-
Notifications
You must be signed in to change notification settings - Fork 4
/
partialconv.h
197 lines (162 loc) · 8.01 KB
/
partialconv.h
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
#ifndef PARTIALCONV_H
#define PARTIALCONV_H
// Partial Convolution for PyTorch C++
// -------------------------------------------------------
// Robin Lobel, March 2020 - requires libtorch 1.4 and higher - Qt compatible
// https://github.com/divideconcept/PyTorch-libtorch-U-Net
//
// Partial Convolution fixes the errors introduced by zero-padding convolution, while keeping the convenience of padding
// Partial Mask Convolution supports additionally a mask
// Adapted from https://github.com/NVIDIA/partialconv
// NVIDIA Paper: https://arxiv.org/pdf/1811.11718.pdf
//Qt compatibility
#ifdef QT_CORE_LIB
#undef slots
#endif
#include<torch/torch.h>
#ifdef QT_CORE_LIB
#define slots Q_SLOTS
#endif
class PartialMaskConv2dImpl : public torch::nn::Conv2dImpl {
public:
PartialMaskConv2dImpl(torch::nn::Conv2dOptions options_, bool multi_channel=false) : torch::nn::Conv2dImpl(options_)
{
// whether the mask is multi-channel or not
this->multi_channel = multi_channel;
if(multi_channel)
weight_maskUpdater = torch::ones({options.out_channels(), options.in_channels(), options.kernel_size()->at(0), options.kernel_size()->at(1)});
else
weight_maskUpdater = torch::ones({1, 1, options.kernel_size()->at(0), options.kernel_size()->at(1)});
slide_winsize = weight_maskUpdater.size(1) * weight_maskUpdater.size(2) * weight_maskUpdater.size(3);
last_size={0,0,0,0};
update_mask = torch::Tensor();
mask_ratio = torch::Tensor();
}
std::tuple<torch::Tensor,torch::Tensor> forward(std::tuple<torch::Tensor,torch::Tensor> inputs) {
torch::Tensor input=std::get<0>(inputs); torch::Tensor mask_in=std::get<1>(inputs);
torch::Tensor mask;
if(mask_in.defined() || last_size != input.sizes())
{
last_size=input.sizes();
{
torch::NoGradGuard no_grad;
if(weight_maskUpdater.scalar_type()!=input.scalar_type() || weight_maskUpdater.device()!=input.device())
weight_maskUpdater = weight_maskUpdater.to(input);
if(!mask_in.defined())
{
// if mask is not provided, create a mask
if(multi_channel)
mask = torch::ones({input.data().size(0), input.data().size(1), input.data().size(2), input.data().size(3)}).to(input);
else
mask = torch::ones({1, 1, input.data().size(2), input.data().size(3)}).to(input);
} else {
mask = mask_in;
}
update_mask = torch::nn::functional::conv2d(mask, weight_maskUpdater, torch::nn::functional::Conv2dFuncOptions().bias(torch::Tensor()).stride(options.stride()).padding(options.padding()).dilation(options.dilation()).groups(1));
if(!mask_in.defined())
mask_ratio = slide_winsize/update_mask;
else {
mask_ratio = slide_winsize/(update_mask + (mask.type().scalarType()==torch::kFloat?1e-8:1e-6));
update_mask = torch::clamp(update_mask, 0, 1);
mask_ratio = torch::mul(mask_ratio, update_mask);
}
}
}
torch::Tensor raw_out = Conv2dImpl::forward(mask_in.defined()?torch::mul(input, mask):input);
torch::Tensor output;
if(this->bias.defined())
{
torch::Tensor bias_view = this->bias.view({1, options.out_channels(), 1, 1});
output = torch::mul(raw_out - bias_view, mask_ratio) + bias_view;
if(mask_in.defined()) output = torch::mul(output, update_mask);
} else
output = torch::mul(raw_out, mask_ratio);
return std::make_tuple(output, update_mask);
}
private:
bool multi_channel;
torch::Tensor weight_maskUpdater;
float slide_winsize;
torch::IntArrayRef last_size;
torch::Tensor update_mask;
torch::Tensor mask_ratio;
};
TORCH_MODULE(PartialMaskConv2d);
class PartialConv2dImpl : public PartialMaskConv2dImpl {
public:
PartialConv2dImpl(torch::nn::Conv2dOptions options_) : PartialMaskConv2dImpl(options_,false) { }
torch::Tensor forward(torch::Tensor input) { return std::get<0>(PartialMaskConv2dImpl::forward(std::make_tuple(input,torch::Tensor()))); }
};
TORCH_MODULE(PartialConv2d);
class PartialMaskConv1dImpl : public torch::nn::Conv1dImpl {
public:
PartialMaskConv1dImpl(torch::nn::Conv1dOptions options_, bool multi_channel=false) : torch::nn::Conv1dImpl(options_)
{
// whether the mask is multi-channel or not
this->multi_channel = multi_channel;
if(multi_channel)
weight_maskUpdater = torch::ones({options.out_channels(), options.in_channels(), options.kernel_size()->at(0)});
else
weight_maskUpdater = torch::ones({1, 1, options.kernel_size()->at(0)});
slide_winsize = weight_maskUpdater.size(1) * weight_maskUpdater.size(2);
last_size={0,0,0};
update_mask = torch::Tensor();
mask_ratio = torch::Tensor();
}
std::tuple<torch::Tensor,torch::Tensor> forward(std::tuple<torch::Tensor,torch::Tensor> inputs) {
torch::Tensor input=std::get<0>(inputs); torch::Tensor mask_in=std::get<1>(inputs);
torch::Tensor mask;
if(mask_in.defined() || last_size != input.sizes())
{
last_size=input.sizes();
{
torch::NoGradGuard no_grad;
if(weight_maskUpdater.scalar_type()!=input.scalar_type() || weight_maskUpdater.device()!=input.device())
weight_maskUpdater = weight_maskUpdater.to(input);
if(!mask_in.defined())
{
// if mask is not provided, create a mask
if(multi_channel)
mask = torch::ones({input.data().size(0), input.data().size(1), input.data().size(2)}).to(input);
else
mask = torch::ones({1, 1, input.data().size(2)}).to(input);
} else {
mask = mask_in;
}
update_mask = torch::nn::functional::conv1d(mask, weight_maskUpdater, torch::nn::functional::Conv1dFuncOptions().bias(torch::Tensor()).stride(options.stride()).padding(options.padding()).dilation(options.dilation()).groups(1));
if(!mask_in.defined())
mask_ratio = slide_winsize/update_mask;
else {
mask_ratio = slide_winsize/(update_mask + (mask.type().scalarType()==torch::kFloat?1e-8:1e-6));
update_mask = torch::clamp(update_mask, 0, 1);
mask_ratio = torch::mul(mask_ratio, update_mask);
}
}
}
torch::Tensor raw_out = Conv1dImpl::forward(mask_in.defined()?torch::mul(input, mask):input);
torch::Tensor output;
if(this->bias.defined())
{
torch::Tensor bias_view = this->bias.view({1, options.out_channels(), 1});
output = torch::mul(raw_out - bias_view, mask_ratio) + bias_view;
if(mask_in.defined()) output = torch::mul(output, update_mask);
} else
output = torch::mul(raw_out, mask_ratio);
return std::make_tuple(output, update_mask);
}
private:
bool multi_channel;
torch::Tensor weight_maskUpdater;
float slide_winsize;
torch::IntArrayRef last_size;
torch::Tensor update_mask;
torch::Tensor mask_ratio;
};
TORCH_MODULE(PartialMaskConv1d);
class PartialConv1dImpl : public PartialMaskConv1dImpl {
public:
PartialConv1dImpl(torch::nn::Conv1dOptions options_) : PartialMaskConv1dImpl(options_,false) { }
torch::Tensor forward(torch::Tensor input) { return std::get<0>(PartialMaskConv1dImpl::forward(std::make_tuple(input,torch::Tensor()))); }
};
TORCH_MODULE(PartialConv1d);
#endif // PARTIALCONV_H