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

16-bit reference kernel operators MAX_POOL_2D and AVERAGE_POOL_2D

Browse Source
This commit is contained in:
Elena Zhelezina 2020-01-22 14:01:14 +00:00
parent 1fb8fe3793
commit 756cc8f62d
3 changed files with 239 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

115
tensorflow/lite/kernels/internal/reference/integer_ops/pooling.h
View File

@ -135,6 +135,121 @@ inline void MaxPool(const PoolParams& params, const RuntimeShape& input_shape,
}
}
inline void AveragePool(const PoolParams& params,
const RuntimeShape& input_shape,
const int16* input_data,
const RuntimeShape& output_shape, int16* output_data) {
TFLITE_DCHECK_LE(params.quantized_activation_min,
params.quantized_activation_max);
TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
const int batches = MatchingDim(input_shape, 0, output_shape, 0);
const int depth = MatchingDim(input_shape, 3, output_shape, 3);
const int input_height = input_shape.Dims(1);
const int input_width = input_shape.Dims(2);
const int output_height = output_shape.Dims(1);
const int output_width = output_shape.Dims(2);
const int stride_height = params.stride_height;
const int stride_width = params.stride_width;
for (int batch = 0; batch < batches; ++batch) {
for (int out_y = 0; out_y < output_height; ++out_y) {
for (int out_x = 0; out_x < output_width; ++out_x) {
for (int channel = 0; channel < depth; ++channel) {
const int in_x_origin =
(out_x * stride_width) - params.padding_values.width;
const int in_y_origin =
(out_y * stride_height) - params.padding_values.height;
// Compute the boundaries of the filter region clamped so as to
// ensure that the filter window fits in the input array.
const int filter_x_start = std::max(0, -in_x_origin);
const int filter_x_end =
std::min(params.filter_width, input_width - in_x_origin);
const int filter_y_start = std::max(0, -in_y_origin);
const int filter_y_end =
std::min(params.filter_height, input_height - in_y_origin);
int32 acc = 0;
int filter_count = 0;
for (int filter_y = filter_y_start; filter_y < filter_y_end;
++filter_y) {
for (int filter_x = filter_x_start; filter_x < filter_x_end;
++filter_x) {
const int in_x = in_x_origin + filter_x;
const int in_y = in_y_origin + filter_y;
acc +=
input_data[Offset(input_shape, batch, in_y, in_x, channel)];
filter_count++;
}
}
// Round to the closest integer value.
acc = acc > 0 ? (acc + filter_count / 2) / filter_count
: (acc - filter_count / 2) / filter_count;
acc = std::max(acc, params.quantized_activation_min);
acc = std::min(acc, params.quantized_activation_max);
output_data[Offset(output_shape, batch, out_y, out_x, channel)] =
static_cast<int16>(acc);
}
}
}
}
}
inline void MaxPool(const PoolParams& params, const RuntimeShape& input_shape,
const int16* input_data, const RuntimeShape& output_shape,
int16* output_data) {
TFLITE_DCHECK_LE(params.quantized_activation_min,
params.quantized_activation_max);
TFLITE_DCHECK_GE(params.quantized_activation_min,
std::numeric_limits<int16_t>::min());
TFLITE_DCHECK_LE(params.quantized_activation_max,
std::numeric_limits<int16_t>::max());
TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
const int batches = MatchingDim(input_shape, 0, output_shape, 0);
const int depth = MatchingDim(input_shape, 3, output_shape, 3);
const int input_height = input_shape.Dims(1);
const int input_width = input_shape.Dims(2);
const int output_height = output_shape.Dims(1);
const int output_width = output_shape.Dims(2);
const int stride_height = params.stride_height;
const int stride_width = params.stride_width;
for (int batch = 0; batch < batches; ++batch) {
for (int out_y = 0; out_y < output_height; ++out_y) {
for (int out_x = 0; out_x < output_width; ++out_x) {
for (int channel = 0; channel < depth; ++channel) {
const int in_x_origin =
(out_x * stride_width) - params.padding_values.width;
const int in_y_origin =
(out_y * stride_height) - params.padding_values.height;
// Compute the boundaries of the filter region clamped so as to
// ensure that the filter window fits in the input array.
const int filter_x_start = std::max(0, -in_x_origin);
const int filter_x_end =
std::min(params.filter_width, input_width - in_x_origin);
const int filter_y_start = std::max(0, -in_y_origin);
const int filter_y_end =
std::min(params.filter_height, input_height - in_y_origin);
int16_t max = std::numeric_limits<int16_t>::lowest();
for (int filter_y = filter_y_start; filter_y < filter_y_end;
++filter_y) {
for (int filter_x = filter_x_start; filter_x < filter_x_end;
++filter_x) {
const int in_x = in_x_origin + filter_x;
const int in_y = in_y_origin + filter_y;
max = std::max(
max,
input_data[Offset(input_shape, batch, in_y, in_x, channel)]);
}
}
max = std::max<int16_t>(max, params.quantized_activation_min);
max = std::min<int16_t>(max, params.quantized_activation_max);
output_data[Offset(output_shape, batch, out_y, out_x, channel)] =
static_cast<int16_t>(max);
}
}
}
}
}
} // namespace reference_integer_ops
} // namespace tflite

59
tensorflow/lite/kernels/pooling.cc
View File

@ -197,6 +197,32 @@ void AverageEvalQuantizedInt8(TfLiteContext* context, TfLiteNode* node,
#undef TF_LITE_AVERAGE_POOL
}
template <KernelType kernel_type>
void AverageEvalQuantizedInt16(TfLiteContext* context, TfLiteNode* node,
TfLitePoolParams* params, OpData* data,
const TfLiteTensor* input,
TfLiteTensor* output) {
int32_t activation_min;
int32_t activation_max;
CalculateActivationRangeQuantized(context, params->activation, output,
&activation_min, &activation_max);
#define TF_LITE_AVERAGE_POOL(type) \
tflite::PoolParams op_params; \
op_params.stride_height = params->stride_height; \
op_params.stride_width = params->stride_width; \
op_params.filter_height = params->filter_height; \
op_params.filter_width = params->filter_width; \
op_params.padding_values.height = data->padding.height; \
op_params.padding_values.width = data->padding.width; \
op_params.quantized_activation_min = activation_min; \
op_params.quantized_activation_max = activation_max; \
type::AveragePool(op_params, GetTensorShape(input), \
GetTensorData<int16_t>(input), GetTensorShape(output), \
GetTensorData<int16_t>(output))
TF_LITE_AVERAGE_POOL(reference_integer_ops);
#undef TF_LITE_AVERAGE_POOL
}
template <KernelType kernel_type>
void MaxEvalFloat(TfLiteContext* context, TfLiteNode* node,
TfLitePoolParams* params, OpData* data,
@ -282,6 +308,31 @@ void MaxEvalQuantizedInt8(TfLiteContext* context, TfLiteNode* node,
#undef TF_LITE_MAX_POOL
}
template <KernelType kernel_type>
void MaxEvalQuantizedInt16(TfLiteContext* context, TfLiteNode* node,
TfLitePoolParams* params, OpData* data,
const TfLiteTensor* input, TfLiteTensor* output) {
int32_t activation_min;
int32_t activation_max;
CalculateActivationRangeQuantized(context, params->activation, output,
&activation_min, &activation_max);
#define TF_LITE_MAX_POOL(type) \
tflite::PoolParams op_params; \
op_params.stride_height = params->stride_height; \
op_params.stride_width = params->stride_width; \
op_params.filter_height = params->filter_height; \
op_params.filter_width = params->filter_width; \
op_params.padding_values.height = data->padding.height; \
op_params.padding_values.width = data->padding.width; \
op_params.quantized_activation_min = activation_min; \
op_params.quantized_activation_max = activation_max; \
type::MaxPool(op_params, GetTensorShape(input), \
GetTensorData<int16_t>(input), GetTensorShape(output), \
GetTensorData<int16_t>(output))
TF_LITE_MAX_POOL(reference_integer_ops);
#undef TF_LITE_MAX_POOL
}
template <KernelType kernel_type>
void L2EvalFloat(TfLiteContext* context, TfLiteNode* node,
TfLitePoolParams* params, OpData* data,
@ -330,6 +381,10 @@ TfLiteStatus AverageEval(TfLiteContext* context, TfLiteNode* node) {
AverageEvalQuantizedInt8<kernel_type>(context, node, params, data, input,
output);
break;
case kTfLiteInt16:
AverageEvalQuantizedInt16<kernel_type>(context, node, params, data, input,
output);
break;
default:
context->ReportError(context, "Type %d not currently supported.",
input->type);
@ -357,6 +412,10 @@ TfLiteStatus MaxEval(TfLiteContext* context, TfLiteNode* node) {
MaxEvalQuantizedInt8<kernel_type>(context, node, params, data, input,
output);
break;
case kTfLiteInt16:
MaxEvalQuantizedInt16<kernel_type>(context, node, params, data, input,
output);
break;
default:
context->ReportError(context, "Type %d not currently supported.",
input->type);

66
tensorflow/lite/kernels/pooling_test.cc
View File

@ -12,8 +12,8 @@ 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 <cstdarg>
#include <gtest/gtest.h>
#include <cstdarg>
#include "tensorflow/lite/interpreter.h"
#include "tensorflow/lite/kernels/register.h"
#include "tensorflow/lite/kernels/test_util.h"
@ -96,6 +96,25 @@ class SymmetricQuantizedPoolingOpModel : public BasePoolingOpModel {
}
};
class SymmetricQuantizedPoolingOpModel16 : public BasePoolingOpModel {
public:
using BasePoolingOpModel::BasePoolingOpModel;
void SetInput(std::initializer_list<float> data) {
QuantizeAndPopulate<int16_t>(input_, data);
}
void SetInput(const std::vector<float>& data) {
QuantizeAndPopulate<int16_t>(input_, data);
}
std::vector<int16_t> GetOutput() { return ExtractVector<int16_t>(output_); }
std::vector<float> GetDequantizedOutput() {
return Dequantize<int16_t>(ExtractVector<int16_t>(output_),
GetScale(output_), GetZeroPoint(output_));
}
};
// Replicate each entry in a vector n times along depth (innermost dimension).
// The values are incremented by delta, creating ramps offset by each input
// value. This is used to create simple and predicatable variation.
@ -398,6 +417,29 @@ TEST(QuantizedPoolingOpTest, AveragePoolLargeDepth) {
ReplicateDepthRamp(output_image_plane, depth, 1.f / 512.f),
1. / 32.f)));
}
// Test quantized AveragePool with int16 input and output. The input is the same
// as the uint8 test QuantizedPoolingOpTest.AveragePool but with a scale of
// 1/4096 rather than 1/16.
TEST(QuantizedPoolingOpTest, SymmetricAveragePool16) {
const float ulp = (float)1 / (float)4096;
SymmetricQuantizedPoolingOpModel16 m(
BuiltinOperator_AVERAGE_POOL_2D,
/*input=*/{TensorType_INT16, {1, 2, 4, 1}, 0, 16 - ulp},
/*filter_width=*/2, /*filter_height=*/2,
/*output=*/{TensorType_INT16, {}, 0, 16 - ulp});
m.SetInput({
0, 6, 2, 4, //
3, 2, 10, 7, //
});
m.Invoke();
EXPECT_THAT(m.GetDequantizedOutput(),
ElementsAreArray(ArrayFloatNear({2.75, 5.75})));
EXPECT_THAT(m.GetOutput(),
ElementsAreArray({(44 - 128) << 8, (92 - 128) << 8}));
}
// Test quantized AveragePool with int8 input and output. The input is the same
// as the uint8 test QuantizedPoolingOpTest.AveragePool. The float output is
// identical to uint8 test and quantized output is identical to uint8 test with
@ -858,6 +900,28 @@ TEST(QuantizedInt8PoolingOpTest, MaxPool) {
EXPECT_THAT(m.GetOutput(), ElementsAreArray({96 - 128, 160 - 128}));
}
TEST(QuantizedInt8PoolingOpTest16, MaxPool) {
// Choose the input ranges carefully so that the dequantized output matches
// the results of the float model above.
// Input Range[0, 16-(1/4096)] --> [Scale{(1/4096)}, zero_point{-32768}]
const float ulp = (float)1 / (float)4096;
SymmetricQuantizedPoolingOpModel16 m(
BuiltinOperator_MAX_POOL_2D,
/*input=*/{TensorType_INT16, {1, 2, 4, 1}, 0, 16 - ulp},
/*filter_width=*/2, /*filter_height=*/2,
/*output=*/{TensorType_INT16, {}, 0, 16 - ulp});
m.SetInput({
0, 6, 2, 4, //
3, 2, 10, 7, //
});
m.Invoke();
EXPECT_THAT(m.GetDequantizedOutput(),
ElementsAreArray(ArrayFloatNear({6, 10})));
EXPECT_THAT(m.GetOutput(),
ElementsAreArray({(96 - 128) << 8, (160 - 128) << 8}));
}
TEST(QuantizedInt8PoolingOpTest, MaxPoolActivationRelu) {
// Choose the input ranges carefully so that the dequantized output matches
// the results of the float model above.