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
|
// Copyright 2004-present Facebook. All Rights Reserved.
#ifndef CAFFE2_OPERATORS_BATCH_SPARSE_TO_DENSE_OP_H_
#define CAFFE2_OPERATORS_BATCH_SPARSE_TO_DENSE_OP_H_
#include "caffe2/core/context.h"
#include "caffe2/core/operator.h"
#include "caffe2/utils/math.h"
namespace caffe2 {
template <typename T, class Context>
class BatchSparseToDenseOp : public Operator<Context> {
public:
USE_OPERATOR_CONTEXT_FUNCTIONS;
template <class... Args>
explicit BatchSparseToDenseOp(Args&&... args)
: Operator<Context>(std::forward<Args>(args)...),
OP_SINGLE_ARG(int64_t, "dense_last_dim", dense_last_dim_, -1),
OP_SINGLE_ARG(T, "default_value", default_value_, static_cast<T>(0)) {}
bool RunOnDevice() {
auto& lengths = Input(LENGTHS);
auto& indices = Input(INDICES);
auto& values = Input(VALUES);
CAFFE_ENFORCE_EQ(indices.numel(), values.numel());
CAFFE_ENFORCE_EQ(lengths.dim(), 1);
CAFFE_ENFORCE_EQ(indices.dim(), 1);
const int64_t* lengths_data = lengths.template data<int64_t>();
const int64_t* indices_data = indices.template data<int64_t>();
const T* values_data = values.template data<T>();
int64_t batch_size = lengths.numel();
vector<int64_t> output_shape = {batch_size};
if (InputSize() == 4) {
auto& shaper = Input(3);
CAFFE_ENFORCE_EQ(shaper.dim(), 2);
if (dense_last_dim_ == -1) {
dense_last_dim_ = shaper.size(1);
} else {
CAFFE_ENFORCE(
dense_last_dim_ == shaper.size(1),
"The last dim argument is not aligned with the shape input last dim");
}
} else {
CAFFE_ENFORCE(dense_last_dim_ >= 1, "The last dim of dense must be >= 1");
}
output_shape.push_back(dense_last_dim_);
auto* output = Output(0, output_shape, at::dtype<T>());
T* output_data = output->template mutable_data<T>();
math::Set(
output->numel(),
static_cast<T>(default_value_),
output_data,
&context_);
FillInDenseValues(
batch_size,
indices.numel(),
lengths_data,
indices_data,
values_data,
output_data,
&context_);
return true;
}
private:
void FillInDenseValues(
const int64_t batch_size,
const int64_t indice_lengths,
const int64_t* lengths_data,
const int64_t* indices_data,
const T* values_data,
T* output_data,
Context* context);
int64_t dense_last_dim_;
T default_value_;
INPUT_TAGS(LENGTHS, INDICES, VALUES);
// len_prefix_sum_ and len_prefix_tmp_ are buffers on the GPU. It is not used
// in the CPUContext implementation.
Tensor len_prefix_sum_{Context::GetDeviceType()};
Tensor len_prefix_tmp_{Context::GetDeviceType()};
};
template <typename T, class Context>
class BatchDenseToSparseOp : public Operator<Context> {
public:
USE_OPERATOR_CONTEXT_FUNCTIONS;
template <class... Args>
explicit BatchDenseToSparseOp(Args&&... args)
: Operator<Context>(std::forward<Args>(args)...) {}
bool RunOnDevice() {
auto& lengths = Input(LENGTHS);
auto& indices = Input(INDICES);
auto& dense = Input(DENSE);
CAFFE_ENFORCE_EQ(lengths.dim(), 1);
CAFFE_ENFORCE_EQ(indices.dim(), 1);
CAFFE_ENFORCE_EQ(dense.dim(), 2);
const int64_t* lengths_data = lengths.template data<int64_t>();
const int64_t* indices_data = indices.template data<int64_t>();
const T* dense_data = dense.template data<T>();
int64_t batch_size = lengths.numel();
CAFFE_ENFORCE_EQ(batch_size, dense.size(0));
dense_last_dim_ = dense.size(1);
vector<int64_t> output_shape = indices.sizes().vec();
auto* output = Output(0, output_shape, at::dtype<T>());
T* output_data = output->template mutable_data<T>();
FillInSparseValues(
batch_size,
indices.numel(),
lengths_data,
indices_data,
dense_data,
output_data,
&context_);
return true;
}
private:
void FillInSparseValues(
const int64_t batch_size,
const int64_t indice_lengths,
const int64_t* lengths_data,
const int64_t* indices_data,
const T* dense_data,
T* output_data,
Context* context);
int64_t dense_last_dim_;
INPUT_TAGS(LENGTHS, INDICES, DENSE);
// len_prefix_sum_ and len_prefix_tmp_ are buffers on the GPU. It is not used
// in the CPUContext implementation.
Tensor len_prefix_sum_{Context::GetDeviceType()};
Tensor len_prefix_tmp_{Context::GetDeviceType()};
};
} // namespace caffe2
#endif // CAFFE2_OPERATORS_BATCH_SPARSE_TO_DENSE_OP_H_
|
