-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathinput_norm_losses.py
251 lines (203 loc) · 9.49 KB
/
input_norm_losses.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
import math
import numpy as np
import scipy
import torch
import torch.nn as nn
import torch.nn.functional as F
# from functorch import vmap
# veigh = vmap(torch.linalg.eigh)
class DBP(nn.Module):
def __init__(self, eps=4./255., std=0.225) -> None:
super().__init__()
self.eps = eps/std
def forward(self, gradients, inputs):
batch_size = gradients.shape[0]
return self.eps*batch_size*gradients.abs().sum((-3, -2, -1)).mean()
# class DBPAMHM(nn.Module):
# def __init__(self, eps=4./255., std=0.225, w_hm=0.25e1, tol=1e-12) -> None:
# super().__init__()
# self.eps = eps/std
# self.w_hm = w_hm
# self.tol = tol
# def forward(self, gradients, inputs):
# batch_size = gradients.shape[0]
# gradients = self.eps*batch_size*gradients.abs().flatten(1)
# am = gradients.sum(1).mean()
# hm = self.w_hm * (gradients.clamp(min=self.tol).pow(-1).mean(-1).pow(-1).mean() - self.tol)
# return am, hm
# def __repr__(self):
# return f'{self.__class__.__name__}(w_hm={self.w_hm}, tol={self.tol})'
# class DBPSparsity(nn.Module):
# def __init__(self, eps=4./255., std=0.225) -> None:
# super().__init__()
# self.eps = eps/std
# def forward(self, gradients, inputs):
# batch_size = gradients.shape[0]
# l1 = gradients.abs().sum((-3, -2, -1))
# l2 = gradients.abs().square().sum((-3, -2, -1)).sqrt()
# norm_term = self.eps*batch_size*l1.mean()
# sparsity_term = 0.005* (l1/l2).mean()
# return norm_term, sparsity_term
class DBPChannel(nn.Module):
def __init__(self, eps=4./255., std=0.225, weight_r=1., weight_g=2., weight_b=1.) -> None:
super().__init__()
self.eps = eps/std
self.weight_r = weight_r
self.weight_g = weight_g
self.weight_b = weight_b
def forward(self, gradients, inputs):
batch_size = gradients.shape[0]
gradients = self.eps*batch_size*gradients.abs().sum((-2, -1)).mean(0)
return self.weight_r * gradients[0] + self.weight_g * gradients[1] + self.weight_b * gradients[2]
def __repr__(self):
return f'{self.__class__.__name__}(weight_r={self.weight_r}, weight_g={self.weight_g}, weight_b={self.weight_b})'
class DBPThresholded(nn.Module):
def __init__(self, th=0.001, eps=4./255., std=0.225) -> None:
super().__init__()
self.th = th
self.eps = eps/std
def forward(self, gradients, inputs):
batch_size = gradients.shape[0]
return torch.relu(batch_size*self.eps*gradients.abs().sum((-3, -2, -1)) - self.th).mean()
# class DBPPow(nn.Module):
# def __init__(self, eps=4./255., std=0.225, p=0.7, th=0.01, tol=1e-12) -> None:
# super().__init__()
# self.eps = eps/std
# self.p = p
# self.th = th
# self.tol = tol
# def forward(self, gradients, inputs):
# batch_size = gradients.shape[0]
# return torch.relu(self.eps*batch_size*(gradients.abs() + self.tol).pow(self.p).sum((-3, -2, -1)) - self.th).mean()
# def __repr__(self):
# return f'{self.__class__.__name__}(p={self.p}, th={self.th}, tol={self.tol})'
# class DBPEdgeWeight(nn.Module):
# def __init__(self, eps=4./255., std=0.225, theta=0.1, rho=0.5) -> None:
# super().__init__()
# self.eps = eps/std
# self.theta = theta
# self.rho = rho
# self.register_buffer('inner_gaussian', torch.from_numpy(make_gaussian_filter(self.theta)))
# self.register_buffer('left_diff', torch.Tensor([[-1., 1., 0.]]), persistent=False)
# self.register_buffer('right_diff', torch.Tensor([[0., -1., 1.]]), persistent=False)
# self.register_buffer('outer_gaussian', torch.from_numpy(make_gaussian_filter(self.rho)))
# def forward(self, gradients, inputs):
# batch_size = gradients.shape[0]
# with torch.no_grad():
# SI = structure_tensor(inputs, self.inner_gaussian, self.left_diff, self.right_diff, self.outer_gaussian)
# p_edge = get_edge_probability(SI)
# q_edge = 1. - p_edge
# return self.eps*batch_size*(q_edge * gradients).abs().sum((-3, -2, -1)).mean()
# class DBPEdgeWeightNorm(nn.Module):
# def __init__(self, eps=4./255., std=0.225, theta=0.1, rho=0.5) -> None:
# super().__init__()
# self.eps = eps/std
# self.theta = theta
# self.rho = rho
# self.register_buffer('inner_gaussian', torch.from_numpy(make_gaussian_filter(self.theta)))
# self.register_buffer('left_diff', torch.Tensor([[-1., 1., 0.]]), persistent=False)
# self.register_buffer('right_diff', torch.Tensor([[0., -1., 1.]]), persistent=False)
# self.register_buffer('outer_gaussian', torch.from_numpy(make_gaussian_filter(self.rho)))
# def forward(self, gradients, inputs):
# batch_size = gradients.shape[0]
# with torch.no_grad():
# SI = structure_tensor(inputs, self.inner_gaussian, self.left_diff, self.right_diff, self.outer_gaussian)
# p_edge = get_edge_probability(SI)
# q_edge = 1. - p_edge
# q_edge = q_edge / q_edge.mean((-3, -2, -1), keepdim=True)
# return self.eps*batch_size*(q_edge * gradients).abs().sum((-3, -2, -1)).mean()
# class DBPTangent(nn.Module):
# def __init__(self, eps=4./255., std=0.225, theta=0.1, rho=0.5) -> None:
# super().__init__()
# self.eps = eps/std
# self.theta = theta
# self.rho = rho
# self.register_buffer('inner_gaussian', torch.from_numpy(make_gaussian_filter(self.theta)))
# self.register_buffer('left_diff', torch.Tensor([[-1., 1., 0.]]), persistent=False)
# self.register_buffer('right_diff', torch.Tensor([[0., -1., 1.]]), persistent=False)
# self.register_buffer('outer_gaussian', torch.from_numpy(make_gaussian_filter(self.rho)))
# def forward(self, gradients, inputs):
# SI = structure_tensor(inputs, self.inner_gaussian, self.left_diff, self.right_diff, self.outer_gaussian)
# #print(SI.shape)
# tangential = get_tangential_direction(SI)
# #print(tangential.shape)
# gradients_dx = filter_2d_with_reflect_pad(gradients, self.left_diff)
# gradients_dy = filter_2d_with_reflect_pad(gradients, self.left_diff.T)
# gradients_dtan = tangential[:, None, ..., 0]*gradients_dx + tangential[:, None, ..., 1]*gradients_dy
# batch_size = gradients.shape[0]
# d0_term = batch_size*self.eps*gradients.abs().sum((-3, -2, -1)).mean()
# d1_term = batch_size*self.eps*gradients_dtan.abs().sum((-3, -2, -1)).mean()
# return d0_term, d1_term
# class DBPChange(nn.Module):
# def __init__(self, eps=4./255., std=0.225) -> None:
# super().__init__()
# self.eps = eps/std
# self.register_buffer('left_diff', torch.Tensor([[-1., 1., 0.]]), persistent=False)
# self.register_buffer('right_diff', torch.Tensor([[0., -1., 1.]]), persistent=False)
# def forward(self, gradients, inputs):
# batch_size = gradients.shape[0]
# gradients_dx = filter_2d_with_reflect_pad(gradients, self.left_diff)
# gradients_dy = filter_2d_with_reflect_pad(gradients, self.left_diff.T)
# d0_term = self.eps*batch_size*gradients.abs().sum((-3, -2, -1)).mean()
# d1_term = batch_size*self.eps*0.5*(gradients_dx.abs() + gradients_dy.abs()).sum((-3, -2, -1)).mean()
# return d0_term, d1_term
# ## Structure tensor
# def get_tangential_direction(SI):
# with torch.no_grad():
# eigenvalues, eigenvectors = veigh(SI)
# uv = eigenvectors[..., 0, :] #.chunk(2, dim=-1)
# return uv
# def get_orthogonal_direction(SI):
# with torch.no_grad():
# eigenvalues, eigenvectors = veigh(SI)
# uv = eigenvectors[..., 1, :] #.chunk(2, dim=-1)
# return uv
# def binarize(x, q):
# return torch.where(x.gt(x.flatten(start_dim=1).quantile(q=q, dim=1)[:, None, None, None]), 1., 0.)
# def get_edge_probability(SI, quantiles=[0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9]):
# p_edge = sum(binarize((SI[..., 0, 0] + SI[..., 1, 1]).unsqueeze(-1), q=q) for q in quantiles) / len(quantiles)
# return p_edge[:, None, :, :, 0]
# def xQxt(Q, x):
# Qx = torch.matmul(Q, x.unsqueeze(-1))
# xQx = torch.matmul(x[..., None].transpose(-1, -2), Qx)
# return xQx
# def tr(Q):
# return Q[..., 0, 0] + Q[..., 1, 1]
# def structure_tensor(I, inner_gaussian, left_diff, right_diff, outer_gaussian):
# # Compute spatial derivatives
# I = filter_separable_2d_with_reflect_pad(I, inner_gaussian)
# Ixl = filter_2d_with_reflect_pad(I, left_diff)
# Iyl = filter_2d_with_reflect_pad(I, left_diff.T)
# Ixr = filter_2d_with_reflect_pad(I, right_diff)
# Iyr = filter_2d_with_reflect_pad(I, right_diff.T)
# # Structure tensor components
# IxIx = Ixl*Ixl + Ixr*Ixr
# IxIy = Ixl*Iyl + Ixr*Iyr
# IyIy = Iyl*Iyl + Iyr*Iyr
# # Outer blur
# IxIx = filter_separable_2d_with_reflect_pad(IxIx, outer_gaussian)
# IxIy = filter_separable_2d_with_reflect_pad(IxIy, outer_gaussian)
# IyIy = filter_separable_2d_with_reflect_pad(IyIy, outer_gaussian)
# # Keep only one channel
# IxIx, IxIy, IyIy = IxIx.mean(1), IxIy.mean(1), IyIy.mean(1)
# # Structure tensor matrix
# SI = torch.stack([IxIx, IxIy, IxIy, IyIy], dim=-1).view(*IxIx.shape, 2, 2)
# return SI
def filter_separable_2d_with_reflect_pad(x, h):
x = filter_2d_with_reflect_pad(x, h[None, :])
x = filter_2d_with_reflect_pad(x, h[:, None])
return x
def filter_2d_with_reflect_pad(x, h):
_, C, _, _ = x.shape
p2d = tuple([(h.size(1)-1)//2] * 2 + [(h.size(0)-1)//2] * 2)
x = F.pad(x, p2d, "reflect")
return F.conv2d(x, h[None, None, :, :].expand((C, -1, -1, -1)), padding='valid', groups=C)
def binarize(x, q):
return torch.where(x.gt(x.flatten(start_dim=1).quantile(q=q, dim=1)[:, None, None, None]), 1., 0.)
def make_gaussian_filter(stddev):
order = math.ceil(3*stddev)
n = np.arange(-order, order+1)
h = math.exp(-stddev) * scipy.special.iv(n, stddev)
h = np.array(h, dtype=np.float32)
h = h / np.sum(h)
return h