forked from styler00dollar/pytorch-loss-functions
/
example.py
139 lines (92 loc) · 4.77 KB
/
example.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
"""
This is just an example and should show basic usage
Write proper syntax yourself / Add it to existing code
Examples with working code:
DFNet:
https://github.com/styler00dollar/Colab-DFNet
deepfillv2:
https://github.com/styler00dollar/Colab-mmediting/tree/master/mmedit/models/inpaintors
"""
from vic.loss import CharbonnierLoss, GANLoss, GradientPenaltyLoss, HFENLoss, TVLoss, GradientLoss, ElasticLoss, RelativeL1, L1CosineSim, ClipL1, MaskedL1Loss, MultiscalePixelLoss, FFTloss, OFLoss, L1_regularization, ColorLoss, AverageLoss, GPLoss, CPLoss, SPL_ComputeWithTrace, SPLoss, Contextual_Loss, StyleLoss
from vic.perceptual_loss import PerceptualLoss
from vic.filters import *
from vic.colors import *
from vic.discriminators import *
from .diffaug import *
from metrics import *
from tensorboardX import SummaryWriter
logdir='/path/'
writer = SummaryWriter(logdir=logdir)
class CalcLoss(nn.Module):
def __init__(self):
super().__init__()
"""
if self.config.HFEN_TYPE == 'L1':
l_hfen_type = nn.L1Loss()
if self.config.HFEN_TYPE == 'MSE':
l_hfen_type = nn.MSELoss()
if self.config.HFEN_TYPE == 'Charbonnier':
l_hfen_type = CharbonnierLoss()
if self.config.HFEN_TYPE == 'ElasticLoss':
l_hfen_type = ElasticLoss()
if self.config.HFEN_TYPE == 'RelativeL1':
l_hfen_type = RelativeL1()
if self.config.HFEN_TYPE == 'L1CosineSim':
l_hfen_type = L1CosineSim()
"""
l_hfen_type = L1CosineSim()
self.HFENLoss = HFENLoss(loss_f=l_hfen_type, kernel='log', kernel_size=15, sigma = 2.5, norm = False)
self.ElasticLoss = ElasticLoss(a=0.2, reduction='mean')
self.RelativeL1 = RelativeL1(eps=.01, reduction='mean')
self.L1CosineSim = L1CosineSim(loss_lambda=5, reduction='mean')
self.ClipL1 = ClipL1(clip_min=0.0, clip_max=10.0)
self.FFTloss = FFTloss(loss_f = torch.nn.L1Loss, reduction='mean')
self.OFLoss = OFLoss()
self.GPLoss = GPLoss(trace=False, spl_denorm=False)
self.CPLoss = CPLoss(rgb=True, yuv=True, yuvgrad=True, trace=False, spl_denorm=False, yuv_denorm=False)
self.StyleLoss = StyleLoss() # Warning: does not support AMP
self.TVLoss = TVLoss(tv_type='tv', p = 1)
self.PerceptualLoss = PerceptualLoss(model='net-lin', net='alex', colorspace='rgb', spatial=False, use_gpu=True, gpu_ids=[0], model_path=None)
layers_weights = {'conv_1_1': 1.0, 'conv_3_2': 1.0}
self.Contextual_Loss = Contextual_Loss(layers_weights, crop_quarter=False, max_1d_size=100,
distance_type = 'cosine', b=1.0, band_width=0.5,
use_vgg = True, net = 'vgg19', calc_type = 'regular')
self.psnr_metric = PSNR()
self.ssim_metric = SSIM()
self.ae_metric = AE()
self.mse_metric = MSE()
def forward(self, input, gt, iteration):
# out is generated output, gt_res is original image
# loss functions
# You can also add them to tensorboard with writer.add_scalar
HFENLoss_forward = self.HFENLoss(out, gt_res)
ElasticLoss_forward = self.ElasticLoss(out, gt_res)
RelativeL1_forward = self.RelativeL1(out, gt_res)
L1CosineSim_forward += 6*self.L1CosineSim(out, gt_res)
ClipL1_forward = self.ClipL1(out, gt_res)
FFTloss_forward = self.FFTloss(out, gt_res)
OFLoss_forward = self.OFLoss(out)
GPLoss_forward = self.GPLoss(out, gt_res)
CPLoss_forward = self.CPLoss(out, gt_res)
Contextual_Loss_forward = self.Contextual_Loss(out, gt_res)
style_forward = self.StyleLoss(out, gt_res) # Warning: does not support AMP
tv_forward = self.TVLoss(out)
total_loss = tv_forward
perceptual_forward = self.PerceptualLoss(out, gt_res)
# improving GANloss by adding DiffAug
# Discriminator loop
out = Discriminator(DiffAugment(out, policy=policy))
gt_res = Discriminator(DiffAugment(gt_res, policy=policy))
# calc GANloss
# Generator loop
out = Discriminator(DiffAugment(out, policy=policy))
# calc GANloss
# writing metrics to tensorboard file
# PSNR (Peak Signal-to-Noise Ratio)
writer.add_scalar('metrics/PSNR', self.psnr_metric(gt_res, out), iteration)
# SSIM (Structural Similarity)
writer.add_scalar('metrics/SSIM', self.ssim_metric(gt_res, out), iteration)
# AE (Average Angular Error)
writer.add_scalar('metrics/AE', self.ae_metric(gt_res, out), iteration)
# MSE (Mean Square Error)
writer.add_scalar('metrics/MSE', self.mse_metric(gt_res, out), iteration)