-
Notifications
You must be signed in to change notification settings - Fork 3
/
HapkeCNN_Simulated2_Git.py
196 lines (170 loc) · 7.16 KB
/
HapkeCNN_Simulated2_Git.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
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Fri Aug 26 15:58:49 2022
@author: Behnood
"""
import matplotlib.pyplot as plt
import os
import numpy as np
import torch
import torch.optim
import torch.nn as nn
from common import *
from UnmixArch import UnmixArch
from UtilityMine import *
import scipy.io
import scipy.linalg
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark =True
dtype = torch.cuda.FloatTensor
PLOT = True
save_result=False
#%%
fname2 = "Git_Data/Simulated2/Y_clean.mat"
mat2 = scipy.io.loadmat(fname2)
img_np_gt = mat2["Y_clean"]
img_np_gt = img_np_gt.transpose(2,0,1)
[p1, nr1, nc1] = img_np_gt.shape
data = scipy.io.loadmat('Git_Data/Noisy/Simulated_non_pure_pixels_reflectance.mat')
YY =data['YY']
A_true_np = data['A_true'] # true abundance
A_true_np = A_true_np.transpose(2,0,1)
E = data['E'] # true endmembers
m=1
m0=1
W=np.power((m0+m)**2*np.power(E,2)+np.multiply(1+4*m*m0*E, 1-E),0.5)-(m0+m)*E
W1=np.divide(W,1+4*m*m0*E)
EE=1-np.power(W1,2)
# Number of endmemebrs
rmax=6#E_np.shape[1]
tol1=1 # comment out the line below for all SNRs
#tol1=YY.shape[1]
tol2=1
# comment out the line below for 10 runs
# tol2=YY.shape[0]
# Selecting tuning parameters for the loss function
lamb=0.1
alpha=0.0001
from tqdm import tqdm
for fi in tqdm(range(tol1)):
for fj in tqdm(range(tol2)):
#%%
img_noisy_np = YY[fj][fi]
img_noisy_np = img_noisy_np.transpose(2,0,1)
[p1, nr1, nc1] = img_noisy_np.shape
#print(compare_snr(img_np_gt, img_noisy_np))
img_resh=np.reshape(img_noisy_np,(p1,nr1*nc1))
V, SS, U = scipy.linalg.svd(img_resh, full_matrices=False)
PC=np.diag(SS)@U
img_resh_DN=V[:,:rmax]@V[:,:rmax].transpose(1,0)@img_resh
img_resh_np_clip=np.clip(img_resh_DN, 0, 1)
II,III = Endmember_extract(img_resh_np_clip,rmax)
E_np1=img_resh_np_clip[:,II]
#%% Set up Simulated
INPUT = 'noise' # 'meshgrid'
pad = 'reflection'
need_bias=True
OPT_OVER = 'net'
#
LR1 = 0.001
show_every = 500
exp_weight=0.99
num_iter1 = 8000
input_depth = img_noisy_np.shape[0]
class CAE_EndEst(nn.Module):
def __init__(self):
super(CAE_EndEst, self).__init__()
# encoding layers
self.conv1 = nn.Sequential(
UnmixArch(
input_depth, EE.shape[1],
num_channels_down = [ 256],
num_channels_up = [ 256],
num_channels_skip = [ 4],
filter_size_up = 3,filter_size_down = 3, filter_skip_size=1,
upsample_mode='bilinear', # downsample_mode='avg',
need1x1_up=False,
need_sigmoid=True, need_bias=True, pad=pad, act_fun='LeakyReLU').type(dtype)
)
self.dconv4 = nn.Sequential(
nn.Conv2d(rmax, p1, 1,1, padding="same",bias=False),
)
def forward(self, x):
x = self.conv1(x)
x1 = self.dconv4(x)
return x, x1
net1 = CAE_EndEst()
net1.cuda()
# Loss
def my_loss(target, End2, alpha,lamb, out_,out_spec):
# Albedo Loss
W=torch.pow((m0+m)**2*torch.pow(End2,2)+torch.mul(1+4*m*m0*End2, 1-End2),0.5)-(m0+m)*End2
W1=torch.div(W,1+4*m*m0*End2)
End3=1-torch.pow(W1,2)
HR=torch.mm(End3.view(p1,rmax),out_.view(rmax,nr1*nc1))
Temp1=1+2*m*torch.pow((1-HR),0.5)
Temp2=1+2*m0*torch.pow((1-HR),0.5)
out_HR=torch.div(HR, torch.mul(Temp1,Temp2))
loss = 0.5*torch.norm((out_HR.view(1,p1,nr1,nc1) - target), 'fro')**2
#---- Net Loss-------
loss1 = 0.5*torch.norm((out_spec.view(1,p1,nr1,nc1) - target), 'fro')**2
#------Minimum Volume Penalty: TV-----
O = torch.from_numpy(np.zeros((p1, rmax))).type(dtype)
B = np_to_torch(np.identity(rmax) - np.ones((rmax))/rmax).type(dtype)
loss2 = torch.norm(torch.mm(End3,B.view((rmax,rmax)))-O, 'fro')**2
return loss+alpha*loss1+lamb*loss2
img_noisy_torch = torch.from_numpy(img_resh_DN).view(1,p1,nr1,nc1).type(dtype)
net_input1 = get_noise(input_depth, INPUT,
(img_noisy_np.shape[1], img_noisy_np.shape[2])).type(dtype).detach()
E11=np.random.rand(p1,rmax);
E_torch = torch.from_numpy(E_np1).type(dtype)
#%%
out_avg = True
i = 0
def closure1():
global i, out_LR_np, out_avg, out_avg_np, Eest
out_LR,out_spec = net1(net_input1)
# Smoothing
if out_avg is None:
out_avg = out_LR.detach()
else:
out_avg = out_avg * exp_weight + out_LR.detach() * (1 - exp_weight)
#%%
total_loss = my_loss(img_noisy_torch, net1.dconv4[0].weight.view(p1,rmax),alpha, lamb,out_LR,out_spec)
total_loss.backward()
print ('Iteration %05d Loss %f ' % (i, total_loss.item()), '\r', end='')
if PLOT and i % show_every == 0:
out_LR_np = out_LR.detach().cpu().squeeze().numpy()
out_avg_np = out_avg.detach().cpu().squeeze().numpy()
out_LR_np = np.clip(out_LR_np, 0, 1)
out_avg_np = np.clip(out_avg_np, 0, 1)
f, ((ax1, ax2)) = plt.subplots(1, 2, sharey=True, figsize=(10,10))
ax1.imshow(np.stack((out_LR_np[2,:,:],out_LR_np[1,:,:],out_LR_np[0,:,:]),2))
ax2.imshow(np.stack((out_avg_np[2,:,:],out_avg_np[1,:,:],out_avg_np[0,:,:]),2))
plt.show()
i += 1
return total_loss
net1.dconv4[0].weight=torch.nn.Parameter(E_torch.view(p1,rmax,1,1))
p11 = get_params(OPT_OVER, net1, net_input1)
optimizer = torch.optim.Adam(p11, lr=LR1, betas=(0.9, 0.999), eps=1e-8,
weight_decay= 0, amsgrad=False)
for j in range(num_iter1):
optimizer.zero_grad()
closure1()
optimizer.step()
net1.dconv4[0].weight.data[net1.dconv4[0].weight <= 0] = 0
net1.dconv4[0].weight.data[net1.dconv4[0].weight >= 1] = 1
if j>0:
Eest=net1.dconv4[0].weight.detach().cpu().squeeze().numpy()
if PLOT and j % show_every== 0:
plt.plot(Eest)
plt.show()
out_avg_np = out_avg.detach().cpu().squeeze().numpy()
#%%
if save_result is True:
scipy.io.savemat("ResultsNoPure/EestdB%01d%01d.mat" % (fi+2, fj+1),
{'Eest%01d%01d' % (fi+2, fj+1):Eest})
scipy.io.savemat("ResultsNoPure/out_avg_npdB%01d%01d.mat" % (fi+2, fj+1),
{'out_avg_np%01d%01d' % (fi+2, fj+1):out_avg_np.transpose(1,2,0)})
#