-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathtrain.py
More file actions
255 lines (212 loc) · 12.4 KB
/
train.py
File metadata and controls
255 lines (212 loc) · 12.4 KB
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
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
from torch.optim.lr_scheduler import StepLR
from torch.autograd import Variable
import torch.autograd as autograd
import matplotlib.pyplot as plt
import random
import numpy as np
import time
from model import Encoder, Decoder, Encoder_Mnist, Decoder_Mnist
from dataset import inf_train_gen, mnist_loader
from loss import imq_kernel, rbf_kernel, jenson_shannon_divergence
import config
import argparse
import os
import sys
def train_g(gps, js, exp, beta, gauss, mnist):
#-------initialising device--------
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print(device)
if mnist == 0:
#------data loading--------
if config.data_pts_load == False:
data_all = inf_train_gen(data_pts)
torch.save(torch.tensor(data_all),'datapts.pth')
print('data generated')
elif config.data_pts_load == True:
data_all = torch.load('datapts.pth').numpy()
print('data loaded')
for i in config.n_samples:
for model_num in range(config.total_epoch):
rec_loss = []
wae_cust_loss = []
tot_loss = []
rn = np.random.choice(config.total_corpus, i)
data = data_all[rn]
print(data.shape)
encoder, decoder = Encoder(), Decoder()
pytorch_total_params = sum(p.numel() for p in encoder.parameters())
print(pytorch_total_params)
pytorch_total_params = sum(p.numel() for p in decoder.parameters())
print(pytorch_total_params)
criterion = nn.MSELoss()
criterion.to(device)
encoder.to(device)
decoder.to(device)
encoder.train()
decoder.train()
enc_optim = optim.Adam(encoder.parameters(), lr= config.lr)
dec_optim = optim.Adam(decoder.parameters(), lr= config.lr)
enc_scheduler = StepLR(enc_optim, step_size=30, gamma=0.5)
dec_scheduler = StepLR(dec_optim, step_size=30, gamma=0.5)
if torch.cuda.is_available():
encoder, decoder = encoder.cuda(), decoder.cuda()
data = torch.tensor(data)
ub_list = []
ub_list = np.arange(0, data.shape[0] + config.size_of_batch_g, config.size_of_batch_g)
for epoch in range(config.ITERS_g):
total_steps = data.shape[0] // config.size_of_batch_g
if epoch ==0:
start = time.process_time()
for steps in range(data.shape[0] // config.size_of_batch_g):
_data = data[ub_list[steps]:ub_list[steps+1]][:]
images = autograd.Variable(torch.Tensor(_data)).to(device)
enc_optim.zero_grad()
dec_optim.zero_grad()
# ================Recons loss============ #
batch_size = images.size()[0]
z = encoder(images, gps)
x_recon = decoder(z, gps)
recon_loss = criterion(x_recon.cuda(), images.cuda())
# ======== Kernel Loss ======== #
z_fake = Variable(torch.rand(images.size()[0], config.n_z_g) * config.sigma).to(device)
z_real = encoder(images).to(device)
if js == 0 :
cust_loss = imq_kernel(z_real.cuda(), z_fake.cuda(), h_dim=2)
cust_loss = cust_loss / batch_size
total_loss = recon_loss + config.alpha * cust_loss
elif js == 1:
if exp == 1 and beta == 0 and gauss == 0:
p = Variable(torch.exp(torch.rand(images.size()[0], config.n_z_g)) * config.sigma).to(device)
elif exp ==0 and beta == 1 and gauss == 0:
beta_dis = torch.distributions.Beta(2,2).sample(torch.Size([images.size()[0], config.n_z_g])).cuda()
p = Variable(beta_dis * config.sigma).cuda()
elif exp == 0 and beta == 0 and gauss == 1:
p = Variable(torch.rand(images.size()[0], config.n_z_g) * config.sigma).to(device)
else:
print("please change your selections")
exit()
q = encoder(images).to(device)
cust_loss = jenson_shannon_divergence(p,q)
total_loss = recon_loss + config.alpha * cust_loss
total_loss.backward()
enc_optim.step()
dec_optim.step()
rec_loss.append(recon_loss.data.item())
wae_cust_loss.append(cust_loss.data.item())
tot_loss.append(total_loss.data.item())
if epoch == 0:
print("epoch "+ str(epoch) +" ",time.process_time() - start)
if (epoch + 1) % config.recons_ep == 0:
print("Model Number: [%d/%d] Epoch: [%d/%d], Reconstruction Loss: %.4f, MMD Loss %.4f, TOTAL Loss %.4f" %
(model_num + 1, config.total_epoch, epoch + 1, config.ITERS_g, recon_loss.data.item(),
cust_loss.item(), total_loss.item()))
if epoch == config.ITERS_g -1:
path = str(i)
torch.save(torch.tensor(rec_loss),path+'/rec_loss/rec_'+str(model_num)+'_'+str(i)+'.pth')
torch.save(torch.tensor(wae_cust_loss),path+'/cust_loss/cust_'+str(model_num)+'_'+str(i)+'.pth')
torch.save(torch.tensor(tot_loss),path+'/total_loss/total_'+str(model_num)+'_'+str(i)+'.pth')
torch.save(encoder.state_dict(),path+'/encoder/enc_'+str(model_num)+'_'+str(i)+'.pth')
torch.save(decoder.state_dict(),path+'/decoder/dec_'+str(model_num)+'_'+str(i)+'.pth')
elif mnist == 1:
train_m(gps, js, exp, beta, gauss, mnist)
def train_m(gps, js, exp, beta, gauss, mnist):
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print(device)
if mnist == 1:
#Converting data to torch.FloatTensor
train_data, test_data, classes = mnist_loader()
encoder, decoder = Encoder_Mnist().to(device), Decoder_Mnist().to(device)
criterion = nn.MSELoss()
criterion.to(device)
encoder.to(device)
decoder.to(device)
encoder.train()
decoder.train()
enc_optim = optim.Adam(encoder.parameters(), lr= config.lr)
dec_optim = optim.Adam(decoder.parameters(), lr= config.lr)
enc_scheduler = StepLR(enc_optim, step_size=30, gamma=0.5)
dec_scheduler = StepLR(dec_optim, step_size=30, gamma=0.5)
path = "test_v2_bs_"+str(config.size_of_batch_m)+"ls_"+str(config.n_z_m)+"ac_relu/"
os.mkdir(path)
original_stdout = sys.stdout
for sample_size in config.n_samples:
rec_loss_list = []
cust_loss_list = []
total_loss_list = []
total_models = 20
for model_num in range(total_models):
cnt = -1
subset_indices = random.sample(range(0, len(train_data)), sample_size)
subset = torch.utils.data.Subset(train_data, subset_indices)
train_loader_subset = torch.utils.data.DataLoader(subset, batch_size=1000, num_workers=0, shuffle=False)
for epoch in range(config.ITERS_m):
for data in train_loader_subset:
images, _ = data
images = images.to(device)
enc_optim.zero_grad()
dec_optim.zero_grad()
x_latent = encoder(images, gps = 1).to(device)
x_recon = decoder(x_latent, gps = 1).to(device)
recon_loss = criterion(x_recon, images).to(device)
if js == 0 :
x_fake = Variable(torch.rand(images.size()[0], config.n_z_m) * config.sigma).to(device)
cust_loss = imq_kernel(x_latent.cuda(), x_fake.cuda(), h_dim=2)
cust_loss = cust_loss / batch_size
total_loss = recon_loss + config.alpha * cust_loss
if js == 1:
if exp == 1 and beta == 0 and gauss == 0:
p = Variable(torch.exp(torch.rand(images.size()[0], config.n_z_m)) * config.sigma).to(device)
elif exp ==0 and beta == 1 and gauss == 0:
beta_dis = torch.distributions.Beta(2,2).sample(torch.Size([images.size()[0], config.n_z_m])).cuda()
p = Variable(beta_dis * config.sigma).cuda()
elif exp == 0 and beta == 0 and gauss == 1:
p = Variable(torch.rand(images.size()[0], config.n_z_m) * config.sigma).to(device)
else:
print("please change your selections")
exit()
cust_loss = jenson_shannon_divergence(p,x_latent).to(device)
total_loss = recon_loss + config.alpha * cust_loss
total_loss.backward()
enc_optim.step()
dec_optim.step()
if epoch+1 == 1 and model_num + 1 ==1:
os.makedirs(path+'loss_'+str(sample_size))
os.makedirs(path+'./loss_'+str(sample_size)+'/rec_loss')
os.makedirs(path+'./loss_'+str(sample_size)+'/cust_loss')
os.makedirs(path+'./loss_'+str(sample_size)+'/total_loss')
os.makedirs(path+'./loss_'+str(sample_size)+'/encoder')
os.makedirs(path+'./loss_'+str(sample_size)+'/decoder')
if (epoch+1) % 60 == 0:
print("Model Number [%d/%d], Epoch: [%d/%d], Reconstruction Loss: %.6f JS Loss: %.6f Total Loss: %.6f" %(model_num+1, total_models, epoch + 1, config.ITERS_m, recon_loss.data.item(), cust_loss.item(), total_loss))
if epoch == config.ITERS_m - 1:
torch.save(torch.tensor(rec_loss_list), path+'./loss_'+str(sample_size)+'/rec_loss'+'/rec_'+str(model_num)+'_loss.pth')
torch.save(torch.tensor(cust_loss_list), path+'./loss_'+str(sample_size)+'/cust_loss'+'/cust_'+str(model_num)+'_loss.pth')
torch.save(torch.tensor(total_loss_list), path+'./loss_'+str(sample_size)+'/total_loss'+'/total_'+str(model_num)+'_loss.pth')
torch.save(encoder.state_dict(),path+'./loss_'+str(sample_size)+'/encoder/enc_'+str(model_num)+'.pth')
torch.save(decoder.state_dict(),path+'./loss_'+str(sample_size)+'/decoder/dec_'+str(model_num)+'.pth')
print(str(sample_size)+' Done!')
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--groupsort", type=int, default = 0)
parser.add_argument("--js", type=int, default = 0)
parser.add_argument("--beta", type=int, default = 0)
parser.add_argument("--exp", type=int, default = 0)
parser.add_argument("--gauss", type=int, default = 0)
parser.add_argument("--mnist", type=int, default = 0)
# Parse the command-line arguments
args = parser.parse_args()
# Access the parsed argument
c_gps = args.groupsort
c_js = args.js
c_beta = args.beta
c_gauss = args.gauss
c_exp = args.exp
c_mnist = args.mnist
if c_mnist == 0:
train_g(c_gps, c_js, c_exp, c_beta, c_gauss, c_mnist)
else:
train_m(c_gps, c_js, c_exp, c_beta, c_gauss, c_mnist)