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
|
/* 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 "tensorflow/compiler/tf2xla/xla_helpers.h"
#include "tensorflow/compiler/tf2xla/xla_op_kernel.h"
#include "tensorflow/compiler/tf2xla/xla_op_registry.h"
#include "tensorflow/compiler/xla/client/xla_builder.h"
#include "tensorflow/core/framework/kernel_def_builder.h"
namespace tensorflow {
namespace {
// Local response normalization
class LRNOp : public XlaOpKernel {
public:
explicit LRNOp(OpKernelConstruction* ctx) : XlaOpKernel(ctx) {
OP_REQUIRES_OK(ctx, ctx->GetAttr("depth_radius", &depth_radius_));
// TODO(phawkins): handle non-float types for attributes.
OP_REQUIRES_OK(ctx, ctx->GetAttr("bias", &bias_));
OP_REQUIRES_OK(ctx, ctx->GetAttr("alpha", &alpha_));
OP_REQUIRES_OK(ctx, ctx->GetAttr("beta", &beta_));
}
void Compile(XlaOpKernelContext* ctx) override {
const TensorShape in_shape = ctx->InputShape(0);
OP_REQUIRES(ctx, in_shape.dims() == 4,
errors::InvalidArgument("in must be 4-dimensional"));
xla::XlaBuilder* builder = ctx->builder();
xla::XlaOp input = ctx->Input(0);
// sqr_sum[a, b, c, d] =
// sum(input[a, b, c, d - depth_radius : d + depth_radius + 1] ** 2)
// output = input / (bias + alpha * sqr_sum) ** beta
// We use a window of depth_radius_ * 2 + 1, to account for the current
// element and a depth_radius_ on either side.
auto accumulation_type = XlaHelpers::SumAccumulationType(input_type(0));
auto converted =
XlaHelpers::ConvertElementType(builder, input, accumulation_type);
auto squared = xla::Mul(converted, converted);
auto reduce = xla::ReduceWindow(
squared, XlaHelpers::Zero(builder, accumulation_type),
*ctx->GetOrCreateAdd(accumulation_type),
/* window_dimensions = */ {1, 1, 1, depth_radius_ * 2 + 1},
/* window_strides = */ {1, 1, 1, 1}, xla::Padding::kSame);
auto sqr_sum =
XlaHelpers::ConvertElementType(builder, reduce, input_type(0));
auto scale = xla::Pow(
xla::Add(xla::ConstantR0<float>(builder, bias_),
xla::Mul(xla::ConstantR0<float>(builder, alpha_), sqr_sum)),
xla::ConstantR0<float>(builder, -beta_));
ctx->SetOutput(0, xla::Mul(input, scale));
}
private:
int64 depth_radius_;
float bias_;
float alpha_;
float beta_;
};
REGISTER_XLA_OP(Name("LRN"), LRNOp);
class LRNGradOp : public XlaOpKernel {
public:
explicit LRNGradOp(OpKernelConstruction* ctx) : XlaOpKernel(ctx) {
OP_REQUIRES_OK(ctx, ctx->GetAttr("depth_radius", &depth_radius_));
// TODO(phawkins): handle non-float types for attributes.
OP_REQUIRES_OK(ctx, ctx->GetAttr("bias", &bias_));
OP_REQUIRES_OK(ctx, ctx->GetAttr("alpha", &alpha_));
OP_REQUIRES_OK(ctx, ctx->GetAttr("beta", &beta_));
}
void Compile(XlaOpKernelContext* ctx) override {
const TensorShape in_grads_shape = ctx->InputShape(0);
const TensorShape in_image_shape = ctx->InputShape(1);
const TensorShape out_image_shape = ctx->InputShape(2);
OP_REQUIRES(ctx, in_grads_shape.dims() == 4 && in_image_shape.dims() == 4,
errors::InvalidArgument("inputs must be 4-dimensional"));
const int64 batch = in_grads_shape.dim_size(0);
const int64 rows = in_grads_shape.dim_size(1);
const int64 cols = in_grads_shape.dim_size(2);
const int64 depth = in_grads_shape.dim_size(3);
OP_REQUIRES(
ctx, in_image_shape.dim_size(0) == batch &&
in_image_shape.dim_size(1) == rows &&
in_image_shape.dim_size(2) == cols &&
in_image_shape.dim_size(3) == depth &&
out_image_shape.dim_size(0) == batch &&
out_image_shape.dim_size(1) == rows &&
out_image_shape.dim_size(2) == cols &&
out_image_shape.dim_size(3) == depth,
errors::InvalidArgument(
"input_grads, input_image, and out_image should have the same "
"shape"));
xla::XlaBuilder* builder = ctx->builder();
xla::XlaOp in_grads = ctx->Input(0);
xla::XlaOp in_image = ctx->Input(1);
xla::XlaOp out_image = ctx->Input(2);
// This code is ported from tensorflow/core/kernels/lrn_op.cc. In Python
// pseudo-code, the Eigen code does this for each spatial position:
// grads = [0.0] * depth
// for j in range(depth):
// depth_begin = max(0, j - depth_radius)
// depth_end = min(depth, j + depth_radius + 1)
//
// norm = 0
// for k in range(depth_begin, depth_end):
// norm += in_image[k] * in_image[k]
// norm = alpha * norm + bias
//
// for k in range(depth_begin, depth_end):
// dyi = -2.0 * alpha * beta * in_image[k] * out_image[j] / norm
// if k == j:
// dyi += norm ** (-beta)
// dyi *= out_grads[j]
// grads[k] += dyi
auto accumulation_type = XlaHelpers::SumAccumulationType(input_type(0));
auto converted =
XlaHelpers::ConvertElementType(builder, in_image, accumulation_type);
auto squared = xla::Mul(converted, converted);
auto reduce = xla::ReduceWindow(
squared, XlaHelpers::Zero(builder, accumulation_type),
*ctx->GetOrCreateAdd(accumulation_type),
/* window_dimensions = */ {1, 1, 1, depth_radius_ * 2 + 1},
/* window_strides = */ {1, 1, 1, 1}, xla::Padding::kSame);
auto sqr_sum =
XlaHelpers::ConvertElementType(builder, reduce, input_type(0));
auto norm =
xla::Add(xla::ConstantR0<float>(builder, bias_),
xla::Mul(xla::ConstantR0<float>(builder, alpha_), sqr_sum));
auto dy = xla::Mul(
xla::Mul(xla::ConstantR0<float>(builder, -2.0f * alpha_ * beta_),
xla::Div(out_image, norm)),
in_grads);
auto converted_dy =
XlaHelpers::ConvertElementType(builder, dy, accumulation_type);
auto dy_reduce = xla::ReduceWindow(
converted_dy, XlaHelpers::Zero(builder, accumulation_type),
*ctx->GetOrCreateAdd(accumulation_type),
/* window_dimensions = */ {1, 1, 1, depth_radius_ * 2 + 1},
/* window_strides = */ {1, 1, 1, 1}, xla::Padding::kSame);
auto dy_reduced =
XlaHelpers::ConvertElementType(builder, dy_reduce, input_type(0));
xla::XlaOp gradients = xla::Add(
xla::Mul(in_image, dy_reduced),
xla::Mul(in_grads,
xla::Pow(norm, xla::ConstantR0<float>(builder, -beta_))));
ctx->SetOutput(0, gradients);
}
private:
int64 depth_radius_;
float bias_;
float alpha_;
float beta_;
};
REGISTER_XLA_OP(Name("LRNGrad"), LRNGradOp);
} // anonymous namespace
} // namespace tensorflow
|
