-
Notifications
You must be signed in to change notification settings - Fork 2
/
Ours_BlindHSISR2020.py
413 lines (343 loc) · 14.8 KB
/
Ours_BlindHSISR2020.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
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
import torch
import torch.nn as nn
from torch.autograd import Variable
import torch.nn.functional as F
from downsampler import *
import math
import numpy as np
import scipy.io as sio
from model import *
from model_1 import *
from scipy.misc import imresize
import time
import copy
import torch.utils.data as data
from torch.utils.data import DataLoader
from torch.nn.parameter import Parameter
import matplotlib.pyplot as plt
import os
import torch.nn.functional as func
from DownKernel import *
##parameter setting
U_spa = 1 # 0 indicates the spatial degeneration model is known, 1 is unknown
U_spc = 1 # 0 indicates the spatial degeneration model is known, 1 is unknown
pretrain = 1 # image specific prior learning
num_steps = 5001 # iterations
lr_dc = 1.1e-4 # learning rate of spectral degeneration network
lr_da = 1.1e-4 # learning rate of spatial degeneration network
lr_i = 1.2e-3 # learning rate of the generator network
factor = 8
kers = 1 # the order of the spatial blur kernel in test phase
N_h = 40 # SNR of the observed LR HSI
N_m = 40 # SNR of the observed HR MSI
scale = 0 # the noise intensity in test spectral response matrix
Dsets_name = 'Both Unknown X{} K{} HSI_{}N MSI_{}N P{}N'.format(factor, kers, N_h, N_m, scale)
CAVE_dataset = False
names = names_ICLV_test
dir_data = '' # The path of ground truth HSI.
H_path = '' # The path of the LR HSI .
M_path = '' # The path of the HR MSI.
save_path = '' # The path of where to save results.
if not os.path.exists(save_path):
os.mkdir(save_path)
#get the spectual downsample func
def create_P():
P_ = [[2, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 6, 11, 17, 21, 22, 21, 20, 20, 19, 19, 18, 18, 17, 17],
[1, 1, 1, 1, 1, 1, 2, 4, 6, 8, 11, 16, 19, 21, 20, 18, 16, 14, 11, 7, 5, 3, 2, 2, 1, 1, 2, 2, 2, 2, 2],
[7, 10, 15, 19, 25, 29, 30, 29, 27, 22, 16, 9, 2, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]]
P = np.array(P_, dtype=np.float32)
div_t = np.sum(P, axis=1)
for i in range(3):
P[i,] = P[i,]/div_t[i]
return P
#get PSNR
def PSNR_GPU(im_true, im_fake):
data_range = 1
C = im_true.size()[0]
H = im_true.size()[1]
W = im_true.size()[2]
Itrue = im_true.clone().resize_(C*H*W)
Ifake = im_fake.clone().resize_(C*H*W)
#mse = nn.MSELoss(reduce=False)
err = torch.pow(Itrue-Ifake,2).sum(dim=0, keepdim=True).div_(C*H*W)
psnr = 10. * torch.log((data_range**2)/err) / np.log(10.)
return torch.mean(psnr)
#get SAM
def SAM_GPU(im_true, im_fake):
C = im_true.size()[0]
H = im_true.size()[1]
W = im_true.size()[2]
Itrue = im_true.clone().resize_(C, H*W)
Ifake = im_fake.clone().resize_(C, H*W)
nom = torch.mul(Itrue, Ifake).sum(dim=0).resize_(H*W)
denom1 = torch.pow(Itrue,2).sum(dim=0).sqrt_().resize_(H*W)
denom2 = torch.pow(Ifake,2).sum(dim=0).sqrt_().resize_(H*W)
sam = torch.div(nom, torch.mul(denom1, denom2)).acos_().resize_(H*W)
sam = sam / np.pi * 180
sam = torch.sum(sam) / (H*W)
return sam
# The dataset of spatial downsampled HR MSI to HR MSI
class RGB_sets(data.Dataset):
def __init__(self,RGB,Label,P_N,P_S,factor):
self.RGB = RGB
self.size = RGB.shape
self.P_S = P_S
self.Label = Label
self.P_N = P_N
self.D_spa = Downsampler(n_planes=31,factor=factor,kernel_type='gauss12',phase=0,preserve_size=True).type(torch.cuda.FloatTensor)
def __getitem__(self, index):
[p,q] = self.Label[index,:]
P_S = self.P_S
RGB = self.RGB[:,p:p+P_S,q:q+P_S]
D_RGB = self.D_spa(torch.unsqueeze(RGB,0))
return torch.squeeze(D_RGB,0),RGB
def __len__(self):
return self.P_N
# The dataset of spectral downsampled LR HSI to LR HSI
class HSI_sets(data.Dataset):
def __init__(self,HSI,Label,P_N,P_S):
self.HSI = HSI
self.size = HSI.shape
self.P_S = P_S
self.Label = Label
self.P_N = P_N
self.P = create_P()
def __getitem__(self, index):
[p,q] = self.Label[index,:]
P_S = self.P_S
HSI = self.HSI[:,p:p+P_S,q:q+P_S]
P = torch.from_numpy(self.P).cuda()
HSI = torch.reshape(HSI,(HSI.shape[0],P_S**2))
D_HSI = torch.matmul(P,HSI).view(3,P_S,P_S)
HSI = HSI.view(HSI.shape[0],P_S,P_S)
return D_HSI,HSI
def __len__(self):
return self.P_N
# Crop the test image into patches.
def Get_Label(im_size, patch_size):
m, n = 0, 0
P_number = 0
Lable_table = []
while 1:
if m + patch_size < im_size[1] and n + patch_size < im_size[0]:
Lable_table.append([m, n])
m = m + patch_size
P_number += 1
elif m + patch_size >= im_size[1] and n + patch_size < im_size[0]:
m = im_size[1] - patch_size
Lable_table.append([m, n])
m, n = 0, n + patch_size
P_number += 1
elif m + patch_size < im_size[1] and n + patch_size >= im_size[0]:
Lable_table.append([m, im_size[0] - patch_size])
m += patch_size
P_number += 1
elif m + patch_size >= im_size[1] and n + patch_size >= im_size[0]:
Lable_table.append([im_size[1] - patch_size, im_size[0] - patch_size])
P_number += 1
break
return np.array(Lable_table), P_number
# Spatial upsample model
def Spa_UpNet(image):
if factor == 8:
P_S = 32
elif factor == 16:
P_S = 64
else:
P_S = 128
Label, P_n = Get_Label(image.shape[1:], P_S)
dataset = RGB_sets(image, Label, P_n, P_S,factor=factor)
data = DataLoader(dataset=dataset, batch_size=4, shuffle=True)
N1 = nn.Sequential()
N1.add(get_net_1(31,'skip','reflection',n_channels=31,skip_n33d=40,skip_n33u=40,skip_n11=1,num_scales=2,upsample_mode='bilinear'))
N1.add(nn.Upsample(scale_factor=factor,mode='bilinear'))
N1.add(get_net_1(31,'skip','reflection',n_channels=31,skip_n33d=40,skip_n33u=40,skip_n11=1,num_scales=2,upsample_mode='bilinear'))
N1 = N1.cuda()
L1Loss = nn.L1Loss(reduce=True)
optimizer_N1 = torch.optim.Adam(N1.parameters(), lr=5e-4)
for epoch in range(500):
running_loss = 0
for i,batch in enumerate(data,1):
lr,hr = batch[0],batch[1]
lr,hr = Variable(lr).cuda(),Variable(hr).cuda()
out = N1(lr)
loss = L1Loss(out,hr)
running_loss += loss.detach()
optimizer_N1.zero_grad()
loss.backward()
optimizer_N1.step()
return N1
# spectral upsample model
def Spc_UpNet(image):
P_S = 16
Label,P_n = Get_Label(image.shape[1:],P_S)
dataset = HSI_sets(image,Label,P_n,P_S)
data = DataLoader(dataset=dataset, batch_size=4, shuffle=True)
N2 = get_net_1(3,'skip','reflection',n_channels=31,skip_n33d=31,skip_n33u=31,skip_n11=4,num_scales=5,upsample_mode='bilinear')
N2 = N2.cuda()
L1Loss = nn.L1Loss(reduce=True)
optimizer_N2 = torch.optim.Adam(N2.parameters(), lr=5e-4)
for epoch in range(500):
running_loss = 0
for i,batch in enumerate(data,1):
lr,hr = batch[0],batch[1]
lr,hr = Variable(lr).cuda(),Variable(hr).cuda()
out = N2(lr)
loss = L1Loss(out,hr)
running_loss += loss.detach()
optimizer_N2.zero_grad()
loss.backward()
optimizer_N2.step()
return N2
# Learnable spectral response matrix
class L_Dspec(nn.Module):
def __init__(self,in_channel,out_channel,P_init):
super(L_Dspec, self).__init__()
self.in_channle = in_channel
self.out_channel = out_channel
self.P = Parameter(P_init)
def forward(self,input):
S = input.shape
out = torch.reshape(input,[S[1],S[2]*S[3]])
out = torch.matmul(self.P,out)
return torch.reshape(out,[self.out_channel,S[2],S[3]])
if __name__ =='__main__':
# define the lossfunction
L1Loss = nn.L1Loss()
k = 0
files = os.listdir(dir_data)
for name in names:
lr = lr_i
k += 1
# Load the Ground Truth HSI
if CAVE_dataset:
data = sio.loadmat(dir_data+name+'_ms.mat')
else:
data = sio.loadmat(dir_data+name+'.mat')
print('Producing with the {} image'.format(name))
im_gt = data['data']
# Generate the spectral response matrix
p = create_P()
p = Variable(torch.from_numpy(p.copy()), requires_grad=False).cuda()
# Trans the GT into Variable and nomalize it
im_gt = Variable(torch.from_numpy(im_gt.copy()),requires_grad=False).type(torch.cuda.FloatTensor).cuda()
im_gt = im_gt/(torch.max(im_gt)-torch.min(im_gt))
s = im_gt.shape
GT = im_gt.view(s[0],s[1]*s[2])
# Load the observed HR MSI
M_data = sio.loadmat(M_path+name+'.mat')
im_m = M_data['HR_MSI']
im_m = Variable(torch.from_numpy(im_m), requires_grad=False).type(torch.cuda.FloatTensor)
if U_spc == 1:
P_N = sio.loadmat('P_N.mat')
P_N = torch.from_numpy(P_N['P'])
down_spc = L_Dspec(31,3,P_N).type(torch.cuda.FloatTensor).cuda()
optimizer_spc = torch.optim.Adam(down_spc.parameters(),lr=lr_dc,weight_decay=1e-5)
# Load the Observed LR HSI
H_data = sio.loadmat(H_path+name+'.mat')
im_h = H_data['LR_HSI']
im_h = Variable(torch.from_numpy(im_h), requires_grad=False).cuda()
start_time = time.time()
if pretrain == 1:
print('Stage one : Pretrain the Spc_UpNet.')
Spc_up = Spc_UpNet(im_h)
H_RGB = Spc_up(torch.unsqueeze(im_m,0))
print('Stage two : Pretrain the Spa_UpNet.')
Spa_up = Spa_UpNet(torch.squeeze(H_RGB,0))
H_HSI = Spa_up(torch.unsqueeze(im_h,0))
net_input = Variable(0.8*H_RGB + 0.2*H_HSI).cuda()
else:
net_input = Variable(torch.unsqueeze(torch.rand_like(im_gt), 0)).cuda()
if U_spa == 1:
#Learnable spatial downsampler
KS = 32
dow = nn.Sequential(nn.ReplicationPad2d(int((KS - factor)/2.)), nn.Conv2d(1,1,KS,factor))
class Apply(nn.Module):
def __init__(self, what, dim, *args):
super(Apply, self).__init__()
self.dim = dim
self.what = what
def forward(self, input):
inputs = []
for i in range(input.size(self.dim)):
inputs.append(self.what(input.narrow(self.dim, i, 1)))
return torch.cat(inputs, dim=self.dim)
def __len__(self):
return len(self._modules)
downs = Apply(dow, 1)
downs = downs.cuda()
optimizer_d = torch.optim.Adam(downs.parameters(),lr=lr_da,weight_decay=1e-5)
else:
down_spa = Downsampler(n_planes=im_gt.shape[0], factor=factor, kernel_type='gauss12', phase=0,preserve_size=True).type(torch.cuda.FloatTensor)
# get_input
net = get_net(im_gt.shape[0], 'skip', 'reflection', n_channels=im_gt.shape[0], skip_n33d=256, skip_n33u=256,skip_n11=1, num_scales=2, upsample_mode='bilinear')
net = net.cuda()
optimizer = torch.optim.Adam(net.parameters(), lr=lr)
f = open(save_path+name+'_result.txt','a+')
f.write('\n\n\nThe result of PSNR & SAM is :\n\n')
f.write('The experiment:At {} times SR & All the downsampler is unknown & the DSpa is {}.\n\n'.format(factor,Dsets_name))
f2 = open(save_path+name+'_loss.txt','w')
f2.write('\n\n\nThe experiment loss is :\n\n')
psnr_max = 0
sam_max = 0
print('Stage three : Producing with the {} image'.format(name))
for i in range(num_steps):
#input data
output = net(net_input)
#procese of output
S = output.shape
if U_spa == 0:
Dspa_O = down_spa(output)
Dspa_O = Dspa_O.view(Dspa_O.shape[1],Dspa_O.shape[2],Dspa_O.shape[3])
else:
Dspa_O = downs(output)
Dspa_O = torch.squeeze(Dspa_O,0)
if U_spc == 0:
out = output.view(S[1],S[2]*S[3])
Dspc_O = torch.matmul(p,out).view(im_m.shape[0],S[2],S[3])
else:
Dspc_O = down_spc(output)
#zero the grad
optimizer.zero_grad()
if U_spc==1:
optimizer_spc.zero_grad()
if U_spa==1:
optimizer_d.zero_grad()
loss = L1Loss(Dspa_O,im_h) + L1Loss(Dspc_O,im_m)
#backward the loss
loss.backward()
#optimize the parameter
optimizer.step()
if U_spc==1:
optimizer_spc.step()
if U_spa==1:
optimizer_d.step()
#print('At step {},the loss is {}.'.format(i,loss.data.cpu()))
if i%10 == 0:
f2.write('At step {},the loss is {}\n'.format(i,loss.data.cpu()))
if i%100 == 0:
output = Variable(output,requires_grad=False).cuda()
output = output.view(S[1],S[2],S[3])
psnr = PSNR_GPU(im_gt,output)
sam = SAM_GPU(im_gt,output)
f.write('{},{}\n'.format(psnr,sam))
print('**{}**{}**At the {}th loop the PSNR&SAM is {},{}.'.format(k,Dsets_name,i,psnr,sam))
if i%1000 == 0:
lr = 0.7*lr
for param_group in optimizer.param_groups:
param_group['lr'] = lr
if i == num_steps-1:
data = {}
out = np.array(output.squeeze().detach().cpu())
data['data'] = out
sio.savemat(save_path+name+'_r.mat',data)
if U_spa == 1:
torch.save(downs,save_path+name+'_DA.pth')
if U_spc == 1:
torch.save(down_spc,save_path+name+'_DC.pth')
used_time = time.time()-start_time
print('The training time is :{}.'.format(used_time))
f.write('The training time is :{}.'.format(used_time))
f.close()
f2.close()
torch.cuda.empty_cache()