-
Notifications
You must be signed in to change notification settings - Fork 0
/
Crogen.py
251 lines (207 loc) · 9.94 KB
/
Crogen.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
import sys
import argparse
import os
import numpy as np
import math
import pickle
import warnings
import random
import itertools
from model import *
from dataset import BpDataset, HivDataset, PpmiDataset, Ppmi622Dataset
from utils import *
from torch.autograd import Variable
import torch.nn as nn
import torch.nn.functional as F
import torch
from evaluate import print_stats, evaluate, clean_lists
warnings.filterwarnings("ignore")
parser = argparse.ArgumentParser()
parser.add_argument("--n_epochs", type=int, default=30, help="number of epochs of training")
parser.add_argument("--to1_epochs", type=int, default=30, help="number of epochs of training")
parser.add_argument("--gamma", type=int, default=15, help="number of epochs of training")
parser.add_argument("--gamma_cyc", type=int, default=5, help="number of epochs of training")
parser.add_argument("--batch_size", type=int, default=1, help="size of the batches")
parser.add_argument("--lr1", type=float, default=0.00006 , help="adam: learning rate")
parser.add_argument("--lr2", type=float, default=0.00006, help="adam: learning rate")
parser.add_argument("--dlr1", type=float, default=0.001, help="adam: learning rate discriminatro")
parser.add_argument("--dlr2", type=float, default=0.001, help="adam: learning rate dicriminator2")
parser.add_argument('--alpha', type=float, default=0.1, help='alpha')
parser.add_argument("--b1", type=float, default=0.9, help="adam: decay of first order momentum of gradient")
parser.add_argument("--b2", type=float, default=0.999, help="adam: decay of first order momentum of gradient")
parser.add_argument("--k_fold", type=int, default=5, help="size of each image dimension")
parser.add_argument("--dropout", type=float, default=0.3, help="...")
parser.add_argument("--dataset", type=str, default="bp", help="dataset: choose dataset by name")
parser.add_argument("--generator", type=str, default="cat", help="generator: choose generator by name")
parser.add_argument('--d_size', type=int, default=32, help="generator input size")
parser.add_argument('--h_size', type=int, default=64, help="generator hidden size")
parser.add_argument('--o_size', type=int, default=64, help="generator output size")
parser.add_argument('--rep_size', type=int, default=32, help="generator output size")
parser.add_argument('--is_neg', type=bool, default=False, help="is negative")
parser.add_argument('--load', type=bool, default=False, help="load saved model")
parser.add_argument("--agg_way", type=str, default="cat", help="aggregation way")
parser.add_argument('--seed', type=int, default=100, help="random seed")
parser.add_argument('--cuda_seed', type=int, default=345, help="random cuda seed")
opt = parser.parse_args()
print(opt)
random.seed(opt.seed)
np.random.seed(opt.seed)
# torch.manual_seed(opt.seed)
# torch.cuda.manual_seed(opt.cuda_seed)
#torch.cuda.manual_seed_all(opt.cuda_seed)
cuda = True if torch.cuda.is_available() else False
# Loss function
adversarial_loss = torch.nn.BCELoss()
mse_loss = torch.nn.MSELoss()
criterion_recon = torch.nn.MSELoss()
criterion_cycle = torch.nn.L1Loss()
# Initialize generator and discriminator
generator_1to2 = Generator(opt.d_size, 64, False, opt.generator)
generator_2to1 = Generator(32, opt.h_size, False, opt.generator)
discriminator1 = Discriminator(opt.d_size, opt.h_size, opt.rep_size, opt.dropout)
discriminator2 = Discriminator(opt.d_size, opt.h_size, opt.rep_size, opt.dropout)
if opt.load:
model_path = f'./checkpoint/cycle_{opt.generator}/'
generator_1to2.encoder = pickle.load(open(model_path + 'g1to2_encoder', 'rb'))
generator_1to2.decoder = pickle.load(open(model_path + 'g2to1_decoder', 'rb'))
generator_2to1.encoder = pickle.load(open(model_path + 'g2to1_encoder', 'rb'))
generator_2to1.decoder = pickle.load(open(model_path + 'g1to2_decoder', 'rb'))
discriminator1 = pickle.load(open(model_path + 'discriminator1', 'rb'))
discriminator2 = pickle.load(open(model_path + 'discriminator2', 'rb'))
print('Load pre-train model successfully.')
if cuda:
generator_1to2.cuda()
generator_2to1.cuda()
discriminator1.cuda()
discriminator2.cuda()
# Optimizers
optimizer_G = torch.optim.Adam(
itertools.chain(generator_1to2.parameters(), generator_2to1.parameters()),
lr=opt.lr1,
betas=(opt.b1, opt.b2)
)
optimizer_d1 = torch.optim.Adam(discriminator1.parameters(), lr=opt.dlr1, betas=(opt.b1, opt.b2))
optimizer_d2 = torch.optim.Adam(discriminator2.parameters(), lr=opt.dlr2, betas=(opt.b1, opt.b2))
Tensor = torch.cuda.FloatTensor if cuda else torch.FloatTensor
# ----------
# Training
# ----------
if opt.dataset == 'bp':
print('load bp dataset')
dataset = BpDataset(train='train', K=opt.k_fold)
elif opt.dataset == 'hiv':
print('load hiv dataset')
dataset = HivDataset(train='train')
elif opt.dataset == 'ppmi':
print('load ppmi dataset')
dataset = PpmiDataset(train='train')
elif opt.dataset == 'ppmi_622':
print('load ppmi_622 dataset')
dataset = Ppmi622Dataset(train='train')
else:
print('valid dataset')
dataTrain = torch.utils.data.DataLoader(
dataset=dataset,
batch_size=opt.batch_size,
shuffle=True,
)
for epoch in range(opt.n_epochs):
generator_1to2.train(); generator_2to1.train()
discriminator1.train(); discriminator2.train()
rec1_list, rec2_list, disc1_list, disc2_list, gen1_disc_list, gen2_disc_list = [], [], [], [], [], []
cycle1_list, cycle2_list = [], []
gen_list = []
for i, (views_A, views_B, labels) in enumerate(dataTrain):
views_A, views_B, labels = views_A[0], views_B[0], labels[0]
views_A_edge_num = torch.sum(views_A != 0).item()
views_B_edge_num = torch.sum(views_B != 0).item()
# Adversarial ground truths
valid = Variable(torch.ones(1), requires_grad=False).cuda()
fake = Variable(torch.zeros(1), requires_grad=False).cuda()
# Configure input
real_views_A = Variable(views_A.type(Tensor)).cuda()
real_views_B = Variable(views_B.type(Tensor)).cuda()
# -----------------
# Train Generator
# -----------------
optimizer_G.zero_grad()
gen_views_B = generator_1to2(real_views_A, real_views_A)
gen_views_B = topk_adj(gen_views_B, views_B_edge_num)
gen_views_A = generator_2to1(real_views_B, real_views_B)
gen_views_A = topk_adj(gen_views_A, views_A_edge_num)
# GAN loss
disc_AB = discriminator2(gen_views_B, gen_views_B)
disc_BA = discriminator1(gen_views_A, gen_views_A)
gan_loss_AB = adversarial_loss(disc_AB, valid)
gan_loss_BA = adversarial_loss(disc_BA, valid)
gan_loss = (gan_loss_AB + gan_loss_BA) / 2
gen2_disc_list.append(disc_AB.data)
gen1_disc_list.append(disc_BA.data)
# reconstruct loss
rec_loss_AB = criterion_recon(gen_views_B, real_views_B)
rec_loss_BA = criterion_recon(gen_views_A, real_views_A)
rec_loss = (rec_loss_AB + rec_loss_BA) / 2
rec1_list.append(rec_loss_AB.data)
rec2_list.append(rec_loss_BA.data)
cyc_views_A = generator_2to1(gen_views_B, gen_views_B)
cyc_views_A = topk_adj(cyc_views_A, views_A_edge_num)
cyc_views_B = generator_1to2(gen_views_A, gen_views_A)
cyc_views_B = topk_adj(cyc_views_B, views_B_edge_num)
# cycle loss
cycle_loss_A = criterion_cycle(cyc_views_A, real_views_A)
cycle_loss_B = criterion_cycle(cyc_views_B, real_views_B)
cycle_loss = (cycle_loss_A + cycle_loss_B) / 2
cycle1_list.append(cycle_loss_A)
cycle2_list.append(cycle_loss_B)
gen_loss = gan_loss + opt.gamma * rec_loss + opt.gamma_cyc * cycle_loss
if epoch % 2 == 0:
gen_loss.backward()
optimizer_G.step()
# ---------------------
# Train Discriminator_1
# ---------------------
optimizer_d1.zero_grad()
d_out = discriminator1(real_views_A, real_views_A)
real_loss = adversarial_loss(d_out, valid)
# print("real {}".format(d_out))
disc1_list.append(real_loss.data)
d_out = discriminator1(gen_views_A.detach(), gen_views_A.detach())
fake_loss = adversarial_loss(d_out, fake)
# print("fake {}".format(d_out))
disc1_list.append(fake_loss.data)
d_loss = real_loss + fake_loss
d_loss.backward()
optimizer_d1.step()
# ---------------------
# Train Discriminator_2
# ---------------------
optimizer_d2.zero_grad()
d_out = discriminator2(real_views_B, real_views_B)
real_loss = adversarial_loss(d_out, valid)
disc2_list.append(real_loss.data)
d_out = discriminator2(gen_views_B.detach(), gen_views_B.detach())
fake_loss = adversarial_loss(d_out, fake)
disc2_list.append(fake_loss.data)
d_loss = real_loss + fake_loss
d_loss.backward()
optimizer_d2.step()
print("[Epoch %d/%d] [rec1 loss: %f] [rec2 loss: %f] [cyc1: %f] [cyc2: %f] [dis1: %f] [dis2: %f] [gen1: %f] [gen2: %f]"
% (
epoch, opt.n_epochs,
torch.mean(torch.stack(rec1_list)),
torch.mean(torch.stack(rec2_list)),
torch.mean(torch.stack(cycle1_list)),
torch.mean(torch.stack(cycle2_list)),
torch.mean(torch.stack(disc1_list)),
torch.mean(torch.stack(disc2_list)),
torch.mean(torch.stack(gen1_disc_list)),
torch.mean(torch.stack(gen2_disc_list)),
)
)
#val(opt, generator_1to2, generator_2to1)
# test(opt, generator_1to2, generator_2to1)
pretrain = 'pretrain_' if opt.load else ''
if not os.path.exists(f'./checkpoint/Experiments/Mygan_cycle_{pretrain}{opt.generator}/'):
os.mkdir(f'./checkpoint/Experiments/Mygan_cycle_{pretrain}{opt.generator}/')
pickle.dump(generator_1to2, open(f'./checkpoint/Experiments/Mygan_cycle_{pretrain}{opt.generator}/generator_1to2', 'wb'))
pickle.dump(generator_2to1, open(f'./checkpoint/Experiments/Mygan_cycle_{pretrain}{opt.generator}/generator_2to1', 'wb'))