Support LayerNormalization
Layer normalization only support float32 data type. Signed-off-by: zhao.xia <zhao.xia@verisilicon.com>
This commit is contained in:
zhao.xia
Kainan Cha
parent
c3858af4fc
commit
88f7141cfe
|
|
@ -0,0 +1,45 @@
|
|||
/****************************************************************************
|
||||
*
|
||||
* Copyright (c) 2021 Vivante Corporation
|
||||
*
|
||||
* Permission is hereby granted, free of charge, to any person obtaining a
|
||||
* copy of this software and associated documentation files (the "Software"),
|
||||
* to deal in the Software without restriction, including without limitation
|
||||
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
|
||||
* and/or sell copies of the Software, and to permit persons to whom the
|
||||
* Software is furnished to do so, subject to the following conditions:
|
||||
*
|
||||
* The above copyright notice and this permission notice shall be included in
|
||||
* all copies or substantial portions of the Software.
|
||||
*
|
||||
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
||||
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
||||
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
||||
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
||||
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
|
||||
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
|
||||
* DEALINGS IN THE SOFTWARE.
|
||||
*
|
||||
*****************************************************************************/
|
||||
#ifndef TIM_VX_OPS_LAYERNOMALIZATION_H_
|
||||
#define TIM_VX_OPS_LAYERNOMALIZATION_H_
|
||||
|
||||
#include <cstdint>
|
||||
#include "tim/vx/operation.h"
|
||||
|
||||
namespace tim {
|
||||
namespace vx {
|
||||
namespace ops {
|
||||
class LayerNormalization : public Operation {
|
||||
public:
|
||||
LayerNormalization(Graph* graph, int32_t axis = 0, float eps = 1e-5f);
|
||||
|
||||
protected:
|
||||
int32_t axis_;
|
||||
int32_t eps_;
|
||||
};
|
||||
|
||||
} // namespace ops
|
||||
} // namespace vx
|
||||
} // namespace tim
|
||||
#endif
|
||||
|
|
@ -54,7 +54,7 @@ Pad|PAD|Mapped|[tf.pad](https://tensorflow.google.cn/api_docs/python/tf/pad)
|
|||
||IMAGEPROCESS|Unmapped
|
||||
||MATRIXMUL|Unmapped|[tf.experimental.numpy.matmul](https://www.tensorflow.org/api_docs/python/tf/experimental/numpy/matmul)
|
||||
||LSTMUNIT|Unmapped
|
||||
||LAYER_NORM|Unmapped|[tf.keras.layers.LayerNormalization](https://tensorflow.google.cn/api_docs/python/tf/keras/layers/LayerNormalization)
|
||||
|LayerNormalization|LAYER_NORM|Mapped|[tf.keras.layers.LayerNormalization](https://tensorflow.google.cn/api_docs/python/tf/keras/layers/LayerNormalization)
|
||||
ReduceMin|REDUCE_MIN|Mapped|[tf.math.reduce_min](https://tensorflow.google.cn/api_docs/python/tf/math/reduce_min)
|
||||
ReduceMax|REDUCE_MAX|Mapped|[tf.math.reduce_max](https://tensorflow.google.cn/api_docs/python/tf/math/reduce_max)
|
||||
ReduceAny|REDUCE_ANY|Mapped|[tf.math.reduce_any](https://tensorflow.google.cn/api_docs/python/tf/math/reduce_any)
|
||||
|
|
|
|||
|
|
@ -0,0 +1,45 @@
|
|||
/****************************************************************************
|
||||
*
|
||||
* Copyright (c) 2021 Vivante Corporation
|
||||
*
|
||||
* Permission is hereby granted, free of charge, to any person obtaining a
|
||||
* copy of this software and associated documentation files (the "Software"),
|
||||
* to deal in the Software without restriction, including without limitation
|
||||
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
|
||||
* and/or sell copies of the Software, and to permit persons to whom the
|
||||
* Software is furnished to do so, subject to the following conditions:
|
||||
*
|
||||
* The above copyright notice and this permission notice shall be included in
|
||||
* all copies or substantial portions of the Software.
|
||||
*
|
||||
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
||||
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
||||
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
||||
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
||||
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
|
||||
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
|
||||
* DEALINGS IN THE SOFTWARE.
|
||||
*
|
||||
*****************************************************************************/
|
||||
#include "tim/vx/ops/layernormalization.h"
|
||||
|
||||
#include <cassert>
|
||||
#include "operation_private.h"
|
||||
#include "vsi_nn_pub.h"
|
||||
|
||||
namespace tim {
|
||||
namespace vx {
|
||||
namespace ops {
|
||||
LayerNormalization::LayerNormalization(Graph* graph, int32_t axis, float eps)
|
||||
: Operation(graph, VSI_NN_OP_LAYER_NORM), axis_(axis), eps_(eps) {
|
||||
// Layer normalization shares the parameters of instance normalization.
|
||||
if (axis != 0) {
|
||||
VSILOGE("Layer norm only support axis 0.");
|
||||
assert(false);
|
||||
}
|
||||
this->impl()->node()->nn_param.instancenorm.eps = eps_;
|
||||
}
|
||||
|
||||
} // namespace ops
|
||||
} // namespace vx
|
||||
} // namespace tim
|
||||
|
|
@ -0,0 +1,218 @@
|
|||
/****************************************************************************
|
||||
*
|
||||
* Copyright (c) 2021 Vivante Corporation
|
||||
*
|
||||
* Permission is hereby granted, free of charge, to any person obtaining a
|
||||
* copy of this software and associated documentation files (the "Software"),
|
||||
* to deal in the Software without restriction, including without limitation
|
||||
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
|
||||
* and/or sell copies of the Software, and to permit persons to whom the
|
||||
* Software is furnished to do so, subject to the following conditions:
|
||||
*
|
||||
* The above copyright notice and this permission notice shall be included in
|
||||
* all copies or substantial portions of the Software.
|
||||
*
|
||||
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
||||
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
||||
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
||||
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
||||
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
|
||||
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
|
||||
* DEALINGS IN THE SOFTWARE.
|
||||
*
|
||||
*****************************************************************************/
|
||||
#include "tim/vx/context.h"
|
||||
#include "tim/vx/graph.h"
|
||||
#include "tim/vx/ops/layernormalization.h"
|
||||
|
||||
#include "gtest/gtest.h"
|
||||
|
||||
namespace {
|
||||
template<typename T>
|
||||
::testing::AssertionResult ArraysMatch(const std::vector<T>& expected,
|
||||
const std::vector<T>& actual,
|
||||
T abs_error){
|
||||
for (size_t i = 0; i < expected.size(); ++i){
|
||||
EXPECT_NEAR(expected[i], actual[i], abs_error) << "at index:" << i;
|
||||
}
|
||||
|
||||
return ::testing::AssertionSuccess();
|
||||
}
|
||||
}
|
||||
|
||||
TEST(LayerNorm, axis_0_shape_3_6_1_float) {
|
||||
auto ctx = tim::vx::Context::Create();
|
||||
auto graph = ctx->CreateGraph();
|
||||
|
||||
tim::vx::ShapeType io_shape({3, 6, 1});
|
||||
tim::vx::ShapeType param_shape({6});
|
||||
tim::vx::TensorSpec input_spec(tim::vx::DataType::FLOAT32,
|
||||
io_shape, tim::vx::TensorAttribute::INPUT);
|
||||
tim::vx::TensorSpec param_spec(tim::vx::DataType::FLOAT32,
|
||||
param_shape, tim::vx::TensorAttribute::INPUT);
|
||||
tim::vx::TensorSpec output_spec(tim::vx::DataType::FLOAT32,
|
||||
io_shape, tim::vx::TensorAttribute::OUTPUT);
|
||||
|
||||
auto input_tensor = graph->CreateTensor(input_spec);
|
||||
auto gamma_tensor = graph->CreateTensor(param_spec);
|
||||
auto beta_tensor = graph->CreateTensor(param_spec);
|
||||
auto output_tensor = graph->CreateTensor(output_spec);
|
||||
|
||||
std::vector<float> in_data = {
|
||||
-2, 0, 2,
|
||||
-3, 0, 3,
|
||||
-4, 0, 4,
|
||||
-5, 0, 5,
|
||||
-6, 0, 6,
|
||||
-7, 0, 7 };
|
||||
std::vector<float> gamma = {
|
||||
1.0f, 1.0f, 1.0f,
|
||||
1.0f, 1.0f, 1.0f
|
||||
};
|
||||
std::vector<float> beta = {
|
||||
.0f, .0f, .0f,
|
||||
.0f, .0f, .0f
|
||||
};
|
||||
std::vector<float> golden = {
|
||||
-1.22474f, 0, 1.22474f,
|
||||
-1.22474f, 0, 1.22474f,
|
||||
-1.22474f, 0, 1.22474f,
|
||||
-1.22474f, 0, 1.22474f,
|
||||
-1.22474f, 0, 1.22474f,
|
||||
-1.22474f, 0, 1.22474f,
|
||||
};
|
||||
|
||||
EXPECT_TRUE(input_tensor->CopyDataToTensor(in_data.data(), in_data.size() * sizeof(float)));
|
||||
EXPECT_TRUE(gamma_tensor->CopyDataToTensor(gamma.data(), gamma.size() * sizeof(float)));
|
||||
EXPECT_TRUE(beta_tensor->CopyDataToTensor(beta.data(), beta.size() * sizeof(float)));
|
||||
|
||||
auto op = graph->CreateOperation<tim::vx::ops::LayerNormalization>(0, 2e-5f);
|
||||
(*op).BindInputs({input_tensor, beta_tensor, gamma_tensor}).BindOutputs({output_tensor});
|
||||
|
||||
EXPECT_TRUE(graph->Compile());
|
||||
EXPECT_TRUE(graph->Run());
|
||||
|
||||
std::vector<float> output(18);
|
||||
EXPECT_TRUE(output_tensor->CopyDataFromTensor(output.data()));
|
||||
EXPECT_TRUE(ArraysMatch(golden, output, 1e-5f));
|
||||
}
|
||||
|
||||
TEST(LayerNorm, axis_0_shape_2_3_6_1_float) {
|
||||
auto ctx = tim::vx::Context::Create();
|
||||
auto graph = ctx->CreateGraph();
|
||||
|
||||
tim::vx::ShapeType io_shape({2, 3, 6, 1});
|
||||
tim::vx::ShapeType param_shape({6});
|
||||
tim::vx::TensorSpec input_spec(tim::vx::DataType::FLOAT32,
|
||||
io_shape, tim::vx::TensorAttribute::INPUT);
|
||||
tim::vx::TensorSpec param_spec(tim::vx::DataType::FLOAT32,
|
||||
param_shape, tim::vx::TensorAttribute::INPUT);
|
||||
tim::vx::TensorSpec output_spec(tim::vx::DataType::FLOAT32,
|
||||
io_shape, tim::vx::TensorAttribute::OUTPUT);
|
||||
|
||||
auto input_tensor = graph->CreateTensor(input_spec);
|
||||
auto gamma_tensor = graph->CreateTensor(param_spec);
|
||||
auto beta_tensor = graph->CreateTensor(param_spec);
|
||||
auto output_tensor = graph->CreateTensor(output_spec);
|
||||
|
||||
std::vector<float> in_data = {
|
||||
-2, 2, -2, 2, -2, 2,
|
||||
-3, 3, -3, 3, -3, 3,
|
||||
-4, 4, -4, 4, -4, 4,
|
||||
-5, 5, -5, 5, -5, 5,
|
||||
-6, 6, -6, 6, -6, 6,
|
||||
-7, 7, -7, 7, -7, 7
|
||||
};
|
||||
std::vector<float> gamma = {
|
||||
1.0f, 1.0f, 1.0f,
|
||||
1.0f, 1.0f, 1.0f
|
||||
};
|
||||
std::vector<float> beta = {
|
||||
.0f, .0f, .0f,
|
||||
.0f, .0f, .0f
|
||||
};
|
||||
std::vector<float> golden = {
|
||||
-1.f, 1.f, -1.f, 1.f, -1.f, 1.f,
|
||||
-1.f, 1.f, -1.f, 1.f, -1.f, 1.f,
|
||||
-1.f, 1.f, -1.f, 1.f, -1.f, 1.f,
|
||||
-1.f, 1.f, -1.f, 1.f, -1.f, 1.f,
|
||||
-1.f, 1.f, -1.f, 1.f, -1.f, 1.f,
|
||||
-1.f, 1.f, -1.f, 1.f, -1.f, 1.f,
|
||||
};
|
||||
|
||||
EXPECT_TRUE(input_tensor->CopyDataToTensor(in_data.data(), in_data.size() * sizeof(float)));
|
||||
EXPECT_TRUE(gamma_tensor->CopyDataToTensor(gamma.data(), gamma.size() * sizeof(float)));
|
||||
EXPECT_TRUE(beta_tensor->CopyDataToTensor(beta.data(), beta.size() * sizeof(float)));
|
||||
|
||||
auto op = graph->CreateOperation<tim::vx::ops::LayerNormalization>(0, 2e-5f);
|
||||
(*op).BindInputs({input_tensor, beta_tensor, gamma_tensor}).BindOutputs({output_tensor});
|
||||
|
||||
EXPECT_TRUE(graph->Compile());
|
||||
EXPECT_TRUE(graph->Run());
|
||||
|
||||
std::vector<float> output(36);
|
||||
EXPECT_TRUE(output_tensor->CopyDataFromTensor(output.data()));
|
||||
EXPECT_TRUE(ArraysMatch(golden, output, 1e-5f));
|
||||
}
|
||||
|
||||
#if 0
|
||||
// Fail case
|
||||
TEST(LayerNorm, axis_0_shape_3_6_1_uint8) {
|
||||
auto ctx = tim::vx::Context::Create();
|
||||
auto graph = ctx->CreateGraph();
|
||||
|
||||
tim::vx::ShapeType io_shape({3, 6, 1});
|
||||
tim::vx::ShapeType param_shape({6});
|
||||
tim::vx::Quantization input_quant(tim::vx::QuantType::ASYMMETRIC, 1, 7);
|
||||
tim::vx::Quantization output_quant(tim::vx::QuantType::ASYMMETRIC, 1.22474f, 1);
|
||||
tim::vx::TensorSpec input_spec(tim::vx::DataType::UINT8,
|
||||
io_shape, tim::vx::TensorAttribute::INPUT, input_quant);
|
||||
tim::vx::TensorSpec param_spec(tim::vx::DataType::FLOAT32,
|
||||
param_shape, tim::vx::TensorAttribute::INPUT);
|
||||
tim::vx::TensorSpec output_spec(tim::vx::DataType::UINT8,
|
||||
io_shape, tim::vx::TensorAttribute::OUTPUT, output_quant);
|
||||
|
||||
auto input_tensor = graph->CreateTensor(input_spec);
|
||||
auto gamma_tensor = graph->CreateTensor(param_spec);
|
||||
auto beta_tensor = graph->CreateTensor(param_spec);
|
||||
auto output_tensor = graph->CreateTensor(output_spec);
|
||||
|
||||
std::vector<float> in_data = {
|
||||
5, 7, 9,
|
||||
4, 7, 10,
|
||||
3, 7, 11,
|
||||
2, 7, 12,
|
||||
1, 7, 13,
|
||||
0, 7, 14 };
|
||||
std::vector<float> gamma = {
|
||||
1.0f, 1.0f, 1.0f,
|
||||
1.0f, 1.0f, 1.0f
|
||||
};
|
||||
std::vector<float> beta = {
|
||||
.0f, .0f, .0f,
|
||||
.0f, .0f, .0f
|
||||
};
|
||||
std::vector<float> golden = {
|
||||
0, 1, 2,
|
||||
0, 1, 2,
|
||||
0, 1, 2,
|
||||
0, 1, 2,
|
||||
0, 1, 2,
|
||||
0, 1, 2,
|
||||
};
|
||||
|
||||
EXPECT_TRUE(input_tensor->CopyDataToTensor(in_data.data(), in_data.size()));
|
||||
EXPECT_TRUE(gamma_tensor->CopyDataToTensor(gamma.data(), gamma.size() * sizeof(float)));
|
||||
EXPECT_TRUE(beta_tensor->CopyDataToTensor(beta.data(), beta.size() * sizeof(float)));
|
||||
|
||||
auto op = graph->CreateOperation<tim::vx::ops::LayerNormalization>(0, 2e-5f);
|
||||
(*op).BindInputs({input_tensor, beta_tensor, gamma_tensor}).BindOutputs({output_tensor});
|
||||
|
||||
EXPECT_TRUE(graph->Compile());
|
||||
EXPECT_TRUE(graph->Run());
|
||||
|
||||
std::vector<float> output(18);
|
||||
EXPECT_TRUE(output_tensor->CopyDataFromTensor(output.data()));
|
||||
EXPECT_EQ(golden, output);
|
||||
}
|
||||
#endif
|
||||
Loading…
Reference in New Issue