DeepLearningProject/testmse.py at main · badesar1/DeepLearningProject · GitHub

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
import os
import torch
import torch.fft as fft
import torch.nn as nn
import torch.nn.functional as F
import numpy as np

class PS(torch.nn.Module):
    def __init__(self):
        super(PS, self).__init__()
        # self.iscuda = True
        # self.N = patchsize
        # self.batchsize = batchsize
        # self.A = patchsize ** 2
        # self.eps = torch.finfo(torch.float64).eps

        # datarange = torch.arange(0, patchsize)
        # if self.iscuda:
        #     datarange = datarange.cuda()
        # self.idatarange = torch.flip(datarange+1, dims=[0])
        # gammp = (patchsize - datarange) / (patchsize)
        # self.rangeMP = 4 * torch.sqrt(gammp)

        # nx = patchsize
        # ny = patchsize
        # kx = torch.arange(-nx/2, nx/2)
        # ky = torch.arange(-ny/2, ny/2)
        # self.kxgrid, self.kygrid = torch.meshgrid(kx, ky)
        # if self.iscuda:
        #     self.kxgrid = self.kxgrid.cuda()
        #     self.kygrid = self.kygrid.cuda()

    def powerspectrum(self, residual):
        sy = residual.size(-1)
        sx = residual.size(-2)
        imarea = sx * sy
        epf = fft.fft2(residual, s=(sx*2-1,sy*2-1))
        gamma = torch.abs(epf * torch.conj(epf)) / imarea
        gamma = fft.fftshift(gamma)
        gamma = torch.clip(gamma,min=-1,max=3)

        # autocorrelation = F.conv2d(residual, residual, padding=sx//2) / imarea
        # fftfromautocorr = torch.abs(fft.fft2(autocorrelation))

        #gamma = gamma[:,:,sx//4:3*sx//4,sy//4:3*sy//4]

        #mask = torch.ones_like(gamma)
        #mask[:,:,sx//4:3*sx//4,sy//4:3*sy//4] = 0
        #mask[:,:,sx//2,sy//2] = 0
        #gamma = gamma[mask==1]

        return gamma.flatten()

    def mp(self, residual):
        if residual.ndim == 4:
            residual = residual.squeeze()

        u,vals,v = torch.svd(residual)

        vals = (vals**2) / self.N
        csum = torch.cumsum(torch.flip(vals, dims=[1]), 1)
        sigmasq_1 = torch.flip(csum, dims=[1]) / self.idatarange

        rangeData = vals[:,:self.N] - vals[:,self.N-1].unsqueeze(1)
        sigmasq_2 = rangeData / self.rangeMP

        zeros = torch.zeros((sigmasq_1.size()))
        if self.iscuda:
            zeros = zeros.cuda()
        t1 = torch.where(sigmasq_2 < sigmasq_1, torch.sqrt(sigmasq_1), zeros)
        sigma_mp1 = torch.max(t1, axis=-1)[0]
        t2 = torch.where(sigmasq_2 < sigmasq_1, torch.sqrt(sigmasq_2), zeros)
        sigma_mp2 = torch.max(t2, axis=-1)[0]

        print(torch.mean(sigma_mp1))
        print(torch.mean(sigma_mp2))

        # min_idx = torch.argmax(torch.abs(sigmasq_1 - sigmasq_2), dim=-1)
        # sigma_mp1 = torch.sqrt(sigmasq_1[...,min_idx])
        # sigma_mp2 = torch.sqrt(sigmasq_2[...,min_idx])

        self.meanmp = torch.mean(sigma_mp1)
        return sigma_mp1, sigma_mp2

    def laplacian(self, residual, imgn):
        imgdn = imgn - residual
        imk = fft.fftshift(fft.fft2(imgdn, dim=[-2,-1]))
        Lapk = -(self.kxgrid**2 + self.kygrid**2) * imk
        Lapx = torch.abs(fft.ifft2(fft.ifftshift(Lapk), dim=[-2,-1]))
        self.meanlap = torch.mean(Lapx)
        return Lapx.flatten()

    def forward(self, out, noise, noisevar):
        batchsize = out.size(0)
        patchsize = out.size(-1)

        normres = out / noisevar.view(-1,1,1,1)
        ps = self.powerspectrum(normres)

        meanps = torch.mean(ps)
        self.meanps = meanps

        # dps = torch.mean(torch.abs(ps[1:] - ps[:-1]))
        # self.dps = dps

        loss1 = torch.sum(torch.abs(ps-1)) / (batchsize * patchsize**2)
        #loss1 = torch.abs(ps.mean()-1) / (2 * batchsize)
        self.ps = loss1

        loss7 = torch.sum((out - noise)**2) / (2 * batchsize)
        self.mse = loss7

        #print(loss1)

        # mp1, mp2 = self.mp(out)
        # loss2 = torch.linalg.norm(mp1 - noisevar) / (0.1)
        # loss3 = torch.linalg.norm(mp1 - mp2) / (2 * self.batchsize)
        # self.loss2 = loss2
        # self.loss3 = loss3
        # print(loss2)

        # lap = self.laplacian(residual, imgn)
        # loss3 = 1 / (50 * self.N * np.sqrt(self.batchsize)) * torch.linalg.norm(lap)


        #print(loss7)

        # loss4 = 0.05 * torch.linalg.norm(residual.flatten()) #0.3 too much
        # loss5 = 1 * torch.linalg.norm(torch.std(residual.flatten()) - noisevar)
        # loss6 = 1 * torch.linalg.norm(torch.mean(residual.flatten()))

        return loss1 + loss7 # + loss3 # loss1 + loss2 + loss6 + loss3 + loss7