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

Enable per tensor quantized weight for per channel int8 transpose conv ops.

Browse Source 
PiperOrigin-RevId: 353147555
Change-Id: I07b2b7f9aa1580ff5e9ac20dbe06bfa61e36171d
This commit is contained in:
Jaehong Kim 2021-01-21 18:24:42 -08:00 committed by TensorFlower Gardener
parent 9045dcaf27
commit f98a42d0a0
2 changed files with 38 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
tensorflow/lite/kernels/transpose_conv.cc
View File

@ -80,8 +80,6 @@ struct OpData {
int output_shift;
// Per channel output multiplier and shift.
// TODO(b/144846950): Add channel dimension index for the kernel to be more
// flexible.
std::vector<int32_t> per_channel_output_multiplier;
std::vector<int32_t> per_channel_output_shift;
@ -374,17 +372,20 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) {
const auto* affine_quantization =
reinterpret_cast<TfLiteAffineQuantization*>(
weights->quantization.params);
const int channels_out = weights->dims->data[0];
TF_LITE_ENSURE(context, affine_quantization);
TF_LITE_ENSURE(context, affine_quantization->scale);
const int number_channel = affine_quantization->scale->size;
data->per_channel_output_multiplier.resize(number_channel);
data->per_channel_output_shift.resize(number_channel);
TF_LITE_ENSURE(context, (affine_quantization->scale->size == 1 ||
affine_quantization->scale->size == channels_out));
data->per_channel_output_multiplier.resize(channels_out);
data->per_channel_output_shift.resize(channels_out);
TF_LITE_ENSURE_STATUS(tflite::PopulateConvolutionQuantizationParams(
context, input, weights, bias, output, kTfLiteActNone,
&data->output_multiplier, &data->output_shift,
&data->output_activation_min, &data->output_activation_max,
data->per_channel_output_multiplier.data(),
data->per_channel_output_shift.data()));
data->per_channel_output_shift.data(), channels_out));
}
return kTfLiteOk;

31
tensorflow/lite/kernels/transpose_conv_test.cc
View File

@ -462,6 +462,37 @@ TEST_P(TransposeConvOpTest, TestQuantizedPerChannelMultiChannel) {
EXPECT_THAT(model.GetOutputShape(), ElementsAreArray({1, 5, 5, 2}));
}
// Test data copied from the float multi-channel test above.
TEST_P(TransposeConvOpTest, TestQuantizedPerTensorMultiChannel) {
const std::initializer_list<float> filter_data = {
1, 3, 5, 7, 9, 11, 13, 15, 17, 2, 4, 6, 8, 10, 12, 14, 16, 18};
const std::initializer_list<int8_t> const_filter_data = {
7, 21, 35, 49, 64, 78, 92, 106, 120,
14, 28, 42, 56, 71, 85, 99, 113, 127};
PerChannelQuantizedTransposeConvOpModel model(
GetRegistration(), {1, 5, 5, 2},
{TensorType_INT8, {2, 3, 3, 1}, 0, 0, 0, 0, true, {18.0 / 127}, {0}, 0},
const_filter_data, {TensorType_INT8, {1, 2, 2, 1}, 0, 0, 4.0 / 255, -128},
{TensorType_INT8, {}, 0, 0, 1, -128}, Padding_VALID, 2, 2, GetTestType(),
/* version */ 2);
model.SetInput({1, 2, 3, 4});
if (GetTestType() == TestType::kDynamic) {
model.SetFilter(filter_data);
}
model.Invoke();
EXPECT_THAT(
model.GetDequantizedOutput(),
ElementsAreArray(ArrayFloatNear(
{1, 2, 3, 4, 7, 10, 6, 8, 10, 12, 7, 8, 9, 10, 25, 28, 18,
20, 22, 24, 16, 20, 24, 28, 62, 72, 42, 48, 54, 60, 21, 24, 27, 30,
61, 68, 36, 40, 44, 48, 39, 42, 45, 48, 103, 110, 60, 64, 68, 72},
1e-5)));
// GetOutputShape() should always be same as model.SetOutputShape(...);
EXPECT_THAT(model.GetOutputShape(), ElementsAreArray({1, 5, 5, 2}));
}
class PerChannelQuantizedTransposeConvOpModel16x8
: public BaseTransposeConvOpModel<int16_t> {
public: