VEMamba/metrics.py at main · I2-Multimedia-Lab/VEMamba · 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
133
import os
import numpy as np
from skimage import io
import torch
import torch.nn as nn
from pytorch_msssim import ssim,ms_ssim
import math
import json

cuda = torch.cuda.is_available()
Tensor = torch.cuda.FloatTensor if cuda else torch.Tensor

def compute_ssim(arr1, arr2,need_2d=True):
    '''ssim calculation on 3D, XY, XZ, YZ'''
    ssim_ls=[]
    ssim_3d = ssim(arr1, arr2,win_size=11,data_range=1.0)
    ssim_ls.append(ssim_3d.item())
    if need_2d:
        ssim_xy,ssim_xz,ssim_yz=0,0,0
        for i in range(arr1.shape[2]):
            ssim_xy += ssim(arr1[:,:,i, :, :], arr2[:,:,i, :, :],win_size=11,data_range=1)
        for j in range(arr1.shape[3]):
            ssim_xz += ssim(arr1[:,:,:, j, :], arr2[:,:,:, j, :],win_size=11,data_range=1)
        for k in range(arr1.shape[4]):
            ssim_yz += ssim(arr1[:,:,:, :, k], arr2[:,:,:, :, k],win_size=11,data_range=1)
        ssim_xy,ssim_xz,ssim_yz =ssim_xy/arr1.shape[2],ssim_xz / arr1.shape[3],ssim_yz / arr1.shape[4]
        ssim_ls.extend([ssim_xy.item(),ssim_xz.item(),ssim_yz.item()])

    return ssim_ls

def compute_ms_ssim(arr1, arr2,need_2d=True):
    '''ms-ssim calculation on 3D, XY, XZ, YZ'''
    ms_ssim_ls=[]
    ms_ssim_3d = ms_ssim(arr1, arr2,win_size=5, data_range=1)
    ms_ssim_ls.append(ms_ssim_3d.item())
    if need_2d:
        ms_ssim_xy,ms_ssim_xz,ms_ssim_yz=0,0,0
        for i in range(arr1.shape[2]):
            ms_ssim_xy += ms_ssim(arr1[:,:,i, :, :], arr2[:,:,i, :, :],win_size=5, data_range=1)
        for j in range(arr1.shape[3]):
            ms_ssim_xz += ms_ssim(arr1[:,:,:, j, :], arr2[:,:,:, j, :],win_size=5, data_range=1)
        for k in range(arr1.shape[4]):
            ms_ssim_yz += ms_ssim(arr1[:,:,:, :, k], arr2[:,:,:, :, k],win_size=5, data_range=1)
        ms_ssim_xy,ms_ssim_xz,ms_ssim_yz =ms_ssim_xy/arr1.shape[2],ms_ssim_xz / arr1.shape[3],ms_ssim_yz / arr1.shape[4]
        ms_ssim_ls.extend([ms_ssim_xy.item(),ms_ssim_xz.item(),ms_ssim_yz.item()])

    return ms_ssim_ls


def compute_psnr(arr1, arr2,need_2d=True):
    '''psnr calculation on 3D, XY, XZ, YZ'''
    psnr_ls=[]
    mse_3d=nn.MSELoss()(arr1, arr2)
    psnr_3d = 20 * math.log10(1 / math.sqrt(mse_3d.item()))
    psnr_ls.append(psnr_3d)
    if need_2d:
        psnr_xy,psnr_xz,psnr_yz=0,0,0
        for i in range(arr1.shape[2]):
            mse_xy = nn.MSELoss()(arr1[:,:,i, :, :], arr2[:,:,i, :, :])
            psnr_xy +=  20 * math.log10(1 / math.sqrt(mse_xy.item()))
        for j in range(arr1.shape[3]):
            mse_xz = nn.MSELoss()(arr1[:, :, :, j, :], arr2[:, :, :, j, :])
            psnr_xz += 20 * math.log10(1 / math.sqrt(mse_xz.item()))
        for k in range(arr1.shape[4]):
            mse_yz = nn.MSELoss()(arr1[:, :, : ,:, k], arr2[:, :, :, :, k])
            psnr_yz += 20 * math.log10(1 / math.sqrt(mse_yz.item()))
        psnr_xy,psnr_xz,psnr_yz =psnr_xy/arr1.shape[2],psnr_xz / arr1.shape[3],psnr_yz / arr1.shape[4]
        psnr_ls.extend([psnr_xy,psnr_xz,psnr_yz])

    return psnr_ls

def compute_lpips(arr1, arr2, need_2d=True):
    '''
    lpips calculation on XY, XZ, YZ
    Usage: conda install piq -c photosynthesis-team -c conda-forge -c PyTorch
    '''
    from piq import LPIPS
    arr1 = arr1.to(torch.float32)
    arr2 = arr2.to(torch.float32)
    loss = LPIPS()

    if need_2d:
        lpips_ls = []
        lpips_xy, lpips_xz, lpips_yz = 0, 0, 0
        for i in range(arr1.shape[2]):
            lpips_xy += loss(arr1[:, :, i, :, :], arr2[:, :, i, :, :])
        # for j in range(arr1.shape[3]):
        #     lpips_xz += loss(arr1[:, :, :, j, :], arr2[:, :, :, j, :])
        # for k in range(arr1.shape[4]):
        #     lpips_yz += loss(arr1[:, :, :, :, k], arr2[:, :, :, :, k])
        # lpips_xy, lpips_xz, lpips_yz = lpips_xy / arr1.shape[2], lpips_xz / arr1.shape[3], lpips_yz / arr1.shape[4]
        lpips_xy = lpips_xy / arr1.shape[2]
        lpips_ls.extend([lpips_xy.item()
                         # , lpips_xz.item(), lpips_yz.item()
                         ])
    return lpips_ls


def error_map(arr1, arr2,save_dir):
    '''visualize the error map'''
    err=torch.abs(arr1-arr2)
    err_np=np.array(err.squeeze()*255).astype('uint8')
    io.imsave(os.path.join(save_dir,'error_map.tif'),err_np)


def calculate_metrics(arr1,arr2,save_json=None,is_cuda=False,vis_error=False):
    '''calculate performance metrics, and save to json.'''
    assert arr1.shape == arr2.shape
    arr1 = torch.tensor(arr1[np.newaxis, np.newaxis, ...]/255.0)
    arr2 = torch.tensor(arr2[np.newaxis, np.newaxis, ...]/255.0)
    if is_cuda:
        arr1=arr1.cuda()
        arr2=arr2.cuda()

    metrics={}
    metrics['ssim']=compute_ssim(arr1, arr2)
    print('ssim done')
    metrics['ms_ssim'] = compute_ms_ssim(arr1, arr2)
    print('ms-ssim done')
    metrics['psnr'] = compute_psnr(arr1, arr2)
    print('psnr done')
    metrics['lpips'] = compute_lpips(arr1, arr2)
    print('lpips done')

    if vis_error:
        error_dir=os.path.dirname(save_json)
        error_map(arr1, arr2,error_dir)

    if save_json:
        with open(save_json, "w") as f:
            f.write(json.dumps(metrics, ensure_ascii=False, indent=4, separators=(',', ':')))

    return metrics