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
233
234
235
236
237
238
239
240
|
/* Copyright 2015 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.
==============================================================================*/
#define EIGEN_USE_THREADS
#include "third_party/eigen3/unsupported/Eigen/CXX11/Tensor"
#include "tensorflow/core/framework/numeric_op.h"
#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/register_types.h"
#include "tensorflow/core/framework/tensor.h"
#include "tensorflow/core/kernels/quantization_utils.h"
namespace tensorflow {
namespace {
// A slow but straightforward implementation of batch normalization.
template <typename T1, typename T2>
void ReferenceBatchNorm(const Tensor& input, const float input_min,
const float input_max, const Tensor& mean,
float mean_min, float mean_max, const Tensor& var,
float var_min, float var_max, const Tensor& beta,
float beta_min, float beta_max, const Tensor& gamma,
float gamma_min, float gamma_max,
float variance_epsilon, bool scale_after_normalization,
Tensor* output, float* output_min, float* output_max) {
auto input_flat = input.flat<T1>();
auto mean_flat = mean.flat<T1>();
auto var_flat = var.flat<T1>();
auto beta_flat = beta.flat<T1>();
auto gamma_flat = gamma.flat<T1>();
auto output_flat = output->flat<T2>();
const int depth = mean.dim_size(0);
const int row_count = input_flat.size() / depth;
*output_min = std::numeric_limits<float>::max();
*output_max = std::numeric_limits<float>::lowest();
for (int pass = 0; pass < 2; ++pass) {
const bool is_range_pass = (pass == 0);
for (int row_index = 0; row_index < row_count; ++row_index) {
for (int channel = 0; channel < depth; ++channel) {
const int input_index = (row_index * depth) + channel;
const float input_value =
QuantizedToFloat(input_flat(input_index), input_min, input_max);
const float mean_value =
QuantizedToFloat(mean_flat(channel), mean_min, mean_max);
const float var_value =
QuantizedToFloat(var_flat(channel), var_min, var_max);
const float beta_value =
QuantizedToFloat(beta_flat(channel), beta_min, beta_max);
const float gamma_value =
QuantizedToFloat(gamma_flat(channel), gamma_min, gamma_max);
float output_value;
if (scale_after_normalization) {
output_value = (((input_value - mean_value) /
sqrtf(var_value + variance_epsilon)) *
gamma_value) +
beta_value;
} else {
output_value = ((input_value - mean_value) /
sqrtf(var_value + variance_epsilon)) +
beta_value;
}
if (is_range_pass) {
*output_min = std::min(output_value, *output_min);
*output_max = std::max(output_value, *output_max);
} else {
output_flat(input_index) =
FloatToQuantized<T2>(output_value, *output_min, *output_max);
}
}
}
}
}
// An implementation of batch normalization that does the main calculations
// using only fixed-point arithmetic. There's a prologue with some floating
// calculations, but assuming the weights are constant these could be hoisted to
// an offline process, or baked into the weights.
template <typename T1, typename T2>
void FixedPointBatchNorm(const Tensor& input, const float input_min,
const float input_max, const Tensor& mean,
float mean_min, float mean_max, const Tensor& var,
float var_min, float var_max, const Tensor& beta,
float beta_min, float beta_max, const Tensor& gamma,
float gamma_min, float gamma_max,
float variance_epsilon, bool scale_after_normalization,
Tensor* output, float* output_min, float* output_max) {
auto input_flat = input.flat<T1>();
auto mean_flat = mean.flat<T1>();
auto var_flat = var.flat<T1>();
auto beta_flat = beta.flat<T1>();
auto gamma_flat = gamma.flat<T1>();
auto output_flat = output->flat<T2>();
const int depth = mean.dim_size(0);
const int row_count = input_flat.size() / depth;
// The range here is chosen so that typical input values fit in without any
// overflow or loss of precision, going from +1m to -1m with 10 bits of fixed
// point precision.
*output_min = -(1 << 20);
*output_max = (1 << 20);
Tensor scale_tensor(DataTypeToEnum<T2>::v(), {depth});
auto scale_flat = scale_tensor.flat<T2>();
Tensor offset_tensor(DataTypeToEnum<T2>::v(), {depth});
auto offset_flat = offset_tensor.flat<T2>();
for (int channel = 0; channel < depth; ++channel) {
const float mean_value =
QuantizedToFloat(mean_flat(channel), mean_min, mean_max);
const float var_value =
QuantizedToFloat(var_flat(channel), var_min, var_max);
const float beta_value =
QuantizedToFloat(beta_flat(channel), beta_min, beta_max);
const float gamma_value =
QuantizedToFloat(gamma_flat(channel), gamma_min, gamma_max);
float scale_value;
if (scale_after_normalization) {
scale_value = (1.0f / sqrtf(var_value + variance_epsilon)) * gamma_value;
} else {
scale_value = (1.0f / sqrtf(var_value + variance_epsilon));
}
const float offset_value = (-mean_value * scale_value) + beta_value;
scale_flat(channel) =
FloatToQuantized<T2>(scale_value, *output_min, *output_max);
offset_flat(channel) =
FloatToQuantized<T2>(offset_value, *output_min, *output_max);
}
const T2 one_in_output_space =
FloatToQuantized<T2>(1.0f, *output_min, *output_max);
for (int row_index = 0; row_index < row_count; ++row_index) {
for (int channel = 0; channel < depth; ++channel) {
const int input_index = (row_index * depth) + channel;
const T2 input_value =
RequantizeInNewRange<T1, T2>(input_flat(input_index), input_min,
input_max, *output_min, *output_max);
const T2 scale_value = scale_flat(channel);
const T2 offset_value = offset_flat(channel);
const T2 output_value =
((input_value * scale_value) / one_in_output_space) + offset_value;
output_flat(input_index) = output_value;
}
}
}
} // namespace
template <typename T1, typename T2>
class QuantizedBatchNormOp : public OpKernel {
public:
explicit QuantizedBatchNormOp(OpKernelConstruction* context)
: OpKernel(context) {
OP_REQUIRES_OK(context,
context->GetAttr("variance_epsilon", &variance_epsilon_));
OP_REQUIRES_OK(context, context->GetAttr("scale_after_normalization",
&scale_after_normalization_));
}
void Compute(OpKernelContext* context) override {
const Tensor& input = context->input(0);
const float input_min = context->input(1).flat<float>()(0);
const float input_max = context->input(2).flat<float>()(0);
const Tensor& mean = context->input(3);
const float mean_min = context->input(4).flat<float>()(0);
const float mean_max = context->input(5).flat<float>()(0);
const Tensor& var = context->input(6);
const float var_min = context->input(7).flat<float>()(0);
const float var_max = context->input(8).flat<float>()(0);
const Tensor& beta = context->input(9);
const float beta_min = context->input(10).flat<float>()(0);
const float beta_max = context->input(11).flat<float>()(0);
const Tensor& gamma = context->input(12);
const float gamma_min = context->input(13).flat<float>()(0);
const float gamma_max = context->input(14).flat<float>()(0);
OP_REQUIRES(context, input.dims() == 4,
errors::InvalidArgument("input must be 4-dimensional",
input.shape().DebugString()));
OP_REQUIRES(context, mean.dims() == 1,
errors::InvalidArgument("mean must be 1-dimensional",
mean.shape().DebugString()));
OP_REQUIRES(context, var.dims() == 1,
errors::InvalidArgument("var must be 1-dimensional",
var.shape().DebugString()));
OP_REQUIRES(context, beta.dims() == 1,
errors::InvalidArgument("beta must be 1-dimensional",
beta.shape().DebugString()));
OP_REQUIRES(context, gamma.dims() == 1,
errors::InvalidArgument("gamma must be 1-dimensional",
gamma.shape().DebugString()));
Tensor* output = nullptr;
OP_REQUIRES_OK(context,
context->allocate_output(0, input.shape(), &output));
float output_min;
float output_max;
FixedPointBatchNorm<T1, T2>(input, input_min, input_max, mean, mean_min,
mean_max, var, var_min, var_max, beta, beta_min,
beta_max, gamma, gamma_min, gamma_max,
variance_epsilon_, scale_after_normalization_,
output, &output_min, &output_max);
Tensor* output_min_tensor = nullptr;
OP_REQUIRES_OK(context,
context->allocate_output(1, {}, &output_min_tensor));
output_min_tensor->flat<float>()(0) = output_min;
Tensor* output_max_tensor = nullptr;
OP_REQUIRES_OK(context,
context->allocate_output(2, {}, &output_max_tensor));
output_max_tensor->flat<float>()(0) = output_max;
}
private:
float variance_epsilon_;
bool scale_after_normalization_;
};
REGISTER_KERNEL_BUILDER(Name("QuantizedBatchNormWithGlobalNormalization")
.Device(DEVICE_CPU)
.TypeConstraint<quint8>("Tinput")
.TypeConstraint<qint32>("out_type"),
QuantizedBatchNormOp<quint8, qint32>);
} // namespace tensorflow
|
