Average loss on the cluster.
        :return node_count: Number of nodes.
        :return prediction: Prediction matrix with probabilities.
        :return target: Target vector.
        """
        edges = self.clustering_machine.sg_edges[cluster].to(self.device)
        macro_nodes = self.clustering_machine.sg_nodes[cluster].to(self.device)
        test_nodes = self.clustering_machine.sg_test_nodes[cluster].to(self.device)
        field_index = self.clustering_machine.sg_field_index[cluster].to(self.device)
        field_adjs = self.generate_field_adjs(field_index.shape[0]).to(self.device)
        target = self.clustering_machine.sg_targets[cluster].to(self.device).squeeze()
        prediction = self.model(edges, field_index, field_adjs)
        average_loss = F.nll_loss(prediction[test_nodes], target[test_nodes], self.class_weight)
        node_count = test_nodes.shape[0]
        target = target[test_nodes]
        prediction = prediction[test_nodes,:]

        return average_loss, node_count, prediction, target

    def update_average_loss(self, batch_average_loss, node_count):
        """
        Updating the average loss in the epoch.
        :param batch_average_loss: Loss of the cluster. 
        :param node_count: Number of nodes in currently processed cluster.
        :return average_loss: Average loss in the epoch.
        """
        self.accumulated_loss = self.accumulated_loss + batch_average_loss.item()*node_count
        self.node_count_seen = self.node_count_seen + node_count
        average_loss = self.accumulated_loss / self.node_count_seen
        return average_loss

    def train_val_test(self):
        """
        Training, validation, and test a model per epoch.
        """
        print("Training, validation, and test started.\n")
        train_start_time = time.perf_counter()
        bad_counter = 0
        best_loss = np.inf
        best_epoch = 0
        self.optimizer = torch.optim.Adam(self.model.parameters(), lr=self.args.lr, weight_decay=self.args.weight_decay)
        # test for epochs
        print(self.args.epochs)
        for epoch in range(1, self.args.epochs+1):
            epoch_start_time = time.time()
            np.random.shuffle(self.clustering_machine.clusters)
            self.model.train()
            self.node_count_seen = 0
            self.accumulated_loss = 0
            for cluster in self.clustering_machine.clusters:
                self.optimizer.zero_grad()
                batch_average_loss, node_count = self.do_forward_pass(cluster)
                batch_average_loss.backward()
                self.optimizer.step()
                average_loss = self.update_average_loss(batch_average_loss, node_count)
            train_loss = average_loss

            self.model.eval()
            self.node_count_seen = 0
            self.accumulated_loss = 0
            for cluster in self.clustering_machine.clusters:
                batch_average_loss, node_count = self.do_validation(cluster, epoch)
                average_loss = self.update_average_loss(batch_average_loss, node_count)
            val_loss = average_loss

            print("Epoch: {:04d}".format(epoch), 
                "||",
                "time cost: {:.2f}s".format(time.time() - epoch_start_time), 
                "||",
                "train loss: {:.4f}".format(train_loss),
                "val loss: {:.4f}".format(val_loss))

            if val_loss < best_loss:
                best_loss = val_loss
                best_epoch = epoch
                bad_counter = 0
                best_model_state = self.model.state_dict()
            else:
                bad_counter += 1

            if bad_counter == self.args.patience:
                break

        self.model.load_state_dict(best_model_state)
        self.model.eval()
        self.node_count_seen = 0
        self.accumulated_loss = 0
        self.predictions = []
        self.targets = []
        for cluster in self.clustering_machine.clusters:
            batch_average_loss, node_count, prediction, target = self.do_prediction(cluster)
            average_loss = self.update_average_loss(batch_average_loss, node_count)
            self.predictions.append(prediction.cpu().detach().numpy())
            self.targets.append(target.cpu().detach().numpy())
        test_loss = average_loss
        self.targets = np.concatenate(self.targets)
        self.predictions = np.concatenate(self.predictions).argmax(1)
        acc_score = metrics.accuracy_score(self.targets, self.predictions)
        macro_f1 = metrics.f1_score(self.targets, self.predictions, average="macro")
        classification_report = metrics.classification_report(self.targets, self.predictions, digits=4)
        print(classification_report)
        # Confusion matrics and AUC
        confusion_matrix = metrics.confusion_matrix(self.targets, self.predictions)
        print(confusion_matrix)
        #F1
        f1 = f1_score(self.targets, self.predictions, average='macro')
        print('F1 score:', f1)
        # fpr, tpr, _ = metrics.roc_curve(self.targets, self.predictions)
        # auc = metrics.auc(fpr, tpr)
        # print("AUC:", auc)

        train_time = (time.perf_counter() - train_start_time)/60


        print("Optimization Finished!")
        print("Total time elapsed: {:.2f}min".format(train_time))
        print("Best Result:\n",
            "best epoch: {:04d}".format(best_epoch),
            "||",
            "test loss: {:.4f}".format(test_loss),
            "||",
            "accuracy: {:.4f}".format(acc_score),
            "macro-f1: {:.4f}".format(macro_f1))        

        
        # Save results on Neptune
        self.neptune_run["best_epoch"] = best_epoch
        self.neptune_run["test/loss"] = test_loss
        self.neptune_run["test/acc"] = acc_score
        self.neptune_run["test/f1"] = macro_f1
        # self.neptune_run["test/auc"] = auc
        # self.neptune_run["test/tpr"] = tpr
        # self.neptune_run["test/fpr"] = fpr
        self.neptune_run["conf_matrix"] = confusion_matrix
        self.neptune_run["train_time"] = train_time