-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathmain.py
237 lines (194 loc) · 9.54 KB
/
main.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
# import stuff
import os
import argparse
import itertools
import numpy as np
import torch
import torchvision
from torchvision import datasets
import torch.nn as nn
import torchvision.transforms as transforms
import torch.nn.functional as F
import matplotlib.pyplot as plt
from fno import *
from metrics import *
from sde import *
from priors import *
from utils import get_samples, get_samples_batched, downsampling_fourier
import os
import datetime
from torch.utils.data import DataLoader
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
def training(seed, model, args):
"""
training the score function
:param seed: random seed, ensure reproducibility within the same seed
:param model: function approximator for the score function
:param args: hyperparameters
:return:
"""
torch.manual_seed(seed)
np.random.seed(seed)
# Define a transform to normalize the data
transform = transforms.Compose([transforms.ToTensor(), transforms.Resize(64), transforms.GaussianBlur(7, 0.1)])
# Download and load the training data
trainset = datasets.MNIST(root='', train=True, download=True, transform=transform)
trainset, val_set = torch.utils.data.random_split(trainset, [55000, 5000])
trainloader = torch.utils.data.DataLoader(trainset, batch_size=args.batch_size, shuffle=True)
valloader2 = torch.utils.data.DataLoader(val_set, batch_size=5000, shuffle=True)
val_test = next(iter(valloader2))[0]
valloader = torch.utils.data.DataLoader(val_set, batch_size=args.batch_size, shuffle=True)
testset = datasets.MNIST(root='', train=False, download=True, transform=transform)
testloader = torch.utils.data.DataLoader(testset, batch_size=args.batch_size, shuffle=True)
testloader2 = torch.utils.data.DataLoader(testset, batch_size=len(testset), shuffle=True)
test_set = next(iter(testloader2))[0]
print('NUMBER OF PARAMETERS:')
print(sum(p.numel() for p in model.parameters()))
fwd_sde = VariancePreservingSDE(args.prior, alpha_min=0.1, alpha_max=20.0, T=1. )
rev_sde = PluginReverseSDE(args.prior, fwd_sde, model, 1., vtype='Rademacher', debias=False).to(device)
optim = torch.optim.Adam(model.parameters(), lr=args.lr, weight_decay = 1e-4)
rev_sde.train()
if args.train ==True:
loss_curve = []
loss_curve_high = []
sw_list = []
min_loss = 10000.
for ep in range(args.n_epochs):
mean_loss = 0.
mean_loss_high = 0.
for k,(x,y) in enumerate(valloader):
x = x.to(device)
with torch.no_grad():
rev_sde.eval()
true_loss = rev_sde.dsm(downsampling_fourier(x)).mean().item()*x.shape[0]
mean_loss = mean_loss + true_loss
true_loss_high = rev_sde.dsm(x).mean().item()*x.shape[0]
mean_loss_high = mean_loss_high + true_loss_high
mean_loss = mean_loss/len(val_set)
mean_loss_high = mean_loss_high/len(val_set)
if ep > 0:
if min_loss > mean_loss:
print("new loss")
min_loss = mean_loss
torch.save(model.state_dict(),'model'+str(args.prior_name))
print('EPOCH: ', ep)
print(mean_loss)
loss_curve.append(mean_loss)
loss_curve_high.append(mean_loss_high)
if ep%10 ==0:
with torch.no_grad():
rev_sde.eval()
test_set = test_set#.cpu().data.numpy()
samples = get_samples_batched(rev_sde, 1, 64, 200, 2000).cpu()#.data.numpy()
for ll in range(len(samples)):
samples[ll] = (samples[ll]-samples[ll].min())/(samples[ll].max()-samples[ll].min())
print('SW')
sw = sw_approx(samples.view(len(samples),-1), test_set.view(len(test_set),-1))
sw_list.append(sw.item())
print(sw)
for k,(x,y) in enumerate(trainloader):
rev_sde.train()
x = x.to(device)
loss = rev_sde.dsm(downsampling_fourier(x)).mean()
optim.zero_grad()
loss.backward()
optim.step()
model.load_state_dict(torch.load('model'+str(args.prior_name)))
rev_sde.eval()
np.save('sw_list'+str(args.prior_name), np.array(sw_list))
with torch.no_grad():
samples = get_samples(rev_sde, 1, 32, 1000, 25)[0]
plt.figure()
fig, axs = plt.subplots(5, 5, figsize=(12, 12))
axs = axs.flatten()
for img, ax in zip(samples, axs):
ax.set_axis_off()
ax.imshow(img.squeeze(0).cpu().data.numpy(), cmap= 'gray')
plt.tight_layout()
plt.savefig(str(args.out_dir)+'/samples32'+str(args.prior_name))
plt.close()
samples = get_samples(rev_sde, 1, 50, 1000, 25)[0]
plt.figure()
fig, axs = plt.subplots(5, 5, figsize=(12, 12))
axs = axs.flatten()
for img, ax in zip(samples, axs):
ax.set_axis_off()
ax.imshow(img.squeeze(0).cpu().data.numpy(), cmap= 'gray')
plt.tight_layout()
plt.savefig(str(args.out_dir)+'/samples50'+str(args.prior_name))
plt.close()
samples = get_samples(rev_sde, 1, 64, 1000, 25)[0]
plt.figure()
fig, axs = plt.subplots(5, 5, figsize=(12, 12))
axs = axs.flatten()
for img, ax in zip(samples, axs):
ax.set_axis_off()
ax.imshow(img.squeeze(0).cpu().data.numpy(), cmap= 'gray')
plt.tight_layout()
plt.savefig(str(args.out_dir)+'/samples64'+str(args.prior_name))
plt.close()
plt.figure()
plt.plot(loss_curve, color = 'red')
plt.plot(loss_curve_high, color = 'blue')
plt.xlabel('Training Epochs', fontsize = "x-large")
plt.ylabel('DSM Loss', fontsize = "x-large")
plt.tight_layout()
plt.savefig(str(args.out_dir)+'/loss_curve'+str(args.prior_name))
plt.close()
else:
model.load_state_dict(torch.load('model'+str(args.prior_name)))
rev_sde = PluginReverseSDE(args.prior, fwd_sde, model, 1., vtype='Rademacher', debias=False).to(device)
with torch.no_grad():
rev_sde.eval()
samples = get_samples(rev_sde, 1, 64, 1000, 25)[0]
plt.figure()
fig, axs = plt.subplots(5, 5, figsize=(12, 12))
axs = axs.flatten()
for img, ax in zip(samples, axs):
ax.set_axis_off()
ax.imshow(img.squeeze(0).cpu().data.numpy(), cmap= 'gray')
plt.tight_layout()
plt.savefig(str(args.out_dir)+'/5x5samples64'+str(args.prior_name))
test_set = test_set#.cpu().data.numpy()
samples = get_samples_batched(rev_sde, 1, 64, 200, 10000).cpu()#.data.numpy()
for ll in range(len(samples)):
samples[ll] = (samples[ll]-samples[ll].min())/(samples[ll].max()-samples[ll].min())
print('SW')
print(sw_approx(samples.view(len(samples),-1), test_set.view(len(test_set),-1)))
print('diversity score')
print(compute_vendi_score(samples.cpu().data.numpy().reshape(len(samples),-1)))
print(compute_vendi_score(test_set.cpu().data.numpy().reshape(len(test_set),-1)))
return loss_curve, loss_curve_high,sw_list
def choose_prior(string):
if string.lower() == "fno":
return FNOprior(k1=args.modes_prior,k2=args.modes_prior//2)
elif string.lower() == "standard":
return StandardNormal()
elif string.lower() == "lap_conv":
return ImplicitConv(scale=20)
elif string.lower() == "combined_conv":
return CombinedConv(k1=args.modes_prior,k2=args.modes_prior//2,scale=10., scale2 = 0.5)
elif string.lower() == "bessel":
return BesselConv(scale = 8, power = 0.55)
else:
raise argparse.ArgumentTypeError(f"Invalid class")
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Training arguments')
parser.add_argument('--n_epochs', type=int, default=200, help='epochs of pde training')
parser.add_argument('--lr', type=float,default=6e-4, help='learning rate')
parser.add_argument('--batch_size', type=int, default=128, help='number of training samples in each batch')
parser.add_argument('--num_samples', type=int, default=16, help='number of samples for visualization')
parser.add_argument('--input_height', type=int, default=32, help='training image resolution')
parser.add_argument('--prior_name', type=str, default='lap_conv', help="prior setup")
parser.add_argument('--width', type=str, default=32, help="prior setup")
parser.add_argument('--modes', type=int, default=14, help='cutoff modes in FNO')
parser.add_argument('--modes_prior', type=int, default=32, help='cutoff modes in FNO')
parser.add_argument('--train', type=bool, default=False, help='train or eval')
parser.add_argument('--seed', type=int, default=0, help='seed for random number generator')
parser.add_argument('--out_dir', type=str, default='mnist_results', help='directory for result')
args = parser.parse_args()
torch.manual_seed(args.seed)
args.prior = choose_prior(args.prior_name)
model = FNO2d(args.modes,args.modes,args.width)
input_channels = 1
loss_curve, loss_curve_high, sw_list = training(args.seed, model, args)