-
Notifications
You must be signed in to change notification settings - Fork 11
/
Copy pathtrain_test_new.py
202 lines (150 loc) · 8.33 KB
/
train_test_new.py
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
import numpy as np
import torch
import random
import torch.nn as nn
import torch.backends.cudnn as cudnn
import argparse
import logging
from torch_geometric.utils import to_dense_batch
from torch_geometric.data import DataLoader
from sklearn.model_selection import StratifiedKFold
import torch.nn.functional as F
from utils import *
from model_GAT_v4 import *
def train(opt, train_dataset, test_dataset):
cudnn.deterministic = True
torch.cuda.manual_seed_all(2020)
torch.manual_seed(2020)
random.seed(2020)
model = MLA_GNN(opt).cuda()
for name, parameters in model.named_parameters():
print(name, ':', parameters.size())
optimizer = define_optimizer(opt, model)
scheduler = define_scheduler(opt, optimizer)
train_loader = DataLoader(dataset=train_dataset, batch_size=opt.batch_size,
shuffle=True, num_workers=8)
metric_logger = {'train': {'loss': [], 'grad_acc': []}, 'test': {'loss': [], 'grad_acc': []}}
# print("============ finish dataloader ===========")
for epoch in range(opt.num_epochs):
model.train()
loss_epoch, grad_acc_epoch = 0, 0
class_edge_weights = torch.zeros(opt.label_dim, opt.num_nodes, opt.num_nodes).cuda()
class_node_importance = torch.zeros(opt.label_dim, opt.num_nodes).cuda()
overall_edge_weights = torch.zeros(opt.num_nodes, opt.num_nodes).cuda()
overall_node_importance = torch.zeros(opt.num_nodes).cuda()
# print(class_edge_weights.shape, class_node_importance.shape)
# print(overall_edge_weights.shape, overall_node_importance.shape)
for batch_idx, data in enumerate(train_loader):
batch_idx += 1
tr_features, tr_fc_features, tr_preds, edge_weights, feature_importance = model(
data.x.cuda(), data.edge_index.cuda(), data.y.cuda(), data.batch.cuda(), opt)
# print(data.y)
sample_weight = torch.tensor(cal_sample_weight(
data.y, opt.label_dim, use_sample_weight=True)).cuda()
# print("===========", data.y, sample_weight)
"""
Compute the edge weights and node importance for each class.
"""
overall_edge_weights += torch.mean(edge_weights, 0)
overall_node_importance += torch.mean(feature_importance, 0)
for i in range(opt.label_dim):
index = np.nonzero(data.y == i).view(1, -1)[0]
# print(torch.index_select(edge_weights, 0, index[0].cuda()).shape)
if index.shape[0] > 0:
# print(index.shape[0])
class_edge_weights[i] += torch.mean(
torch.index_select(edge_weights, 0, index.cuda()), 0)
class_node_importance[i] += torch.mean(
torch.index_select(feature_importance, 0, index.cuda()), 0)
# print(class_node_importance[i])
loss_reg = define_reg(model)
loss_func = nn.CrossEntropyLoss(reduction='none')
grad_loss = torch.mean(torch.mul(loss_func(tr_preds, data.y.cuda()), sample_weight))
# print(grad_loss)
# one_hot_labels = one_hot_tensor(batch_labels.cuda(), opt.label_dim)
# grad_loss = torch.mean(torch.mul(tr_preds - one_hot_labels, tr_preds - one_hot_labels))
loss = opt.lambda_nll * grad_loss + opt.lambda_reg * loss_reg
loss_epoch += loss.data.item()
optimizer.zero_grad()
# tr_features.retain_grad()
loss.backward(retain_graph=True)
optimizer.step()
pred = tr_preds.argmax(dim=1, keepdim=True)
grad_acc_epoch += pred.eq(data.y.cuda().view_as(pred)).sum().item()
scheduler.step()
class_edge_weights = (class_edge_weights/batch_idx).detach().cpu().numpy()
class_node_importance = (class_node_importance/batch_idx).detach().cpu().numpy()
overall_edge_weights = (overall_edge_weights/batch_idx).detach().cpu().numpy()
overall_node_importance = (overall_node_importance/batch_idx).detach().cpu().numpy()
# print(gradients_all.shape, importance_all.shape)
# print(class_edge_weights, class_node_importance)
loss_epoch /= len(train_loader.dataset)
grad_acc_epoch = grad_acc_epoch / len(train_loader.dataset)
loss_test, grad_acc_test, pred_test, features_test, _ = test(opt, model, test_dataset)
# loss_test, grad_acc_test, pred_test, features_test, _ = test(opt, model, train_dataset)
metric_logger['train']['loss'].append(loss_epoch)
metric_logger['train']['grad_acc'].append(grad_acc_epoch)
metric_logger['test']['loss'].append(loss_test)
metric_logger['test']['grad_acc'].append(grad_acc_test)
# pickle.dump(pred_test, open(os.path.join(
# opt.results_dir, 'split%d_%d_pred_test.pkl' % (k, epoch)), 'wb'))
logging.info('\nEpoch {:02d}/{:02d}, [{:s}]\t\tLoss: {:.4f}, {:s}: {:.4f}'.format(
epoch + 1, opt.num_epochs, 'Train', loss_epoch, 'Accuracy', grad_acc_epoch))
logging.info('\nEpoch {:02d}/{:02d}, [{:s}]\t\tLoss: {:.4f}, {:s}: {:.4f}\n'.format(
epoch + 1, opt.num_epochs, 'Test', loss_test, 'Accuracy', grad_acc_test))
# print("=========gradients_all:", gradients_all)
return model, optimizer, metric_logger, class_edge_weights, class_node_importance, \
overall_edge_weights, overall_node_importance
def test(opt, model, test_dataset):
model.eval()
test_loader = DataLoader(dataset=test_dataset, batch_size=opt.batch_size, shuffle=False)
probs_all, gt_all = None, np.array([])
loss_test, grad_acc_test = 0, 0
for batch_idx, (data) in enumerate(test_loader):
te_features, te_fc_features, te_preds, _, _ = model(
data.x.cuda(), data.edge_index.cuda(), data.y.cuda(), data.batch.cuda(), opt)
# print(te_preds, te_fc_features)
# print(data.y, torch.argmax(te_preds, dim=1))
# print(data.y, te_preds)
# print("te_preds:", te_preds)
if batch_idx == 0:
features_all = te_features.detach().cpu().numpy()
fc_features_all = te_fc_features.detach().cpu().numpy()
else:
features_all = np.concatenate((features_all, te_features.detach().cpu().numpy()), axis=0)
fc_features_all = np.concatenate((fc_features_all, te_fc_features.detach().cpu().numpy()), axis=0)
# print(features_all.shape, te_features.shape)
loss_reg = define_reg(model)
loss_func = nn.CrossEntropyLoss()
grad_loss = loss_func(te_preds, data.y.cuda())
# one_hot_labels = one_hot_tensor(batch_labels.cuda(), opt.label_dim)
# grad_loss = torch.mean(torch.mul(te_preds - one_hot_labels, te_preds - one_hot_labels))
loss = opt.lambda_nll * grad_loss + opt.lambda_reg * loss_reg
loss_test += loss.data.item()
gt_all = np.concatenate((gt_all, data.y.reshape(-1))) # Logging Information
pred = te_preds.argmax(dim=1, keepdim=True)
grad_acc_test += pred.eq(data.y.cuda().view_as(pred)).sum().item()
probs_np = te_preds.detach().cpu().numpy()
probs_all = probs_np if probs_all is None else np.concatenate((probs_all, probs_np), axis=0) # Logging Information
# print("total batch:", batch_idx)
# print(survtime_all)
###################################################
# ==== Measuring Test Loss, C-Index, P-Value ==== #
###################################################
loss_test /= len(test_loader.dataset)
grad_acc_test = grad_acc_test / len(test_loader.dataset)
pred_test = [probs_all, gt_all]
# print(probs_all.shape, gt_all.shape)
return loss_test, grad_acc_test, pred_test, features_all, fc_features_all
def external_test(opt, model, test_dataset):
model.eval()
test_loader = DataLoader(dataset=test_dataset, batch_size=opt.batch_size, shuffle=False)
probs_all = None
for batch_idx, (data) in enumerate(test_loader):
te_features, te_fc_features, te_preds, _, _ = model(
data.x.cuda(), data.edge_index.cuda(), data.y.cuda(), data.batch.cuda(), opt)
# print("features:", te_features)
probs_np = F.softmax(te_preds, 1).detach().cpu().numpy()
probs_all = probs_np if probs_all is None else np.concatenate((probs_all, probs_np), axis=0)
# print(probs_all)
return probs_all