Skip to content

Commit

Permalink
[luci-interpreter] Support InstanceNorm 3D kernel
Browse files Browse the repository at this point in the history 
This supports InstanceNorm 3D kernel.

ONE-DCO-1.0-Signed-off-by: ragmani <[email protected]>
  • Loading branch information
@ragmani
ragmani committed Aug 21, 2024
1 parent dad984b commit c1225bd
Show file tree
Hide file tree
Showing 2 changed files with 138 additions and 44 deletions.
161 changes: 117 additions & 44 deletions compiler/luci-interpreter/src/kernels/InstanceNorm.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -34,16 +34,35 @@ InstanceNorm::InstanceNorm(const Tensor *input, const Tensor *gamma, const Tenso

void InstanceNorm::configure()
{
LUCI_INTERPRETER_CHECK(input()->shape().num_dims() == 4);
LUCI_INTERPRETER_CHECK(input()->element_type() == output()->element_type());
LUCI_INTERPRETER_CHECK(gamma()->element_type() == input()->element_type());
LUCI_INTERPRETER_CHECK(gamma()->shape().num_dims() == 1);
LUCI_INTERPRETER_CHECK(gamma()->shape().dim(0) == input()->shape().dim(3) ||
gamma()->shape().dim(0) == 1);
LUCI_INTERPRETER_CHECK(beta()->element_type() == input()->element_type());
LUCI_INTERPRETER_CHECK(beta()->shape().num_dims() == 1);
LUCI_INTERPRETER_CHECK(beta()->shape().dim(0) == input()->shape().dim(3) ||
beta()->shape().dim(0) == 1);
if (input()->shape().num_dims() == 4)
{
LUCI_INTERPRETER_CHECK(input()->shape().num_dims() == 4);
LUCI_INTERPRETER_CHECK(gamma()->shape().num_dims() == 1);
LUCI_INTERPRETER_CHECK(gamma()->shape().dim(0) == input()->shape().dim(3) ||
gamma()->shape().dim(0) == 1);
LUCI_INTERPRETER_CHECK(beta()->element_type() == input()->element_type());
LUCI_INTERPRETER_CHECK(beta()->shape().num_dims() == 1);
LUCI_INTERPRETER_CHECK(beta()->shape().dim(0) == input()->shape().dim(3) ||
beta()->shape().dim(0) == 1);
}
else if (input()->shape().num_dims() == 3)
{
LUCI_INTERPRETER_CHECK(input()->shape().num_dims() == 3);
LUCI_INTERPRETER_CHECK(input()->element_type() == output()->element_type());
LUCI_INTERPRETER_CHECK(gamma()->element_type() == input()->element_type());
LUCI_INTERPRETER_CHECK(gamma()->shape().num_dims() == 1);
LUCI_INTERPRETER_CHECK(gamma()->shape().dim(0) == input()->shape().dim(1) ||
gamma()->shape().dim(0) == 1);
LUCI_INTERPRETER_CHECK(beta()->element_type() == input()->element_type());
LUCI_INTERPRETER_CHECK(beta()->shape().num_dims() == 1);
LUCI_INTERPRETER_CHECK(beta()->shape().dim(0) == input()->shape().dim(1) ||
beta()->shape().dim(0) == 1);
}
else
LUCI_INTERPRETER_CHECK(false && "luci-intp InstanceNorm unsupported rank.");

output()->resize(input()->shape());
}

Expand All @@ -63,58 +82,112 @@ void InstanceNorm::evalFloat() const
{
float activation_min, activation_max;
calculateActivationRange(params().activation, &activation_min, &activation_max);
auto input_shape = getTensorShape(input());
tflite::RuntimeShape input_shape = getTensorShape(input());
auto output_shape = getTensorShape(output());
const int32_t batches = tflite::MatchingDim(input_shape, 0, output_shape, 0);
const int32_t heights = tflite::MatchingDim(input_shape, 1, output_shape, 1);
const int32_t widths = tflite::MatchingDim(input_shape, 2, output_shape, 2);
const int32_t channels = tflite::MatchingDim(input_shape, 3, output_shape, 3);
const float *input_data = getTensorData<float>(input());
const float *gamma_data = getTensorData<float>(gamma());
auto gamma_shape = getTensorShape(gamma());
bool single_gamma = gamma_shape.DimensionsCount() == 1 && gamma_shape.Dims(0) == 1;
const float *beta_data = getTensorData<float>(beta());
auto beta_shape = getTensorShape(beta());
bool single_beta = beta_shape.DimensionsCount() == 1 && beta_shape.Dims(0) == 1;
float *output_data = getTensorData<float>(output());
for (int32_t batch = 0; batch < batches; batch++)

if (input_shape.DimensionsCount() == 4)
{
// Dimensions for image case are (N x H x W x C)
const int32_t batches = tflite::MatchingDim(input_shape, 0, output_shape, 0);
const int32_t heights = tflite::MatchingDim(input_shape, 1, output_shape, 1);
const int32_t widths = tflite::MatchingDim(input_shape, 2, output_shape, 2);
const int32_t channels = tflite::MatchingDim(input_shape, 3, output_shape, 3);
const float *input_data = getTensorData<float>(input());
const float *gamma_data = getTensorData<float>(gamma());
auto gamma_shape = getTensorShape(gamma());
bool single_gamma = gamma_shape.DimensionsCount() == 1 && gamma_shape.Dims(0) == 1;
const float *beta_data = getTensorData<float>(beta());
auto beta_shape = getTensorShape(beta());
bool single_beta = beta_shape.DimensionsCount() == 1 && beta_shape.Dims(0) == 1;
float *output_data = getTensorData<float>(output());
for (int32_t batch = 0; batch < batches; batch++)
{
for (int32_t channel = 0; channel < channels; channel++)
{
double sum = 0.0f;
double square_sum = 0.0f;
int32_t size = heights * widths;
for (int32_t height = 0; height < heights; height++)
{
for (int32_t width = 0; width < widths; width++)
{
double input_val =
input_data[tflite::Offset(input_shape, batch, height, width, channel)];
sum += input_val;
square_sum += (input_val * input_val);
}
}
double mean = sum / size;
double var = square_sum / size - mean * mean;

double gamma = single_gamma ? gamma_data[0] : gamma_data[channel];
double beta = single_beta ? beta_data[0] : beta_data[channel];
double a = gamma / (std::sqrt(var + params().epsilon));
double b = -mean * a + beta;

for (int32_t height = 0; height < heights; height++)
{
for (int32_t width = 0; width < widths; width++)
{
double input_value =
input_data[tflite::Offset(output_shape, batch, height, width, channel)];
double output_value = input_value * a + b;
output_data[tflite::Offset(output_shape, batch, height, width, channel)] =
tflite::ActivationFunctionWithMinMax((float)output_value, activation_min,
activation_max);
}
}
}
}
}
else if (input_shape.DimensionsCount() == 3)
{
for (int32_t channel = 0; channel < channels; channel++)
// Dimensions for non image case are (N x C x D1 x D2 … Dn)
const int32_t batches = tflite::MatchingDim(input_shape, 0, output_shape, 0);
const int32_t channels = tflite::MatchingDim(input_shape, 1, output_shape, 1);
const int32_t size = tflite::MatchingDim(input_shape, 2, output_shape, 2);
const float *input_data = getTensorData<float>(input());
const float *gamma_data = getTensorData<float>(gamma());
auto gamma_shape = getTensorShape(gamma());
bool single_gamma = gamma_shape.DimensionsCount() == 1 && gamma_shape.Dims(0) == 1;
const float *beta_data = getTensorData<float>(beta());
auto beta_shape = getTensorShape(beta());
bool single_beta = beta_shape.DimensionsCount() == 1 && beta_shape.Dims(0) == 1;
float *output_data = getTensorData<float>(output());
for (int32_t batch = 0; batch < batches; batch++)
{
double sum = 0.0f;
double square_sum = 0.0f;
int32_t size = heights * widths;
for (int32_t height = 0; height < heights; height++)
for (int32_t channel = 0; channel < channels; channel++)
{
for (int32_t width = 0; width < widths; width++)
double sum = 0.0f;
double square_sum = 0.0f;
size_t offset =
static_cast<size_t>(batch * channels * size) + static_cast<size_t>(channel * size);
for (int32_t i = 0; i < size; i++)
{
double input_val = input_data[tflite::Offset(input_shape, batch, height, width, channel)];
double input_val = input_data[offset + i];
sum += input_val;
square_sum += (input_val * input_val);
}
}
double mean = sum / size;
double var = square_sum / size - mean * mean;
double mean = sum / size;
double var = square_sum / size - mean * mean;

double gamma = single_gamma ? gamma_data[0] : gamma_data[channel];
double beta = single_beta ? beta_data[0] : beta_data[channel];
double a = gamma / (std::sqrt(var + params().epsilon));
double b = -mean * a + beta;
double gamma = single_gamma ? gamma_data[0] : gamma_data[channel];
double beta = single_beta ? beta_data[0] : beta_data[channel];
double a = gamma / (std::sqrt(var + params().epsilon));
double b = -mean * a + beta;

for (int32_t height = 0; height < heights; height++)
{
for (int32_t width = 0; width < widths; width++)
for (int32_t i = 0; i < size; i++)
{
double input_value =
input_data[tflite::Offset(output_shape, batch, height, width, channel)];
double input_value = input_data[offset + i];
double output_value = input_value * a + b;
output_data[tflite::Offset(output_shape, batch, height, width, channel)] =
tflite::ActivationFunctionWithMinMax((float)output_value, activation_min,
activation_max);
output_data[offset + i] = tflite::ActivationFunctionWithMinMax(
(float)output_value, activation_min, activation_max);
}
}
}
}
else
throw std::runtime_error("luci-intp InstanceNorm unsupported rank.");
}

} // namespace kernels
Expand Down
21 changes: 21 additions & 0 deletions compiler/luci-interpreter/src/kernels/InstanceNorm.test.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -76,6 +76,27 @@ TEST_F(InstanceNormTest, Single_gamma_beta)
EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({1, 2, 1, 2}));
}

TEST_F(InstanceNormTest, Single_gamma_beta_3D)
{
Tensor input_tensor =
makeInputTensor<DataType::FLOAT32>({1, 2, 2}, {1, 1, 1, 1}, _memory_manager.get());
Tensor gamma_tensor = makeInputTensor<DataType::FLOAT32>({1}, {1}, _memory_manager.get());
Tensor beta_tensor = makeInputTensor<DataType::FLOAT32>({1}, {2}, _memory_manager.get());
Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);

InstanceNormParams params{};
params.epsilon = 0.1f;
params.activation = Activation::NONE;

InstanceNorm kernel(&input_tensor, &gamma_tensor, &beta_tensor, &output_tensor, params);
kernel.configure();
_memory_manager->allocate_memory(output_tensor);
kernel.execute();

EXPECT_THAT(extractTensorData<float>(output_tensor), FloatArrayNear({2, 2, 2, 2}));
EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({1, 2, 2}));
}

TEST_F(InstanceNormTest, Wrong_gamma_beta_dim_NEG)
{
Tensor input_tensor =
Expand Down

0 comments on commit c1225bd

Please sign in to comment.