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
|
#include "caffe2/operators/batch_sparse_to_dense_op.h"
namespace caffe2 {
template <>
template <typename TLen, typename TInd>
void BatchSparseToDenseOp<float, CPUContext>::FillInDenseValues(
const int64_t batch_size,
const int64_t indice_lengths,
const TLen* lengths_data,
const TInd* indices_data,
const float* values_data,
float* output_data,
CPUContext* /*context*/) {
TLen lengths_sum = 0;
math::Sum<TLen, CPUContext>(
batch_size, lengths_data, &lengths_sum, &context_);
CAFFE_ENFORCE_EQ(lengths_sum, indice_lengths);
int64_t k = 0;
for (int64_t i = 0; i < batch_size; ++i) {
for (int64_t j = 0; j < lengths_data[i]; ++j) {
CAFFE_ENFORCE(
indices_data[k] < dense_last_dim_,
"An indice (",
indices_data[k],
") is larger then last dim of dense (",
dense_last_dim_,
").");
output_data[i * dense_last_dim_ + indices_data[k]] = values_data[k];
k += 1;
}
}
}
template <>
template <typename TLen, typename TInd>
void BatchDenseToSparseOp<float, CPUContext>::FillInSparseValues(
const int64_t batch_size,
const int64_t indice_lengths,
const TLen* lengths_data,
const TInd* indices_data,
const float* dense_data,
float* output_data,
CPUContext* /*context*/) {
TLen lengths_sum = 0;
math::Sum<TLen, CPUContext>(
batch_size, lengths_data, &lengths_sum, &context_);
CAFFE_ENFORCE_EQ(lengths_sum, indice_lengths);
int64_t k = 0;
for (int64_t i = 0; i < batch_size; ++i) {
for (int64_t j = 0; j < lengths_data[i]; ++j) {
CAFFE_ENFORCE(
indices_data[k] < dense_last_dim_,
"An indice (",
indices_data[k],
") is larger then last dim of dense (",
dense_last_dim_,
").");
output_data[k] = dense_data[i * dense_last_dim_ + indices_data[k]];
k += 1;
}
}
}
REGISTER_CPU_OPERATOR(
BatchSparseToDense,
BatchSparseToDenseOp<float, CPUContext>);
OPERATOR_SCHEMA(BatchSparseToDense)
.NumInputs(3, 4)
.NumOutputs(1)
.DisallowInputFillers() // TODO: enable the filler
.SetDoc(R"DOC(
Convert sparse matrix representation into dense matrix.
A sparse matrix is represented by `lengths` vector, `indices` vector,
and `values` vector. Each element in `lengths` vector (lengths[`i`]) represents
the number of indices in this batch (batch `i`).
With in each batch, `indices` should not have duplicate number.
For example, with input:
lengths = [2, 3, 1]
indices = [0, 1, 2, 3, 4, 5]
values = [6, 7, 8, 9, 10, 11]
dense_dim = 6
default_value = 0
The output is:
output = [[6, 7, 0, 0, 0, 0],
[0, 0, 8, 9, 10, 0],
[0, 0, 0, 0, 0, 11]]
after running this operator.
)DOC")
.Input(
0,
"lengths",
"Flatten tensor, used to break down indices and values into per batch indices and values.")
.Input(
1,
"indices",
"Flatten tensor of total size = \\sum lengths, containing the indices ")
.Input(2, "values", "Data tensor, dimension has to match `indices`")
.Input(
3,
"output_shape_inference",
"Optional, a dense tensor whose shape define the output shape")
.Output(
0,
"dense",
"2-D dense tensor, with 1st dim = len(lengths), 2nd dim = dense_last_dim"
"in the arg list, the tensor is of the same data type as `values`."
"Missing values are filled with default_value")
.TensorInferenceFunction([](const OperatorDef& def,
const vector<TensorShape>& in) {
ArgumentHelper helper(def);
vector<long> output_dims;
if (in.size() == 4) {
const auto& inference_dims = GetDimsVector(in[3]);
output_dims.insert(output_dims.end(), inference_dims.begin(), inference_dims.end());
const int dense_last_dim = helper.GetSingleArgument<int>("dense_last_dim", 0);
if(dense_last_dim > 0) {
CAFFE_ENFORCE(
output_dims.back() == dense_last_dim,
"The last dim of output_shape_inference should be consistent with dense_last_dim");
}
} else {
const int dense_last_dim = helper.GetSingleArgument<int>("dense_last_dim", 0);
CAFFE_ENFORCE(
dense_last_dim > 0,
"dense_last_dim must be set when output shape inference is unavailable");
const auto& lens_dims = GetDimsVector(in[0]);
output_dims.insert(output_dims.end(), lens_dims[0]);
output_dims.insert(output_dims.end(), dense_last_dim);
}
vector<TensorShape> out(1);
out[0] = CreateTensorShape(output_dims, in[2].data_type());
return out;
})
.Arg(
"dense_last_dim",
"Optional, output dense last dimension. "
"If both this argument and output_shape_inference are set, "
"it should be consistent with output_shape_inference's last dim")
.Arg(
"default_value",
"Optional, missing values are filled with this value."
"default_value = 0 when not set");
REGISTER_CPU_OPERATOR(
BatchDenseToSparse,
BatchDenseToSparseOp<float, CPUContext>);
OPERATOR_SCHEMA(BatchDenseToSparse)
.NumInputs(3)
.NumOutputs(1)
.SetDoc(R"DOC(
This Op is a inverse of BatchSparseToDenseOp.
Basically, given a `lengths` vector, a `indices` vector,
and a dense matrix `dense`, output `value` vector so that, along with
`lengths` vector and `indices` vector, forms a sparse representation
of the dense matrix.
A sparse matrix is represented by `lengths` vector, `indices` vector,
and `values` vector. Each element in `lengths` vector (lengths[`i`]) represents
the number of indices in this batch (batch `i`).
With in each batch, `indices` should not have duplicate number.
For example, with input:
lengths = [2, 3, 1]
indices = [0, 1, 2, 3, 4, 5]
output = [[6, 7, 0, 0, 0, 0],
[0, 0, 8, 9, 10, 0],
[0, 0, 0, 0, 0, 11]]
The output is:
values = [6, 7, 8, 9, 10, 11]
after running this operator.
)DOC")
.Input(
0,
"lengths",
"Flatten lengths, Used to break down indices into per batch indices")
.Input(
1,
"indices",
"Flatten indices, tensor of total size = \\sum lengths, containing the indices ")
.Input(
2,
"dense",
"dense 2-D tensor, first dim = len(lengths), last dim > Any(indices)")
.Output(
0,
"values",
"Values, tensor of the same size as `indices` and same data type as dense tensor.");
namespace {
class GetBatchSparseToDenseGradient : public GradientMakerBase {
using GradientMakerBase::GradientMakerBase;
vector<OperatorDef> GetGradientDefs() override {
return SingleGradientDef(
"BatchDenseToSparse",
"",
vector<string>{I(0), I(1), GO(0)},
vector<string>{GI(2)});
}
};
class GetBatchDenseToSparseGradient : public GradientMakerBase {
using GradientMakerBase::GradientMakerBase;
vector<OperatorDef> GetGradientDefs() override {
return SingleGradientDef(
"BatchSparseToDense",
"",
vector<string>{I(0), I(1), GO(0), I(2)},
vector<string>{GI(2)});
}
};
REGISTER_GRADIENT(BatchSparseToDense, GetBatchSparseToDenseGradient);
REGISTER_GRADIENT(BatchDenseToSparse, GetBatchDenseToSparseGradient);
} // namespace
} // namespace caffe2
|
