experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Create fixed point Softmax that uses asymmetric quantization with int8 as input and output.

Browse Source 
PiperOrigin-RevId: 226401161
This commit is contained in:
Jian Li 2018-12-20 15:13:09 -08:00 committed by TensorFlower Gardener
parent 7a0c9559a9
commit 23178bd498
7 changed files with 394 additions and 42 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

128
tensorflow/lite/kernels/activations.cc
View File

@ -23,6 +23,7 @@ limitations under the License.
#include "tensorflow/lite/c/c_api_internal.h" #include "tensorflow/lite/c/c_api_internal.h"
#include "tensorflow/lite/kernels/internal/optimized/optimized_ops.h" #include "tensorflow/lite/kernels/internal/optimized/optimized_ops.h"
#include "tensorflow/lite/kernels/internal/quantization_util.h" #include "tensorflow/lite/kernels/internal/quantization_util.h"
#include "tensorflow/lite/kernels/internal/reference/integer_ops/softmax.h"
#include "tensorflow/lite/kernels/internal/reference/reference_ops.h" #include "tensorflow/lite/kernels/internal/reference/reference_ops.h"
#include "tensorflow/lite/kernels/internal/tensor.h" #include "tensorflow/lite/kernels/internal/tensor.h"
#include "tensorflow/lite/kernels/kernel_util.h" #include "tensorflow/lite/kernels/kernel_util.h"
@ -50,6 +51,20 @@ struct PreluOpData : public OpData {
int output_shift = 0; int output_shift = 0;
}; };
namespace {
TfLiteStatus CheckInputQuantParams(TfLiteContext* context,
const TfLiteTensor* input,
const TfLiteTensor* output) {
if (input->type == kTfLiteUInt8) {
TF_LITE_ENSURE_EQ(context, output->params.zero_point, 0);
TF_LITE_ENSURE(context, output->params.scale == 1. / 256);
} else {
TF_LITE_ENSURE_EQ(context, output->params.zero_point, -128);
}
return kTfLiteOk;
}
} // namespace
void* Init(TfLiteContext* context, const char* buffer, size_t length) { void* Init(TfLiteContext* context, const char* buffer, size_t length) {
// This is a builtin op, so we don't use the contents in 'buffer', if any. // This is a builtin op, so we don't use the contents in 'buffer', if any.
// Instead, we allocate a new object to carry information from Prepare() to // Instead, we allocate a new object to carry information from Prepare() to
@ -215,12 +230,12 @@ TfLiteStatus SoftmaxPrepare(TfLiteContext* context, TfLiteNode* node) {
const int num_dims = NumDimensions(input); const int num_dims = NumDimensions(input);
TF_LITE_ENSURE(context, num_dims >= 1 && num_dims <= 4); TF_LITE_ENSURE(context, num_dims >= 1 && num_dims <= 4);
if (input->type == kTfLiteUInt8) { if (input->type == kTfLiteUInt8 || input->type == kTfLiteInt8) {
TF_LITE_ENSURE_EQ(context, output->params.zero_point, 0); if (CheckInputQuantParams(context, input, output) == kTfLiteError) {
TF_LITE_ENSURE(context, output->params.scale == 1. / 256); return kTfLiteError;
}
static const int kScaledDiffIntegerBits = 5; static const int kScaledDiffIntegerBits = 5;
tflite::PreprocessSoftmaxScaling( tflite::PreprocessSoftmaxScaling(
params->beta, input->params.scale, kScaledDiffIntegerBits, params->beta, input->params.scale, kScaledDiffIntegerBits,
&data->input_multiplier, &data->input_left_shift); &data->input_multiplier, &data->input_left_shift);
@ -505,8 +520,8 @@ void Softmax3DFloat(const TfLiteTensor* input, TfLiteTensor* output,
GetTensorData<float>(output)); GetTensorData<float>(output));
} }
void Softmax1DQuantized(const TfLiteTensor* input, TfLiteTensor* output, void Softmax1DQuantizedUint8(const TfLiteTensor* input, TfLiteTensor* output,
TfLiteSoftmaxParams* params, OpData* data) { TfLiteSoftmaxParams* params, OpData* data) {
// TODO(ahentz): this is arguably a dirty trick. Since the implementation // TODO(ahentz): this is arguably a dirty trick. Since the implementation
// always traverses the last dimension of a 4D tensor, we will pretend our 1D // always traverses the last dimension of a 4D tensor, we will pretend our 1D
// tensor is 4D in a special way. We will convert a (Y) shape into a (1, // tensor is 4D in a special way. We will convert a (Y) shape into a (1,
@ -521,8 +536,8 @@ void Softmax1DQuantized(const TfLiteTensor* input, TfLiteTensor* output,
GetTensorShape({1, 1, 1, input_size}), GetTensorShape({1, 1, 1, input_size}),
GetTensorData<uint8_t>(output)); GetTensorData<uint8_t>(output));
} }
void Softmax2DQuantized(const TfLiteTensor* input, TfLiteTensor* output, void Softmax2DQuantizedUint8(const TfLiteTensor* input, TfLiteTensor* output,
TfLiteSoftmaxParams* params, OpData* data) { TfLiteSoftmaxParams* params, OpData* data) {
// TODO(ahentz): this is arguably a dirty trick. Since the implementation // TODO(ahentz): this is arguably a dirty trick. Since the implementation
// always traverses the last dimension of a 4D tensor, we will pretend our 2D // always traverses the last dimension of a 4D tensor, we will pretend our 2D
// tensor is 4D in a special way. We will convert a (X, Y) shape into a (X, // tensor is 4D in a special way. We will convert a (X, Y) shape into a (X,
@ -540,8 +555,8 @@ void Softmax2DQuantized(const TfLiteTensor* input, TfLiteTensor* output,
GetTensorData<uint8_t>(output)); GetTensorData<uint8_t>(output));
} }
void Softmax3DQuantized(const TfLiteTensor* input, TfLiteTensor* output, void Softmax3DQuantizedUint8(const TfLiteTensor* input, TfLiteTensor* output,
TfLiteSoftmaxParams* params, OpData* data) { TfLiteSoftmaxParams* params, OpData* data) {
const int batch_size = input->dims->data[0]; const int batch_size = input->dims->data[0];
const int intermediate_size = input->dims->data[1]; const int intermediate_size = input->dims->data[1];
const int input_size = input->dims->data[2]; const int input_size = input->dims->data[2];
@ -566,8 +581,8 @@ void Softmax4DFloat(const TfLiteTensor* input, TfLiteTensor* output,
GetTensorData<float>(output)); GetTensorData<float>(output));
} }
void Softmax4DQuantized(const TfLiteTensor* input, TfLiteTensor* output, void Softmax4DQuantizedUint8(const TfLiteTensor* input, TfLiteTensor* output,
TfLiteSoftmaxParams* params, OpData* data) { TfLiteSoftmaxParams* params, OpData* data) {
SoftmaxParams op_params; SoftmaxParams op_params;
op_params.input_multiplier = data->input_multiplier; op_params.input_multiplier = data->input_multiplier;
op_params.input_left_shift = data->input_left_shift; op_params.input_left_shift = data->input_left_shift;
@ -577,6 +592,63 @@ void Softmax4DQuantized(const TfLiteTensor* input, TfLiteTensor* output,
GetTensorData<uint8_t>(output)); GetTensorData<uint8_t>(output));
} }
// TODO(jianlijianli): Try merging Softmax<n>DQuantizedInt8 with
// Softmax<n>DQuantized, which needs a larger refactor.
void Softmax1DQuantizedInt8(const TfLiteTensor* input, TfLiteTensor* output,
TfLiteSoftmaxParams* params, OpData* data) {
const int input_size = input->dims->data[0];
SoftmaxParams op_params;
op_params.input_multiplier = data->input_multiplier;
op_params.input_left_shift = data->input_left_shift;
op_params.diff_min = data->diff_min;
reference_integer_ops::Softmax(
op_params, GetTensorShape({1, 1, 1, input_size}),
GetTensorData<int8_t>(input), GetTensorShape({1, 1, 1, input_size}),
GetTensorData<int8_t>(output));
}
void Softmax2DQuantizedInt8(const TfLiteTensor* input, TfLiteTensor* output,
TfLiteSoftmaxParams* params, OpData* data) {
const int batch_size = input->dims->data[0];
const int input_size = input->dims->data[1];
SoftmaxParams op_params;
op_params.input_multiplier = data->input_multiplier;
op_params.input_left_shift = data->input_left_shift;
op_params.diff_min = data->diff_min;
reference_integer_ops::Softmax(op_params,
GetTensorShape({batch_size, 1, 1, input_size}),
GetTensorData<int8_t>(input),
GetTensorShape({batch_size, 1, 1, input_size}),
GetTensorData<int8_t>(output));
}
void Softmax3DQuantizedInt8(const TfLiteTensor* input, TfLiteTensor* output,
TfLiteSoftmaxParams* params, OpData* data) {
const int batch_size = input->dims->data[0];
const int intermediate_size = input->dims->data[1];
const int input_size = input->dims->data[2];
SoftmaxParams op_params;
op_params.input_multiplier = data->input_multiplier;
op_params.input_left_shift = data->input_left_shift;
op_params.diff_min = data->diff_min;
reference_integer_ops::Softmax(
op_params, GetTensorShape({batch_size, intermediate_size, 1, input_size}),
GetTensorData<int8_t>(input),
GetTensorShape({batch_size, intermediate_size, 1, input_size}),
GetTensorData<int8_t>(output));
}
void Softmax4DQuantizedInt8(const TfLiteTensor* input, TfLiteTensor* output,
TfLiteSoftmaxParams* params, OpData* data) {
SoftmaxParams op_params;
op_params.input_multiplier = data->input_multiplier;
op_params.input_left_shift = data->input_left_shift;
op_params.diff_min = data->diff_min;
reference_integer_ops::Softmax(
op_params, GetTensorShape(input), GetTensorData<int8_t>(input),
GetTensorShape(output), GetTensorData<int8_t>(output));
}
TfLiteStatus SoftmaxEval(TfLiteContext* context, TfLiteNode* node) { TfLiteStatus SoftmaxEval(TfLiteContext* context, TfLiteNode* node) {
auto* params = reinterpret_cast<TfLiteSoftmaxParams*>(node->builtin_data); auto* params = reinterpret_cast<TfLiteSoftmaxParams*>(node->builtin_data);
OpData* data = reinterpret_cast<OpData*>(node->user_data); OpData* data = reinterpret_cast<OpData*>(node->user_data);
@ -611,19 +683,19 @@ TfLiteStatus SoftmaxEval(TfLiteContext* context, TfLiteNode* node) {
} }
case kTfLiteUInt8: { case kTfLiteUInt8: {
if (NumDimensions(input) == 1) { if (NumDimensions(input) == 1) {
Softmax1DQuantized(input, output, params, data); Softmax1DQuantizedUint8(input, output, params, data);
return kTfLiteOk; return kTfLiteOk;
} }
if (NumDimensions(input) == 2) { if (NumDimensions(input) == 2) {
Softmax2DQuantized(input, output, params, data); Softmax2DQuantizedUint8(input, output, params, data);
return kTfLiteOk; return kTfLiteOk;
} }
if (NumDimensions(input) == 3) { if (NumDimensions(input) == 3) {
Softmax3DQuantized(input, output, params, data); Softmax3DQuantizedUint8(input, output, params, data);
return kTfLiteOk; return kTfLiteOk;
} }
if (NumDimensions(input) == 4) { if (NumDimensions(input) == 4) {
Softmax4DQuantized(input, output, params, data); Softmax4DQuantizedUint8(input, output, params, data);
return kTfLiteOk; return kTfLiteOk;
} }
context->ReportError( context->ReportError(
@ -631,6 +703,30 @@ TfLiteStatus SoftmaxEval(TfLiteContext* context, TfLiteNode* node) {
NumDimensions(input)); NumDimensions(input));
return kTfLiteError; return kTfLiteError;
} }
case kTfLiteInt8: {
if (NumDimensions(input) == 1) {
Softmax1DQuantizedInt8(input, output, params, data);
return kTfLiteOk;
}
if (NumDimensions(input) == 2) {
Softmax2DQuantizedInt8(input, output, params, data);
return kTfLiteOk;
}
if (NumDimensions(input) == 3) {
Softmax3DQuantizedInt8(input, output, params, data);
return kTfLiteOk;
}
if (NumDimensions(input) == 4) {
Softmax4DQuantizedInt8(input, output, params, data);
return kTfLiteOk;
}
context->ReportError(
context,
"Only 4D tensors supported currently for Int8 kernels, got %dD.",
NumDimensions(input));
return kTfLiteError;
}
default: default:
context->ReportError( context->ReportError(
context, "Only float32 and uint8_t supported currently, got %s.", context, "Only float32 and uint8_t supported currently, got %s.",

153
tensorflow/lite/kernels/activations_test.cc
View File

@ -44,6 +44,8 @@ class BaseActivationsOpModel : public SingleOpModel {
input_ = AddInput(input); input_ = AddInput(input);
if (input.type == TensorType_UINT8) { if (input.type == TensorType_UINT8) {
output_ = AddOutput({input.type, {}, 0, 0, 1. / 256}); output_ = AddOutput({input.type, {}, 0, 0, 1. / 256});
} else if (input.type == TensorType_INT8) {
output_ = AddOutput({TensorType_INT8, {}, 0, 0, 1. / 256, -128});
} else { } else {
output_ = AddOutput({input.type, {}}); output_ = AddOutput({input.type, {}});
} }
@ -52,8 +54,8 @@ class BaseActivationsOpModel : public SingleOpModel {
BuildInterpreter({GetShape(input_)}); BuildInterpreter({GetShape(input_)});
} }
BaseActivationsOpModel(BuiltinOperator type, const TensorData &input, BaseActivationsOpModel(BuiltinOperator type, const TensorData& input,
const TensorData &output) { const TensorData& output) {
input_ = AddInput(input); input_ = AddInput(input);
output_ = AddOutput(output); output_ = AddOutput(output);
SetBuiltinOp(type, BuiltinOptions_NONE, 0); SetBuiltinOp(type, BuiltinOptions_NONE, 0);
@ -323,7 +325,7 @@ TEST(FloatActivationsOpTest, Softmax4D) {
}))); })));
} }
TEST(QuantizedActivationsOpTest, Softmax4D) { TEST(QuantizedActivationsOpTest, Softmax4DUint8) {
QuantizedActivationsOpModel m( QuantizedActivationsOpModel m(
0.1, 0.1,
/*input=*/{TensorType_UINT8, {1, 2, 1, 4}, -10, 10}); /*input=*/{TensorType_UINT8, {1, 2, 1, 4}, -10, 10});
@ -362,6 +364,145 @@ TEST(QuantizedActivationsOpTest, Softmax4D) {
kQuantizedTolerance))); kQuantizedTolerance)));
} }
// Test quantized softmax with int8 input and output. With the same input as in
// QuantizedActivationsOpTest.Softmax1D, the dequantized output is identical.
TEST(QuantizedActivationsOpTest, Softmax1DInt8) {
QuantizedActivationsOpModel m(0.1,
/*input=*/{TensorType_INT8, {8}, -10, 10});
m.SetInput<int8_t>({0, -6, 2, 4, 3, -2, 10, 1});
m.Invoke();
EXPECT_THAT(
m.GetDequantizedOutput<int8_t>(),
ElementsAreArray(ArrayFloatNear({0.09766, 0.05469, 0.12109, 0.14453,
0.13281, 0.07813, 0.26563, 0.10938},
kQuantizedTolerance)));
}
// Test quantized softmax with int8 input and output. With the same input as in
// QuantizedActivationsOpTest.Softmax2D, the dequantized output is identical.
TEST(QuantizedActivationsOpTest, Softmax2DInt8) {
QuantizedActivationsOpModel m(0.1,
/*input=*/{TensorType_INT8, {2, 4}, -10, 10});
m.SetInput<int8_t>({
0, -6, 2, 4, //
3, -2, 10, 1, //
});
m.Invoke();
EXPECT_THAT(m.GetDequantizedOutput<int8_t>(),
ElementsAreArray(ArrayFloatNear(
{
.23463, .12877, .28658, .35003, //
.22528, .13664, .45365, .18443, //
},
kQuantizedTolerance)));
// Same input, but a different shape.
QuantizedActivationsOpModel m2(0.1,
/*input=*/{TensorType_INT8, {4, 2}, -10, 10});
m2.SetInput<int8_t>({
0, -6, //
2, 4, //
3, -2, //
10, 1, //
});
m2.Invoke();
EXPECT_THAT(m2.GetDequantizedOutput<int8_t>(),
ElementsAreArray(ArrayFloatNear(
{
0.645656, 0.354344, //
0.450166, 0.549834, //
0.622459, 0.377541, //
0.710949, 0.28905, //
},
kQuantizedTolerance)));
}
// Test quantized softmax with int8 input and output. With the same input as in
// QuantizedActivationsOpTest.Softmax3D, the dequantized output is identical.
TEST(QuantizedActivationsOpTest, Softmax3DInt8) {
QuantizedActivationsOpModel m(
0.1,
/*input=*/{TensorType_INT8, {1, 2, 4}, -10, 10});
m.SetInput<int8_t>({
0, -6, 2, 4, // depth = 0
3, -2, 10, 1, // depth = 1
});
m.Invoke();
EXPECT_THAT(m.GetDequantizedOutput<int8_t>(),
ElementsAreArray(ArrayFloatNear(
{
.23463, .12877, .28658, .35003, //
.22528, .13664, .45365, .18443, //
},
kQuantizedTolerance)));
// Same input, but a different shape.
QuantizedActivationsOpModel m2(
0.1,
/*input=*/{TensorType_INT8, {4, 1, 2}, -10, 10});
m2.SetInput<int8_t>({
0, -6, //
2, 4, //
3, -2, //
10, 1, //
});
m2.Invoke();
EXPECT_THAT(m2.GetDequantizedOutput<int8_t>(),
ElementsAreArray(ArrayFloatNear(
{
0.645656, 0.354344, //
0.450166, 0.549834, //
0.622459, 0.377541, //
0.710949, 0.28905, //
},
kQuantizedTolerance)));
}
// Test quantized softmax with int8 input and output. With the same input as in
// QuantizedActivationsOpTest.Softmax4D, the dequantized output is identical.
TEST(QuantizedActivationsOpTest, Softmax4DInt8) {
QuantizedActivationsOpModel m(
0.1,
/*input=*/{TensorType_INT8, {1, 2, 1, 4}, -10, 10});
m.SetInput<int8_t>({
0, -6, 2, 4, // depth = 0
3, -2, 10, 1, // depth = 1
});
m.Invoke();
EXPECT_THAT(m.GetOutput<int8_t>(), ElementsAreArray({
-68, -95, -54, -38, //
-70, -93, -12, -81, //
}));
EXPECT_THAT(m.GetDequantizedOutput<int8_t>(),
ElementsAreArray(ArrayFloatNear(
{
.23463, .12877, .28658, .35003, //
.22528, .13664, .45365, .18443, //
},
kQuantizedTolerance)));
// Same input, but a different shape.
QuantizedActivationsOpModel m2(
0.1,
/*input=*/{TensorType_INT8, {4, 1, 1, 2}, -10, 10});
m2.SetInput<int8_t>({
0, -6, //
2, 4, //
3, -2, //
10, 1, //
});
m2.Invoke();
EXPECT_THAT(m2.GetDequantizedOutput<int8_t>(),
ElementsAreArray(ArrayFloatNear(
{
0.645656, 0.354344, //
0.450166, 0.549834, //
0.622459, 0.377541, //
0.710949, 0.28905, //
},
kQuantizedTolerance)));
}
TEST(FloatActivationsOpTest, Softmax3D) { TEST(FloatActivationsOpTest, Softmax3D) {
FloatActivationsOpModel m(0.1, FloatActivationsOpModel m(0.1,
/*input=*/{TensorType_FLOAT32, {1, 2, 4}}); /*input=*/{TensorType_FLOAT32, {1, 2, 4}});
@ -393,7 +534,7 @@ TEST(FloatActivationsOpTest, Softmax3D) {
}))); })));
} }
TEST(QuantizedActivationsOpTest, Softmax3D) { TEST(QuantizedActivationsOpTest, Softmax3DUint8) {
QuantizedActivationsOpModel m( QuantizedActivationsOpModel m(
0.1, 0.1,
/*input=*/{TensorType_UINT8, {1, 2, 4}, -10, 10}); /*input=*/{TensorType_UINT8, {1, 2, 4}, -10, 10});
@ -443,7 +584,7 @@ TEST(FloatActivationsOpTest, Softmax1D) {
{.09752, .05352, .11911, .14548, .13164, .07984, .26509, .10778}))); {.09752, .05352, .11911, .14548, .13164, .07984, .26509, .10778})));
} }
TEST(QuantizedActivationsOpTest, Softmax1D) { TEST(QuantizedActivationsOpTest, Softmax1DUint8) {
QuantizedActivationsOpModel m(0.1, QuantizedActivationsOpModel m(0.1,
/*input=*/{TensorType_UINT8, {8}, -10, 10}); /*input=*/{TensorType_UINT8, {8}, -10, 10});
m.SetInput<uint8_t>({0, -6, 2, 4, 3, -2, 10, 1}); m.SetInput<uint8_t>({0, -6, 2, 4, 3, -2, 10, 1});
@ -486,7 +627,7 @@ TEST(FloatActivationsOpTest, Softmax2D) {
}))); })));
} }
TEST(QuantizedActivationsOpTest, Softmax2D) { TEST(QuantizedActivationsOpTest, Softmax2DUint8) {
QuantizedActivationsOpModel m(0.1, QuantizedActivationsOpModel m(0.1,
/*input=*/{TensorType_UINT8, {2, 4}, -10, 10}); /*input=*/{TensorType_UINT8, {2, 4}, -10, 10});
m.SetInput<uint8_t>({ m.SetInput<uint8_t>({

7
tensorflow/lite/kernels/internal/BUILD
View File

@ -1,12 +1,12 @@
load("//tensorflow/lite:build_def.bzl", "tflite_copts")
load("//tensorflow/lite:special_rules.bzl", "tflite_portable_test_suite")
package(default_visibility = [ package(default_visibility = [
"//visibility:public", "//visibility:public",
]) ])
licenses(["notice"]) # Apache 2.0 licenses(["notice"]) # Apache 2.0
load("//tensorflow/lite:build_def.bzl", "tflite_copts")
load("//tensorflow/lite:special_rules.bzl", "tflite_portable_test_suite")
tflite_deps_intel = [ tflite_deps_intel = [
"@arm_neon_2_x86_sse", "@arm_neon_2_x86_sse",
] ]
@ -314,6 +314,7 @@ cc_library(
"reference/depthwiseconv_uint8.h", "reference/depthwiseconv_uint8.h",
"reference/fully_connected.h", "reference/fully_connected.h",
"reference/integer_ops/dequantize.h", "reference/integer_ops/dequantize.h",
"reference/integer_ops/softmax.h",
"reference/reference_ops.h", "reference/reference_ops.h",
"reference/softmax.h", "reference/softmax.h",
], ],

17
tensorflow/lite/kernels/internal/common.h
View File

@ -131,6 +131,23 @@ int CountLeadingZeros(T integer_input) {
#endif #endif
} }
inline int32 GetReciprocal(int32 x, int x_integer_digits,
int* num_bits_over_unit) {
int headroom_plus_one = CountLeadingZeros(static_cast<uint32>(x));
// This is the number of bits to the left of the binary point above 1.0.
// Consider x=1.25. In that case shifted_scale=0.8 and
// no later adjustment will be needed.
*num_bits_over_unit = x_integer_digits - headroom_plus_one;
const int32 shifted_sum_minus_one =
static_cast<int32>((static_cast<uint32>(x) << headroom_plus_one) -
(static_cast<uint32>(1) << 31));
gemmlowp::FixedPoint<int32, 0> shifted_scale =
gemmlowp::one_over_one_plus_x_for_x_in_0_1(
gemmlowp::FixedPoint<int32, 0>::FromRaw(shifted_sum_minus_one));
return shifted_scale.raw();
}
// DO NOT USE THIS STRUCT FOR NEW FUNCTIONALITY BEYOND IMPLEMENTING // DO NOT USE THIS STRUCT FOR NEW FUNCTIONALITY BEYOND IMPLEMENTING
// BROADCASTING. // BROADCASTING.
// //

102
tensorflow/lite/kernels/internal/reference/integer_ops/softmax.h Normal file
View File

@ -0,0 +1,102 @@
/* Copyright 2018 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.
==============================================================================*/
#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_SOFTMAX_H_
#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_SOFTMAX_H_
#include "tensorflow/lite/kernels/internal/common.h"
namespace tflite {
namespace reference_integer_ops {
// Quantized softmax with int8 input and output.
inline void Softmax(const SoftmaxParams& params,
const RuntimeShape& input_shape, const int8* input_data,
const RuntimeShape& output_shape, int8* output_data) {
const int32 input_beta_multiplier = params.input_multiplier;
const int32 input_beta_left_shift = params.input_left_shift;
const int diff_min = params.diff_min;
// The representation chosen for the input to the exp() function is Q5.26.
// We need to leave extra space since values that we skip might be as large as
// -32 before multiplying by input_beta_multiplier, and therefore as large as
// -16 afterwards. Note that exp(-8) is definitely not insignificant to
// accumulation, but exp(-16) definitely is.
static const int kScaledDiffIntegerBits = 5;
static const int kAccumulationIntegerBits = 12;
using FixedPointScaledDiff =
gemmlowp::FixedPoint<int32, kScaledDiffIntegerBits>;
using FixedPointAccum = gemmlowp::FixedPoint<int32, kAccumulationIntegerBits>;
using FixedPoint0 = gemmlowp::FixedPoint<int32, 0>;
const int trailing_dim = input_shape.DimensionsCount() - 1;
const int outer_size =
MatchingFlatSizeSkipDim(input_shape, trailing_dim, output_shape);
const int depth =
MatchingDim(input_shape, trailing_dim, output_shape, trailing_dim);
for (int i = 0; i < outer_size; ++i) {
int8 max_in_row = -128;
for (int c = 0; c < depth; ++c) {
max_in_row = std::max(max_in_row, input_data[i * depth + c]);
}
FixedPointAccum sum_of_exps = FixedPointAccum::Zero();
for (int c = 0; c < depth; ++c) {
int32 input_diff =
static_cast<int32>(input_data[i * depth + c]) - max_in_row;
if (input_diff >= diff_min) {
const int32 input_diff_rescaled =
MultiplyByQuantizedMultiplierGreaterThanOne(
input_diff, input_beta_multiplier, input_beta_left_shift);
const FixedPointScaledDiff scaled_diff_f8 =
FixedPointScaledDiff::FromRaw(input_diff_rescaled);
sum_of_exps = sum_of_exps + gemmlowp::Rescale<kAccumulationIntegerBits>(
exp_on_negative_values(scaled_diff_f8));
}
}
int num_bits_over_unit;
FixedPoint0 shifted_scale = FixedPoint0::FromRaw(GetReciprocal(
sum_of_exps.raw(), kAccumulationIntegerBits, &num_bits_over_unit));
for (int c = 0; c < depth; ++c) {
int32 input_diff =
static_cast<int32>(input_data[i * depth + c]) - max_in_row;
if (input_diff >= diff_min) {
const int32 input_diff_rescaled =
MultiplyByQuantizedMultiplierGreaterThanOne(
input_diff, input_beta_multiplier, input_beta_left_shift);
const FixedPointScaledDiff scaled_diff_f8 =
FixedPointScaledDiff::FromRaw(input_diff_rescaled);
FixedPoint0 exp_in_0 = exp_on_negative_values(scaled_diff_f8);
const int32 unsat_output = gemmlowp::RoundingDivideByPOT(
(shifted_scale * exp_in_0).raw(), num_bits_over_unit + 31 - 8);
const int32 shifted_output = unsat_output - 128;
output_data[i * depth + c] = static_cast<int8>(
std::max(std::min(shifted_output, static_cast<int32>(127)),
static_cast<int32>(-128)));
} else {
output_data[i * depth + c] = -128;
}
}
}
}
} // namespace reference_integer_ops
} // namespace tflite
#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_SOFTMAX_H_

16
tensorflow/lite/kernels/internal/reference/softmax.h
View File

@ -102,19 +102,9 @@ inline void Softmax(const SoftmaxParams& params,
} }
} }
int32 fixed_sum_of_exps = sum_of_exps.raw(); int num_bits_over_unit;
int headroom_plus_one = FixedPoint0 shifted_scale = FixedPoint0::FromRaw(GetReciprocal(
CountLeadingZeros(static_cast<uint32>(fixed_sum_of_exps)); sum_of_exps.raw(), kAccumulationIntegerBits, &num_bits_over_unit));
// This is the number of bits to the left of the binary point above 1.0.
// Consider fixed_sum_of_exps=1.25. In that case shifted_scale=0.8 and
// no later adjustment will be needed.
int num_bits_over_unit = kAccumulationIntegerBits - headroom_plus_one;
int32 shifted_sum_minus_one = static_cast<int32>(
(static_cast<uint32>(fixed_sum_of_exps) << headroom_plus_one) -
(static_cast<uint32>(1) << 31));
FixedPoint0 shifted_scale = gemmlowp::one_over_one_plus_x_for_x_in_0_1(
FixedPoint0::FromRaw(shifted_sum_minus_one));
for (int c = 0; c < depth; ++c) { for (int c = 0; c < depth; ++c) {
int32 input_diff = int32 input_diff =

13
tensorflow/lite/toco/tflite/operator.cc
View File

@ -706,6 +706,11 @@ class Softmax
} }
int GetVersion(const OperatorSignature& op_signature) const override { int GetVersion(const OperatorSignature& op_signature) const override {
const string& input_name = op_signature.op->inputs[0];
const Array& input_array = op_signature.model->GetArray(input_name);
if (input_array.data_type == ArrayDataType::kInt8) {
return 2;
}
return 1; return 1;
} }
}; };
@ -1556,10 +1561,6 @@ class TensorFlowUnsupported : public BaseOperator {
return std::unique_ptr<flexbuffers::Builder>(fbb.release()); return std::unique_ptr<flexbuffers::Builder>(fbb.release());
} }
// TODO(wvo): hack to make this code compile with 2 different API versions.
// Please remove once OS/internal versions are in sync.
// See hardcoded values in the switch below.
void ReadOptions(const flexbuffers::Map& m, void ReadOptions(const flexbuffers::Map& m,
TensorFlowUnsupportedOperator* op) const { TensorFlowUnsupportedOperator* op) const {
::tensorflow::NodeDef node_def; ::tensorflow::NodeDef node_def;
@ -1569,6 +1570,10 @@ class TensorFlowUnsupported : public BaseOperator {
for (size_t i = 0; i < keys.size(); ++i) { for (size_t i = 0; i < keys.size(); ++i) {
const auto key = keys[i].AsKey(); const auto key = keys[i].AsKey();
const auto& value = m[key]; const auto& value = m[key];
// TODO(wvo): hack to make this code compile with 2 different API
// versions.
// Please remove once OS/internal versions are in sync.
// See hardcoded values in the switch below.
switch (value.GetType()) { switch (value.GetType()) {
case 5: // flexbuffers::FBT_STRING: case 5: // flexbuffers::FBT_STRING:
(*attr)[key].set_s(value.AsString().c_str()); (*attr)[key].set_s(value.AsString().c_str());