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
|
/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/
#include <memory>
#include <string>
#include <unordered_map>
#include <vector>
#include "tensorflow/contrib/lite/toco/graph_transformations/graph_transformations.h"
#include "tensorflow/contrib/lite/toco/model.h"
#include "tensorflow/contrib/lite/toco/runtime/types.h"
#include "tensorflow/contrib/lite/toco/tooling_util.h"
#include "tensorflow/core/platform/logging.h"
namespace toco {
::tensorflow::Status ResolveBatchNormalization::Run(Model* model,
std::size_t op_index,
bool* modified) {
*modified = false;
auto bn_it = model->operators.begin() + op_index;
if (bn_it->get()->type != OperatorType::kBatchNormalization) {
return ::tensorflow::Status::OK();
}
const auto* bn_op =
static_cast<const BatchNormalizationOperator*>(bn_it->get());
auto& mean_array = model->GetArray(bn_op->inputs[1]);
const auto& multiplier_array = model->GetArray(bn_op->inputs[2]);
const auto& offset_array = model->GetArray(bn_op->inputs[3]);
CHECK(IsConstantParameterArray(*model, bn_op->inputs[1]) &&
IsConstantParameterArray(*model, bn_op->inputs[2]) &&
IsConstantParameterArray(*model, bn_op->inputs[3]))
<< "Batch normalization resolution requires that mean, multiplier and "
"offset arrays be constant.";
// We should only have *float* BatchNormalizations... let's guard this
// assumption by CHECK's.
CHECK(mean_array.data_type == ArrayDataType::kFloat);
CHECK(multiplier_array.data_type == ArrayDataType::kFloat);
CHECK(offset_array.data_type == ArrayDataType::kFloat);
// This graph transformations will need to address constant buffers below,
// so we need to exit early if these buffers don't exist (i.e. if the params
// haven't yet been resolved as constants).
if (!mean_array.buffer || !multiplier_array.buffer || !offset_array.buffer) {
return ::tensorflow::Status::OK();
}
// Create the new Mul, Add operators
auto* mul_op = new MulOperator;
auto* add_op = new AddOperator;
const string mul_name =
AvailableArrayName(*model, bn_op->outputs[0] + "_mul");
const string add_name =
AvailableArrayName(*model, bn_op->outputs[0] + "_add");
const string mul_param_name = AvailableArrayName(*model, mul_name + "_param");
const string add_param_name = AvailableArrayName(*model, add_name + "_param");
mul_op->inputs = {bn_op->inputs[0], mul_param_name};
mul_op->outputs = {mul_name};
add_op->inputs = {mul_name, add_param_name};
add_op->outputs = {bn_op->outputs[0]};
AddMessageF("Splitting %s into %s and %s", LogName(*bn_op), LogName(*mul_op),
LogName(*add_op));
// Create the intermediate activation array (output of mul, input of add)
auto& intermediate_array = model->GetOrCreateArray(mul_op->outputs[0]);
intermediate_array.data_type = model->GetArray(bn_op->inputs[0]).data_type;
// Insert the new operators in the graph
auto add_it = model->operators.emplace(bn_it, add_op);
auto mul_it = model->operators.emplace(add_it, mul_op);
// update invalidated iterators.
DCHECK_EQ(mul_it->get(), mul_op);
add_it = mul_it + 1;
DCHECK_EQ(add_it->get(), add_op);
bn_it = add_it + 1;
DCHECK_EQ(bn_it->get(), bn_op);
// Create the new param arrays
auto& mean_shape = *mean_array.mutable_shape();
const auto& multiplier_shape = multiplier_array.shape();
const auto& offset_shape = offset_array.shape();
if (mean_shape.dims().empty()) {
*mean_shape.mutable_dims() = multiplier_shape.dims();
auto& data = mean_array.GetMutableBuffer<ArrayDataType::kFloat>().data;
CHECK_EQ(data.size(), 1);
data.resize(RequiredBufferSizeForShape(mean_shape), data[0]);
}
CHECK(mean_shape.dims() == multiplier_shape.dims());
CHECK(mean_shape.dims() == offset_shape.dims());
const auto& param_shape = mean_shape;
const int buffer_size = RequiredBufferSizeForShape(param_shape);
auto& mul_param_array = model->GetOrCreateArray(mul_param_name);
auto& add_param_array = model->GetOrCreateArray(add_param_name);
DropMinMax(model, mul_param_name);
DropMinMax(model, add_param_name);
mul_param_array.copy_shape(param_shape);
add_param_array.copy_shape(param_shape);
mul_param_array.data_type = ArrayDataType::kFloat;
add_param_array.data_type = ArrayDataType::kFloat;
auto& mul_float_data =
mul_param_array.GetMutableBuffer<ArrayDataType::kFloat>().data;
auto& add_float_data =
add_param_array.GetMutableBuffer<ArrayDataType::kFloat>().data;
mul_float_data.resize(buffer_size);
add_float_data.resize(buffer_size);
const auto& mean_float_data =
mean_array.GetBuffer<ArrayDataType::kFloat>().data;
const auto& multiplier_float_data =
multiplier_array.GetBuffer<ArrayDataType::kFloat>().data;
const auto& offset_float_data =
offset_array.GetBuffer<ArrayDataType::kFloat>().data;
CHECK(mul_float_data.size() == buffer_size);
CHECK(add_float_data.size() == buffer_size);
CHECK(mean_float_data.size() == buffer_size);
CHECK(multiplier_float_data.size() == buffer_size);
CHECK(offset_float_data.size() == buffer_size);
for (int i = 0; i < buffer_size; i++) {
mul_float_data[i] = multiplier_float_data[i];
add_float_data[i] =
offset_float_data[i] - mean_float_data[i] * multiplier_float_data[i];
}
// Remove the old param arrays
DeleteArrayIfUsedOnce(bn_op->inputs[1], model);
DeleteArrayIfUsedOnce(bn_op->inputs[2], model);
DeleteArrayIfUsedOnce(bn_op->inputs[3], model);
// Remove the old operator
DCHECK_EQ(bn_it->get(), bn_op);
model->operators.erase(bn_it);
*modified = true;
return ::tensorflow::Status::OK();
}
} // namespace toco
|
