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
|
#ifndef CAFFE2_OPERATORS_EXPAND_OP_H_
#define CAFFE2_OPERATORS_EXPAND_OP_H_
#include <vector>
#include "caffe2/core/context.h"
#include "caffe2/core/operator.h"
#include "caffe2/core/types.h"
#include "caffe2/utils/math.h"
#include "c10/util/irange.h"
namespace caffe2 {
template <typename InputTypes, class Context>
class ExpandOp final : public Operator<Context> {
public:
USE_OPERATOR_CONTEXT_FUNCTIONS;
template <class... Args>
explicit ExpandOp(Args&&... args)
: Operator<Context>(std::forward<Args>(args)...),
OP_SINGLE_ARG(bool, "allow_broadcast_fastpath", allow_broadcast_fastpath_, false) {}
bool RunOnDevice() override {
return DispatchHelper<InputTypes>::call(this, Input(0));
}
template <typename T>
bool DoRunWithType() {
const auto& X = Input(0);
const auto& Y_shape_tensor = Input(1);
std::vector<int64_t> shape_dims(Y_shape_tensor.numel());
context_.template CopyToCPU<int64_t>(
Y_shape_tensor.numel(),
Y_shape_tensor.template data<int64_t>(),
shape_dims.data());
const int ndim = shape_dims.size();
const std::vector<int> X_dims(X.sizes().cbegin(), X.sizes().cend());
std::vector<int> Y_dims;
Y_dims.reserve(std::max(ndim, X.dim()));
// ndim, X.ndim() might equal to 0
for (int i = ndim - 1, j = X.dim() - 1; i >= 0 || j >= 0; --i, --j) {
const int shape_x = (j >= 0 ? X_dims[j] : 1);
// In PyTorch expand treats -1 as a special value to indicate
// preserving the size of that dimension.
const int shape_y = ((i >= 0 && shape_dims[i] > 0) ? shape_dims[i] : 1);
CAFFE_ENFORCE(
shape_x == 1 || shape_y == 1 || shape_x == shape_y,
"Dimensions format invalid.");
Y_dims.push_back(std::max(shape_x, shape_y));
}
std::reverse(Y_dims.begin(), Y_dims.end());
// TODO: remove when the function in math are changed to use vector<int64_t>
std::vector<int64_t> Y_dims_int64;
std::copy(Y_dims.begin(), Y_dims.end(), std::back_inserter(Y_dims_int64));
auto* Y = Output(0, Y_dims_int64, at::dtype<T>());
math::Broadcast<T, Context>(
X_dims.size(),
X_dims.data(),
Y_dims.size(),
Y_dims.data(),
T(1),
X.template data<T>(),
Y->template mutable_data<T>(),
&context_,
allow_broadcast_fastpath_);
return true;
}
const bool allow_broadcast_fastpath_;
};
template <typename InputTypes, class Context>
class ExpandGradientOp final : public Operator<Context> {
public:
USE_OPERATOR_CONTEXT_FUNCTIONS;
template <class... Args>
explicit ExpandGradientOp(Args&&... args)
: Operator<Context>(std::forward<Args>(args)...),
OP_SINGLE_ARG(bool, "allow_broadcast_fastpath", allow_broadcast_fastpath_, false) {}
bool RunOnDevice() override {
return DispatchHelper<InputTypes>::call(this, Input(0));
}
template <typename T>
bool DoRunWithType() {
const auto& dY = Input(0);
const auto& X = Input(1);
const int ndim = dY.dim();
const std::vector<int> dX_dims(X.sizes().cbegin(), X.sizes().cend());
const std::vector<int> dY_dims(dY.sizes().cbegin(), dY.sizes().cend());
auto* dX = Output(0, X.sizes(), at::dtype<T>());
std::vector<int> axes;
const int offset = ndim - X.dim();
for (const auto i : c10::irange(ndim)) {
if (i < offset || dX_dims[i - offset] == 1) {
axes.push_back(i);
}
}
std::vector<int> X_dims = dY_dims;
for (const int axis : axes) {
X_dims[axis] = 1;
}
math::ReduceSum<T, Context>(
dY_dims.size(),
dY_dims.data(),
X_dims.data(),
T(1),
dY.template data<T>(),
dX->template mutable_data<T>(),
&context_,
allow_broadcast_fastpath_);
return true;
}
const bool allow_broadcast_fastpath_;
};
} // namespace caffe2
#endif // CAFFE2_OPERATORS_REDUCE_OPS_H_
|
