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

Support float32->int16 and int16->int16 quantization in TFLu

Browse Source
This commit is contained in:
Patrik Laurell 2020-06-17 16:43:22 +02:00
parent e66a20bb44
commit 412da53d65
2 changed files with 76 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

19
tensorflow/lite/micro/kernels/quantize.cc
View File

@ -66,11 +66,13 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) {
TF_LITE_ENSURE(context, input->type == kTfLiteFloat32 ||
input->type == kTfLiteInt16 ||
input->type == kTfLiteInt8);
TF_LITE_ENSURE(context,
output->type == kTfLiteUInt8 || output->type == kTfLiteInt8);
TF_LITE_ENSURE(context, output->type == kTfLiteUInt8 ||
output->type == kTfLiteInt8 ||
output->type == kTfLiteInt16);
if ((input->type == kTfLiteInt16 || input->type == kTfLiteInt8) &&
output->type == kTfLiteInt8) {
output->type == kTfLiteInt8 ||
(input->type == kTfLiteInt16 && output->type == kTfLiteInt16)) {
double effective_scale =
static_cast<double>(input->params.scale / output->params.scale);
@ -103,6 +105,11 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
op_params, GetTensorShape(input), GetTensorData<float>(input),
GetTensorShape(output), GetTensorData<uint8_t>(output));
break;
case kTfLiteInt16:
reference_ops::AffineQuantize(
op_params, GetTensorShape(input), GetTensorData<float>(input),
GetTensorShape(output), GetTensorData<int16_t>(output));
return kTfLiteOk;
default:
TF_LITE_KERNEL_LOG(context, "Input %s, output %s not supported.",
TfLiteTypeGetName(input->type),
@ -118,6 +125,12 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
data->output_shift, input->params.zero_point,
output->params.zero_point, GetTensorData<int8_t>(output));
break;
case kTfLiteInt16:
reference_ops::Requantize(
GetTensorData<int16_t>(input), size, data->output_multiplier,
data->output_shift, input->params.zero_point,
output->params.zero_point, GetTensorData<int16_t>(output));
return kTfLiteOk;
default:
TF_LITE_KERNEL_LOG(context, "Input %s, output %s not supported.",
TfLiteTypeGetName(input->type),

60
tensorflow/lite/micro/kernels/quantize_test.cc
View File

@ -198,6 +198,32 @@ TF_LITE_MICRO_TEST(QuantizeOpTestInt8NoScale) {
dims, values, dims, values, values_quantized, scale, zero_point, output);
}
TF_LITE_MICRO_TEST(QuantizeOpTestInt16) {
const int length = 10;
const int dims[] = {2, 2, 5};
const float values[] = {-63.5, -63, -62.5, -62, -61.5,
62, 62.5, 63, 63.5, 64};
const float scale = 0.5;
const int zero_point = -1;
int16_t output[length];
int16_t values_quantized[length];
tflite::testing::TestQuantizeFloat(
dims, values, dims, values, values_quantized, scale, zero_point, output);
}
TF_LITE_MICRO_TEST(QuantizeOpTestInt16NoScale) {
const int length = 10;
const int dims[] = {2, 2, 5};
const float values[] = {-128, -127, -126, -125, -124,
123, 124, 125, 126, 127};
const float scale = 1.0;
const int zero_point = 0;
int16_t output[length];
int16_t values_quantized[length];
tflite::testing::TestQuantizeFloat(
dims, values, dims, values, values_quantized, scale, zero_point, output);
}
TF_LITE_MICRO_TEST(QuantizeOpTestInt16toInt8) {
const int length = 10;
const int dims[] = {2, 2, 5};
@ -215,6 +241,40 @@ TF_LITE_MICRO_TEST(QuantizeOpTestInt16toInt8) {
output_zero_point, output_quantized);
}
TF_LITE_MICRO_TEST(QuantizeOpTestInt16toInt16) {
const int length = 10;
const int dims[] = {2, 2, 5};
const float values[] = {-64, -62, -60, -58, -56, 54, 56, 58, 60, 62};
const float input_scale = 2.f;
const int input_zero_point = 0;
const float output_scale = 0.5;
const int output_zero_point = 32;
int16_t output_quantized[length];
int16_t values_quantized[length];
int16_t input_quantized[length];
tflite::testing::TestRequantize(dims, values, input_quantized, input_scale,
input_zero_point, dims, values,
values_quantized, output_scale,
output_zero_point, output_quantized);
}
TF_LITE_MICRO_TEST(QuantizeOpTestInt16toInt16NoZeroPoint) {
const int length = 10;
const int dims[] = {2, 2, 5};
const float values[] = {-32, -31, -30, -29, -28, 27, 28, 29, 30, 31};
const float input_scale = 1.f;
const int input_zero_point = 0;
const float output_scale = 0.5;
const int output_zero_point = 0;
int16_t output_quantized[length];
int16_t values_quantized[length];
int16_t input_quantized[length];
tflite::testing::TestRequantize(dims, values, input_quantized, input_scale,
input_zero_point, dims, values,
values_quantized, output_scale,
output_zero_point, output_quantized);
}
TF_LITE_MICRO_TEST(QuantizeOpTestInt8toInt8) {
const int length = 10;
const int dims[] = {2, 2, 5};