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
|
/*
* Copyright (c) 2019 Samsung Electronics Co., Ltd. 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 "BatchNormalization.h"
#include "ONNXHelpers.h"
#include "AttributeHelpers.h"
#include "mir/ShapeRange.h"
#include "mir/Tensor.h"
#include "mir/ops/AddOp.h"
#include "mir/ops/ConstantOp.h"
#include "mir/ops/MulOp.h"
#include "mir/ops/ReshapeOp.h"
#include <cmath>
namespace mir_onnx
{
void convertBatchNormalizationV1(const onnx::NodeProto &onnx_node, ConverterContext *context)
{
// consumed_inputs attribute not used
convertBatchNormalizationV6(onnx_node, context);
}
void convertBatchNormalizationV6(const onnx::NodeProto &onnx_node, ConverterContext *context)
{
const auto is_test = getAttributeValue<std::int64_t>(onnx_node, "is_test", 0);
if (is_test == 0)
throw std::runtime_error("Not supported is_test attribute!");
convertBatchNormalizationV7(onnx_node, context);
}
void convertBatchNormalizationV7(const onnx::NodeProto &onnx_node, ConverterContext *context)
{
// spatial attribute used only for learning
convertBatchNormalizationV9(onnx_node, context);
}
void convertBatchNormalizationV9(const onnx::NodeProto &onnx_node, ConverterContext *context)
{
// momentum attrribute used only for learning
std::vector<mir::Operation::Output *> inputs = context->getNodeInputs(onnx_node);
mir::Graph *graph = context->getGraph();
assert(inputs.size() == 5);
auto input = inputs[0];
auto scale = inputs[1];
auto bias = inputs[2];
auto mean = inputs[3];
auto var = inputs[4];
// 1e-05f is the default epsilon.
const auto epsilon = getAttributeValue<float>(onnx_node, "epsilon", 1e-05f);
// Y = (X - mean) * scale / sqrt(var + epsilon) + bias =
// = (X + C1) * C2 + bias
// We need these to be constants since we are going to change them.
// TODO Implement the formula using ops and let the optimizer constant-fold them.
auto scale_op = dynamic_cast<mir::ops::ConstantOp *>(scale->getNode());
auto mean_op = dynamic_cast<mir::ops::ConstantOp *>(mean->getNode());
auto var_op = dynamic_cast<mir::ops::ConstantOp *>(var->getNode());
if (scale_op == nullptr || mean_op == nullptr || var_op == nullptr)
throw std::runtime_error(
"BatchNormalization: only constant 'scale', 'mean' and 'variance' inputs are supported.");
mir::Tensor<float> scale_accessor(scale_op->getValue());
mir::Tensor<float> mean_accessor(mean_op->getValue());
mir::Tensor<float> var_accessor(var_op->getValue());
// C1 = -mean
for (const auto &idx : mir::ShapeRange(mean_accessor.getShape()))
mean_accessor.at(idx) *= -1;
// C2 = scale / sqrt(var + epsilon)
for (const auto &idx : mir::ShapeRange(scale_accessor.getShape()))
scale_accessor.at(idx) /= std::sqrt(var_accessor.at(idx) + epsilon);
assert(mean_accessor.getShape().rank() == 1);
auto input_rank = input->getShape().rank();
if (input_rank < 2)
throw std::runtime_error("Inputs with shape rank < 2 are not supported for batchnorm");
mir::Shape new_shape(std::vector<std::int32_t>(input_rank, 1));
new_shape.dim(1) = mean_accessor.getShape().dim(0); // set channel dim
auto reshaped_mean = createOp<mir::ops::ReshapeOp>(graph, mean, new_shape)->getOutput(0);
auto reshaped_scale = createOp<mir::ops::ReshapeOp>(graph, scale, new_shape)->getOutput(0);
auto reshaped_bias = createOp<mir::ops::ReshapeOp>(graph, bias, new_shape)->getOutput(0);
// Y = (X + C1) * C2 + bias
auto result = createOp<mir::ops::AddOp>(graph, input, reshaped_mean)->getOutput(0);
result = createOp<mir::ops::MulOp>(graph, result, reshaped_scale)->getOutput(0);
result = createOp<mir::ops::AddOp>(graph, result, reshaped_bias)->getOutput(0);
context->setNodeOutputs(onnx_node, {result});
}
} // namespace mir_onnx
|
