-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
358 lines (329 loc) · 19.1 KB
/
train.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
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
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
import os
import torch
from torch.optim import Adam
import torch.nn as nn
import torch.nn.functional as F
from sklearn.cluster import KMeans
import opt
from util import *
from miniBatch import *
from MyLoss import *
import warnings
warnings.filterwarnings('ignore')
# ===================多进程调用（一次数据集加载，模型共享）===================
from model import *
def parallel_model(model):
if model == 1:
# 1. 分开学 + 线性加
model = MultiAttributedModel_Iso_Combine(gae_n_enc_1=opt.args.gae_n_enc_1, gae_n_enc_2=opt.args.gae_n_enc_2,
gae_n_enc_3=opt.args.gae_n_enc_3,
gae_n_dec_1=opt.args.gae_n_dec_1, gae_n_dec_2=opt.args.gae_n_dec_2,
gae_n_dec_3=opt.args.gae_n_dec_3,
n_input=opt.args.input_dim, n_clusters=opt.args.n_clusters, v=1,
n_nodes=opt.args.n_nodes, device=opt.args.device).to(opt.args.device)
opt.args.model = 1
elif model == 2:
# 2. 分开学 + 逐层加
model = MultiAttributedModel_Iso_Merge(gae_n_enc_1=opt.args.gae_n_enc_1, gae_n_enc_2=opt.args.gae_n_enc_2,
gae_n_enc_3=opt.args.gae_n_enc_3,
gae_n_dec_1=opt.args.gae_n_dec_1, gae_n_dec_2=opt.args.gae_n_dec_2,
gae_n_dec_3=opt.args.gae_n_dec_3,
n_input=opt.args.input_dim, n_clusters=opt.args.n_clusters, v=1,
n_nodes=opt.args.n_nodes, device=opt.args.device).to(opt.args.device)
opt.args.model = 2
elif model == 3: # opt.args.model == 3
# 3. 拼接学 + 普通GCN(topo_adj)
model = MultiAttributedModel_Concatenate(gae_n_enc_1=opt.args.gae_n_enc_1, gae_n_enc_2=opt.args.gae_n_enc_2,
gae_n_enc_3=opt.args.gae_n_enc_3,
gae_n_dec_1=opt.args.gae_n_dec_1, gae_n_dec_2=opt.args.gae_n_dec_2,
gae_n_dec_3=opt.args.gae_n_dec_3,
n_input=opt.args.input_dim, n_clusters=opt.args.n_clusters, v=1,
n_nodes=opt.args.n_nodes, device=opt.args.device).to(opt.args.device)
opt.args.model = 3
print("opt.args.model = {}, 已完成校准！".format(opt.args.model))
return model
# ===================================================================================
def train(model, adj_label, topology_adj, geo_adj, text_adj,
topology_feature, geo_feature, text_feature,
pca_feature, label,
id_map, context_pairs_list,
device):
print("Using train method!")
# ===================多进程调用（一次数据集加载，模型共享）===================
# model = parallel_model(model)
# acc_result = []
# nmi_result = []
# ari_result = []
# f1_result = []
acc_best = None
nmi_best = None
ari_best = None
f1_best = None
optimizer = Adam(model.parameters(), lr=opt.args.lr)
with torch.no_grad():
model_cpu = model.cpu()
# model_cpu = model.module.cpu() # 多卡并行解除DataParallel
model_cpu.device = "cpu"
if opt.args.model != 3:
topology_feature = topology_feature.cpu()
geo_feature = geo_feature.cpu()
text_feature = text_feature.cpu()
topology_adj = topology_adj.cpu()
geo_adj = geo_adj.cpu()
text_adj = text_adj.cpu()
z, q = model_cpu(topology_feature, geo_feature, text_feature, topology_adj, geo_adj, text_adj)
# z, q = model(topology_feature, geo_feature, text_feature, topology_adj, geo_adj, text_adj)
else:
pca_feature = pca_feature.cpu()
topology_adj = topology_adj.cpu()
z, q = model_cpu(pca_feature, topology_adj)
# z, q = model(pca_feature, topology_adj)
p = target_distribution(q.data).to(opt.args.device)
kmeans = KMeans(n_clusters=opt.args.n_clusters, n_init=20)
y_pred = kmeans.fit_predict(z.data.cpu().numpy())
# kmeans.cluster_centers_ 从z中获得初始聚类中心
model.cluster_layer.data = torch.tensor(kmeans.cluster_centers_).to(device)
# model.module.cluster_layer.data = torch.tensor(kmeans.cluster_centers_).to(device) # 多卡并行解除DataParallel
eva(label, y_pred, 'init')
# TODO 新增可视化
from visualizer import Visualizer
# smallBatch_XXX(A)、continue_XXX、BsmallBatch_XXX(B)
# reconstruct_XXX
# KL_XXX
visualizer = Visualizer(visdom_port=opt.args.visdom_port,
env_name="KL_{}_{}_model{}".format(opt.args.dataset, opt.args.order, opt.args.model))
# TODO ---------分批处理------------
# miniBatch = EdgeMinibatchIterator(id_map=id_map,
# context_pairs_list=context_pairs_list,
# batch_size=opt.args.batch_size)
# myloss = MyLoss()
for epoch in range(opt.args.epoch):
# ==============训练阶段======================
model.train()
model = model.to(opt.args.device)
model.device = "cuda"
loss = 0
if opt.args.model != 3:
topology_feature = topology_feature.to(opt.args.device)
geo_feature = geo_feature.to(opt.args.device)
text_feature = text_feature.to(opt.args.device)
topology_adj = topology_adj.to(opt.args.device)
geo_adj = geo_adj.to(opt.args.device)
text_adj = text_adj.to(opt.args.device)
z, q = model(topology_feature, geo_feature, text_feature, topology_adj, geo_adj, text_adj)
else:
pca_feature = pca_feature.to(opt.args.device)
topology_adj = topology_adj.to(opt.args.device)
z, q = model(pca_feature, topology_adj)
# 基础loss
# KL
loss = F.kl_div(q.log(), p, reduction='batchmean')
# # A
# loss = F.binary_cross_entropy(torch.sigmoid(torch.mm(z, z.t())).view(-1), adj_label.to_dense().view(-1))
optimizer.zero_grad()
loss.backward()
optimizer.step()
# # TODO ----------分批处理-----------
#
# # 打乱训练的点边，并将batch_num重置为0
# miniBatch.shuffle()
# while not miniBatch.end():
# # 小批量加载
# feed_dict = miniBatch.next_minibatch()
# batch_size = feed_dict['batch_size']
# pos1 = feed_dict['batch1']
# pos2 = feed_dict['batch2']
# pos_sim = feed_dict['batch3']
# nnegs = feed_dict['batch4']
# nnegs_sim = feed_dict['batch5']
#
# optimizer.zero_grad()
# # 模型
# if opt.args.model != 3:
# # TODO adj为sp
# z_pos1, q_pos1 = model(topology_feature[pos1],
# geo_feature[pos1],
# text_feature[pos1],
# tensor_to_sparse_tensor(topology_adj.to_dense()[pos1,:][:, pos1]),
# tensor_to_sparse_tensor(geo_adj.to_dense()[pos1,:][:, pos1]),
# tensor_to_sparse_tensor(text_adj.to_dense()[pos1,:][:, pos1]))
# z_pos2, q_pos2 = model(topology_feature[pos2],
# geo_feature[pos2],
# text_feature[pos2],
# tensor_to_sparse_tensor(topology_adj.to_dense()[pos2, :][:, pos2]),
# tensor_to_sparse_tensor(geo_adj.to_dense()[pos2, :][:, pos2]),
# tensor_to_sparse_tensor(text_adj.to_dense()[pos2, :][:, pos2]))
# # 稀疏矩阵转换成稠密矩阵，因为并行计算不能用稀疏切片
# # z_pos1, q_pos1 = model(topology_feature[pos1],
# # geo_feature[pos1],
# # text_feature[pos1],
# # topology_adj.to_dense()[pos1, :][:, pos1],
# # geo_adj.to_dense()[pos1, :][:, pos1],
# # text_adj.to_dense()[pos1, :][:, pos1])
# # z_pos2, q_pos2 = model(topology_feature[pos2],
# # geo_feature[pos2],
# # text_feature[pos2],
# # topology_adj.to_dense()[pos2, :][:, pos2],
# # geo_adj.to_dense()[pos2, :][:, pos2],
# # text_adj.to_dense()[pos2, :][:, pos2])
# # z_pos1, q_pos1 = model(topology_feature[pos1],
# # geo_feature[pos1],
# # text_feature[pos1],
# # topology_adj[pos1, :][:, pos1],
# # geo_adj[pos1, :][:, pos1],
# # text_adj[pos1, :][:, pos1])
# # z_pos2, q_pos2 = model(topology_feature[pos2],
# # geo_feature[pos2],
# # text_feature[pos2],
# # topology_adj[pos2, :][:, pos2],
# # geo_adj[pos2, :][:, pos2],
# # text_adj[pos2, :][:, pos2])
# def integrate(neg_sample):
# # TODO adj为sp
# z_neg, q_neg = model(topology_feature[neg_sample],
# geo_feature[neg_sample],
# text_feature[neg_sample],
# tensor_to_sparse_tensor(topology_adj.to_dense()[neg_sample, :][:, neg_sample]),
# tensor_to_sparse_tensor(geo_adj.to_dense()[neg_sample, :][:, neg_sample]),
# tensor_to_sparse_tensor(text_adj.to_dense()[neg_sample, :][:, neg_sample]))
# # 稀疏矩阵转换成稠密矩阵，因为并行计算不能用稀疏切片
# # z_neg, q_neg = model(topology_feature[neg_sample],
# # geo_feature[neg_sample],
# # text_feature[neg_sample],
# # topology_adj.to_dense()[neg_sample, :][:, neg_sample],
# # geo_adj.to_dense()[neg_sample, :][:, neg_sample],
# # text_adj.to_dense()[neg_sample, :][:, neg_sample])
# # z_neg, q_neg = model(topology_feature[neg_sample],
# # geo_feature[neg_sample],
# # text_feature[neg_sample],
# # topology_adj[neg_sample, :][:, neg_sample],
# # geo_adj[neg_sample, :][:, neg_sample],
# # text_adj[neg_sample, :][:, neg_sample])
# return [z_neg, q_neg]
#
# # 负样本 nnegs.shape=(X, 20, dim), 所以是20个20个分别进去integrate
# # neg_outputs.shape = (X, 20, output_dim*2)
# neg_outputs = [integrate(x) for x in nnegs]
# z_neg = torch.stack([tmp[0] for tmp in neg_outputs], dim=0).float()
# # q_neg = torch.stack([tmp[1] for tmp in neg_outputs], dim=0).float()
#
# else:
# # TODO adj为sp
# z_pos1, q_pos1 = model(pca_feature[pos1],
# tensor_to_sparse_tensor(topology_adj.to_dense()[pos1,:][:, pos1]))
# z_pos2, q_pos2 = model(pca_feature[pos2],
# tensor_to_sparse_tensor(topology_adj.to_dense()[pos2, :][:, pos2]))
# # 稀疏矩阵转换成稠密矩阵，因为并行计算不能用稀疏切片
# # z_pos1, q_pos1 = model(pca_feature[pos1],
# # topology_adj.to_dense()[pos1, :][:, pos1])
# # z_pos2, q_pos2 = model(pca_feature[pos2],
# # topology_adj.to_dense()[pos2, :][:, pos2])
# # z_pos1, q_pos1 = model(pca_feature[pos1],
# # topology_adj[pos1, :][:, pos1])
# # z_pos2, q_pos2 = model(pca_feature[pos2],
# # topology_adj[pos2, :][:, pos2])
# def integrate(neg_sample):
# # TODO adj为sp
# z_neg, q_neg = model(pca_feature[neg_sample],
# tensor_to_sparse_tensor(topology_adj.to_dense()[neg_sample,:][:, neg_sample]))
# # 稀疏矩阵转换成稠密矩阵，因为并行计算不能用稀疏切片
# # z_neg, q_neg = model(pca_feature[neg_sample],
# # topology_adj.to_dense()[neg_sample, :][:, neg_sample])
# # z_neg, q_neg = model(pca_feature[neg_sample],
# # topology_adj[neg_sample,:][:, neg_sample])
# return [z_neg, q_neg]
#
# # 负样本 nnegs.shape=(X, 20, dim), 所以是20个20个分别进去integrate
# # neg_outputs.shape = (X, 20, output_dim)
# neg_outputs = [integrate(x) for x in nnegs]
# z_neg = torch.stack([tmp[0] for tmp in neg_outputs], dim=0).float()
# # q_neg = torch.stack([tmp[1] for tmp in neg_outputs], dim=0).float()
#
#
# # TODO:正负样本损失计算的比例问题：负样本会不会因为数量多而占比过大？（已修正）
# cur_loss = myloss(z_pos1, z_pos2, z_neg, pos_sim, nnegs_sim) / batch_size
# loss += cur_loss.item()
# print("epoch = {}, {}/{}, cur loss = {}, loss = {}".format(epoch, miniBatch.batch_num, len(miniBatch),
# cur_loss.item(), loss))
# # TODO 1
# save_directory = '/home/laixy/work2/output/subgraph/{}/{}/BsmallBatch'.format(opt.args.dataset,
# opt.args.order)
# if not os.path.exists(save_directory):
# os.makedirs(save_directory)
# torch.save(model.state_dict(),
# save_directory + "/model{}_last_model.pth"
# .format(opt.args.model))
#
# # loss = loss + myloss(z_pos1, z_pos2, z_neg, pos_sim, nnegs_sim)
# # loss = loss / batch_size
# # print("epoch = {}, {}/{}, loss = {}".format(epoch, miniBatch.batch_num, len(miniBatch),
# # loss.item()))
# # loss.backward()
# cur_loss.backward()
# optimizer.step()
# # optimizer.zero_grad() # 清零梯度
print("epoch = {}, total loss = {}".format(epoch, loss))
# TODO 保存模型
# TODO 2
save_directory = '/home/laixy/work2/output/subgraph/{}/{}/KL'.format(opt.args.dataset, opt.args.order)
if not os.path.exists(save_directory):
os.makedirs(save_directory)
torch.save(model.state_dict(), save_directory + "/model{}_epoch{}.pth".format(opt.args.model, epoch))
# ==============评估阶段======================
if epoch % opt.args.upd == 0:
model.eval()
model_cpu = model.cpu()
# model_cpu = model.module.cpu() # 多卡并行解除DataParallel
model_cpu.device = "cpu"
if opt.args.model != 3:
topology_feature = topology_feature.cpu()
geo_feature = geo_feature.cpu()
text_feature = text_feature.cpu()
topology_adj = topology_adj.cpu()
geo_adj = geo_adj.cpu()
text_adj = text_adj.cpu()
z, q = model_cpu(topology_feature, geo_feature, text_feature, topology_adj, geo_adj, text_adj)
else:
pca_feature = pca_feature.cpu()
topology_adj = topology_adj.cpu()
z, q = model_cpu(pca_feature, topology_adj)
p = target_distribution(q.data).to(opt.args.device)
# TODO
# 聚类1
kmeans = KMeans(n_clusters=opt.args.n_clusters, n_init=20).fit(z.data.cpu().numpy())
res = kmeans.labels_
# 聚类2
# res = q.data.cpu().numpy().argmax(1) # Q
acc, nmi, ari, f1 = eva(label, res, str(epoch))
if acc != -1:
# acc, nmi, ari, f1 = eva(label, kmeans.labels_, epoch)
# acc_result.append(acc)
# nmi_result.append(nmi)
# ari_result.append(ari)
# f1_result.append(f1)
if acc_best is None:
acc_best = acc
nmi_best = nmi
ari_best = ari
f1_best = f1
# TODO 新增
if acc >= acc_best:
acc_best = acc
nmi_best = nmi
ari_best = ari
f1_best = f1
print("maxAcc update! ===> epoch = {}, acc = {}, nmi = {}, ari = {}, f1 = {}".format(epoch, acc_best, nmi_best, ari_best, f1_best))
# TODO 3
save_directory = '/home/laixy/work2/output/subgraph/{}/{}/KL'.format(opt.args.dataset,
opt.args.order)
if not os.path.exists(save_directory):
os.makedirs(save_directory)
torch.save(model.state_dict(),
save_directory + "/model{}_best_model.pth"
.format(opt.args.model))
# TODO 新增
visualizer.plot_acc(epoch, acc)
# visualizer.plot_loss(epoch, loss) # 正负样本交叉熵
visualizer.plot_loss(epoch, loss.item()) # 其他损失函数
# return np.mean(acc_result), np.mean(nmi_result), np.mean(ari_result), np.mean(f1_result), len(acc_result)
return acc_best, nmi_best, ari_best, f1_best