COMPMID-3442: Add support of negative axis in NESoftmaxLayer and reference code

Signed-off-by: Sheri Zhang <sheri.zhang@arm.com>
Change-Id: I285cc3b74ac0a45f0ad5830baed5237cea568f15
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3147
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

9 files changed, 70 insertions, 55 deletions

diff --git a/arm_compute/core/NEON/kernels/NESoftmaxLayerKernel.h b/arm_compute/core/NEON/kernels/NESoftmaxLayerKernel.h
index dfcc2ffe2..0e0be7936 100644
--- a/arm_compute/core/NEON/kernels/NESoftmaxLayerKernel.h
+++ b/arm_compute/core/NEON/kernels/NESoftmaxLayerKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2019 ARM Limited.
+ * Copyright (c) 2017-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -43,13 +43,13 @@ public:
     NELogits1DMaxKernel();
     /** Set the input and output tensors.
      *
-     * @param[in]  input  Source tensor. Data types supported: QASYMM8/F16/F32.
+     * @param[in]  input  Source tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
      * @param[out] output Destination tensor. Data types supported: same as @p input
      */
     void configure(const ITensor *input, ITensor *output);
     /** Static function to check if given info will lead to a valid configuration of @ref NELogits1DMaxKernel
      *
-     * @param[in] input  Source tensor. Data types supported: QASYMM8/F16/F32.
+     * @param[in] input  Source tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
      * @param[in] output Destination tensor. Data types supported: same as @p input
      *
      * @return a status
@@ -98,7 +98,7 @@ public:
     ~NELogits1DSoftmaxKernel() = default;
     /** Set the input and output tensors.
      *
-     * @param[in]  input  Source tensor. Data types supported: QASYMM8/F16/F32.
+     * @param[in]  input  Source tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
      * @param[in]  max    Max values tensor. Same shape as input with dimension 0 set to 1.
      *                    Data types supported: same as @p input.
      * @param[out] output Destination tensor. Data types supported: same as @p input.
@@ -109,7 +109,7 @@ public:
     void configure(const ITensor *input, const ITensor *max, ITensor *output, const float beta, ITensor *tmp);
     /** Static function to check if given info will lead to a valid configuration of @ref NELogits1DSoftmaxKernel
      *
-     * @param[in] input  Source tensor info. Data types supported: QASYMM8/F16/F32.
+     * @param[in] input  Source tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
      * @param[in] max    Max values tensor info. Same shape as input with dimension 0 set to 1.
      *                   Data types supported: same as @p input.
      * @param[in] output Destination tensor info. Data types supported: same as @p input.
diff --git a/arm_compute/runtime/NEON/functions/NESoftmaxLayer.h b/arm_compute/runtime/NEON/functions/NESoftmaxLayer.h
index f89add71e..b80ceaf25 100644
--- a/arm_compute/runtime/NEON/functions/NESoftmaxLayer.h
+++ b/arm_compute/runtime/NEON/functions/NESoftmaxLayer.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2019 ARM Limited.
+ * Copyright (c) 2017-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -65,28 +65,30 @@ public:
     NESoftmaxLayerGeneric &operator=(NESoftmaxLayerGeneric &&) = default;
     /** Set the input and output tensors.
      *
-     * @param[in,out] input  Source tensor. Data types supported: QASYMM8/F16/F32. If the width is not a
+     * @param[in,out] input  Source tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. If the width is not a
      *                       multiple of the internal processing block size, @ref NEFillBorderKernel replicates the
      *                       last value of each row to the nearest multiple.
      * @param[out]    output Destination tensor. Data types supported: same as @p input.
      * @param[in]     beta   (Optional) A scaling factor for the exponent.
-     * @param[in]     axis   (Optional) Reduction axis. Defaults to 1. Must be in range [1, input_num_dimensions).
+     * @param[in]     axis   (Optional) Reduction axis. Defaults to -1.
+     *                       Negative index is used to specify axis from the end (e.g. -1 for the last axis).Must be in range [-input_num_dimensions, input_num_dimensions).
      *                       It has the purpose of squashing the first @p axis dimensions together. For instance, given a [4x4x4x4] image,
      *                       when @p axis is 2, the Softmax reduction will be applied on each of the [4x4] planes of the input image.
      */
-    void configure(ITensor *input, ITensor *output, float beta = 1.0f, size_t axis = 1);
+    void configure(ITensor *input, ITensor *output, float beta = 1.0f, int32_t axis = -1);
     /** Static function to check if given info will lead to a valid configuration of @ref NESoftmaxLayer
      *
-     * @param[in] input  Source tensor info. Data types supported: QASYMM8/F16/F32.
+     * @param[in] input  Source tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
      * @param[in] output Destination tensor info. Data types supported: same as @p input
      * @param[in] beta   (Optional) A scaling factor for the exponent.
-     * @param[in] axis   (Optional) Reduction axis. Defaults to 1. Must be in range [1, input_num_dimensions).
+     * @param[in] axis   (Optional) Reduction axis. Defaults to -1.
+     *                   Negative index is used to specify axis from the end (e.g. -1 for the last axis).Must be in range [-input_num_dimensions, input_num_dimensions).
      *                   It has the purpose of squashing the first @p axis dimensions together. For instance, given a [4x4x4x4] image,
      *                   when @p axis is 2, the Softmax reduction will be applied on each of the [4x4] planes of the input image.
      *
      * @return a status
      */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *output, float beta = 1.0f, size_t axis = 1);
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, float beta = 1.0f, int32_t axis = -1);
 
     // Inherited methods overridden:
     void run() override;
@@ -101,11 +103,12 @@ private:
      *
      * @param[in] input  Original source tensor.
      * @param[in] output Original destination tensor.
-     * @param[in] axis   (Optional) Reduction axis. Defaults to 1. Must be in range [1, input_num_dimensions).
+     * @param[in] axis   (Optional) Reduction axis. Defaults to -1.
+     *                   Negative index is used to specify axis from the end (e.g. -1 for the last axis).Must be in range [-input_num_dimensions, input_num_dimensions).
      *                   It has the purpose of squashing the first @p axis dimensions together. For instance, given a [4x4x4x4] image,
      *                   when @p axis is 2, the Softmax reduction will be applied on each of the [4x4] planes of the input image.
      */
-    void configure_reshape_input_kernel(const ITensor *input, const ITensor *output, size_t axis);
+    void configure_reshape_input_kernel(const ITensor *input, const ITensor *output, int32_t axis);
 
     MemoryGroup                     _memory_group;
     NELogits1DMaxKernel             _max_kernel;
diff --git a/src/runtime/NEON/functions/NESoftmaxLayer.cpp b/src/runtime/NEON/functions/NESoftmaxLayer.cpp
index f530a87d0..57d75af77 100644
--- a/src/runtime/NEON/functions/NESoftmaxLayer.cpp
+++ b/src/runtime/NEON/functions/NESoftmaxLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2019 ARM Limited.
+ * Copyright (c) 2017-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -41,7 +41,7 @@ NESoftmaxLayerGeneric<IS_LOG>::NESoftmaxLayerGeneric(std::shared_ptr<IMemoryMana
 }
 
 template <bool IS_LOG>
-void NESoftmaxLayerGeneric<IS_LOG>::configure_reshape_input_kernel(const ITensor *input, const ITensor *output, size_t axis)
+void NESoftmaxLayerGeneric<IS_LOG>::configure_reshape_input_kernel(const ITensor *input, const ITensor *output, int32_t axis)
 {
     // Flatten the input
     const TensorShape shape_flatten = misc::shape_calculator::compute_softmax_shape(input->info(), axis);
@@ -71,12 +71,15 @@ void NESoftmaxLayerGeneric<IS_LOG>::configure_reshape_input_kernel(const ITensor
 }
 
 template <bool IS_LOG>
-void NESoftmaxLayerGeneric<IS_LOG>::configure(ITensor *input, ITensor *output, float beta, size_t axis)
+void NESoftmaxLayerGeneric<IS_LOG>::configure(ITensor *input, ITensor *output, float beta, int32_t axis)
 {
     // Perform validation step
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
     ARM_COMPUTE_ERROR_THROW_ON(NESoftmaxLayerGeneric::validate(input->info(), output->info(), beta, axis));
 
+    // Handle negative axis, negative index is used to specify axis from the end (e.g. -1 for the last axis).
+    axis = wrap_around(axis, static_cast<int32_t>(input->info()->num_dimensions()));
+
     // We don't need flattening only in the case the input is 2D and axis is 1
     _needs_flattening = axis != 1;
 
@@ -142,13 +145,16 @@ void NESoftmaxLayerGeneric<IS_LOG>::configure(ITensor *input, ITensor *output, f
 }
 
 template <bool IS_LOG>
-Status NESoftmaxLayerGeneric<IS_LOG>::validate(const ITensorInfo *input, const ITensorInfo *output, float beta, size_t axis)
+Status NESoftmaxLayerGeneric<IS_LOG>::validate(const ITensorInfo *input, const ITensorInfo *output, float beta, int32_t axis)
 {
     // Perform validation step
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->num_dimensions() > 4, "Only up to 4 dimensions are supported");
     ARM_COMPUTE_UNUSED(beta);
-    ARM_COMPUTE_RETURN_ERROR_ON(axis < 1 || input->num_dimensions() < axis);
+    ARM_COMPUTE_RETURN_ERROR_ON(axis < static_cast<int32_t>(-input->num_dimensions()) || static_cast<int32_t>(input->num_dimensions()) <= axis);
+
+    // Handle negative axis, negative index is used to specify axis from the end (e.g. -1 for the last axis).
+    axis = wrap_around(axis, static_cast<int32_t>(input->num_dimensions()));
 
     // Create intermediate tensor info
     DataType         tmp_data_type = input->data_type();
diff --git a/tests/validation/NEON/SoftmaxLayer.cpp b/tests/validation/NEON/SoftmaxLayer.cpp
index c429782e6..8af3847cf 100644
--- a/tests/validation/NEON/SoftmaxLayer.cpp
+++ b/tests/validation/NEON/SoftmaxLayer.cpp
@@ -97,9 +97,9 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(
                framework::dataset::make("axis", { 1,
                                                   1,
                                                   1,
+                                                  -1,
                                                   1,
-                                                  1,
-                                                  0,
+                                                  -3,
                                                 })),
                framework::dataset::make("Expected", { false, false, false, true, true, false })),
                input_info, output_info, beta, axis, expected)
@@ -188,7 +188,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall4D, NESoftmaxLayerQuantizedFixture<uint8_t>, fram
                                                                                                                  framework::dataset::make("DataType", DataType::QASYMM8)),
                                                                                                                  combine(framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, -10) }),
                                                                                                                          framework::dataset::make("Beta", { 1.0f, 2.f }))),
-                                                                                                                 framework::dataset::make("Axis", { 1, 2, 3 })))
+                                                                                                                 framework::dataset::make("Axis", { -1, 2, 3 })))
 {
     // Validate output
     validate(Accessor(_target), _reference, tolerance_qasymm8);
@@ -209,7 +209,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall2D, NESoftmaxLayerQuantizedFixture<int8_t>, frame
                                                                                                                         framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)),
                                                                                                                         combine(framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, -10) }),
                                                                                                                                 framework::dataset::make("Beta", { 1.0f, 2.f }))),
-                                                                                                                framework::dataset::make("Axis", { 1 })))
+                                                                                                                framework::dataset::make("Axis", { -1, 1 })))
 {
     // Validate output
     validate(Accessor(_target), _reference, tolerance_qasymm8_signed);
@@ -218,7 +218,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall4D, NESoftmaxLayerQuantizedFixture<int8_t>, frame
                                                                                                                         framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)),
                                                                                                                         combine(framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, -10) }),
                                                                                                                                 framework::dataset::make("Beta", { 1.0f, 2.f }))),
-                                                                                                                framework::dataset::make("Axis", { 1, 2, 3 })))
+                                                                                                                framework::dataset::make("Axis", { -2, 2, 3 })))
 {
     // Validate output
     validate(Accessor(_target), _reference, tolerance_qasymm8_signed);
diff --git a/tests/validation/fixtures/SoftmaxLayerFixture.h b/tests/validation/fixtures/SoftmaxLayerFixture.h
index 82daf34f1..aeff77777 100644
--- a/tests/validation/fixtures/SoftmaxLayerFixture.h
+++ b/tests/validation/fixtures/SoftmaxLayerFixture.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2019 ARM Limited.
+ * Copyright (c) 2017-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -73,7 +73,7 @@ protected:
     }
 
     TensorType compute_target(const TensorShape &shape, DataType data_type,
-                              QuantizationInfo quantization_info, float beta, size_t axis)
+                              QuantizationInfo quantization_info, float beta, int32_t axis)
     {
         // Create tensors
         TensorType src = create_tensor<TensorType>(shape, data_type, 1, quantization_info);
@@ -103,7 +103,7 @@ protected:
     }
 
     SimpleTensor<T> compute_reference(const TensorShape &shape, DataType data_type,
-                                      QuantizationInfo quantization_info, float beta, size_t axis)
+                                      QuantizationInfo quantization_info, float beta, int32_t axis)
     {
         // Create reference
         SimpleTensor<T> src{ shape, data_type, 1, quantization_info };
diff --git a/tests/validation/reference/LogSoftmaxLayer.cpp b/tests/validation/reference/LogSoftmaxLayer.cpp
index e4403956a..edb208e6a 100644
--- a/tests/validation/reference/LogSoftmaxLayer.cpp
+++ b/tests/validation/reference/LogSoftmaxLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019 ARM Limited.
+ * Copyright (c) 2019-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -35,13 +35,13 @@ namespace validation
 namespace reference
 {
 template <typename T, typename std::enable_if<is_floating_point<T>::value, int>::type>
-SimpleTensor<T> log_softmax_layer(const SimpleTensor<T> &src, float beta, size_t axis)
+SimpleTensor<T> log_softmax_layer(const SimpleTensor<T> &src, float beta, int32_t axis)
 {
     return softmax_layer_generic<T>(src, beta, axis, true);
 }
 
-template <typename T, typename std::enable_if<std::is_same<T, uint8_t>::value || std::is_same<T, int8_t>::value, int>::type>
-SimpleTensor<T> log_softmax_layer(const SimpleTensor<T> &src, float beta, size_t axis)
+template < typename T, typename std::enable_if < std::is_same<T, uint8_t>::value || std::is_same<T, int8_t>::value, int >::type >
+SimpleTensor<T> log_softmax_layer(const SimpleTensor<T> &src, float beta, int32_t axis)
 {
     const QuantizationInfo output_quantization_info = arm_compute::get_softmax_output_quantization_info(src.data_type(), true);
 
@@ -51,10 +51,10 @@ SimpleTensor<T> log_softmax_layer(const SimpleTensor<T> &src, float beta, size_t
     return dst;
 }
 
-template SimpleTensor<float> log_softmax_layer(const SimpleTensor<float> &src, float beta, size_t axis);
-template SimpleTensor<half> log_softmax_layer(const SimpleTensor<half> &src, float beta, size_t axis);
-template SimpleTensor<uint8_t> log_softmax_layer(const SimpleTensor<uint8_t> &src, float beta, size_t axis);
-template SimpleTensor<int8_t> log_softmax_layer(const SimpleTensor<int8_t> &src, float beta, size_t axis);
+template SimpleTensor<float> log_softmax_layer(const SimpleTensor<float> &src, float beta, int32_t axis);
+template SimpleTensor<half> log_softmax_layer(const SimpleTensor<half> &src, float beta, int32_t axis);
+template SimpleTensor<uint8_t> log_softmax_layer(const SimpleTensor<uint8_t> &src, float beta, int32_t axis);
+template SimpleTensor<int8_t> log_softmax_layer(const SimpleTensor<int8_t> &src, float beta, int32_t axis);
 } // namespace reference
 } // namespace validation
 } // namespace test
diff --git a/tests/validation/reference/LogSoftmaxLayer.h b/tests/validation/reference/LogSoftmaxLayer.h
index c2e3f5974..48ffdcfbc 100644
--- a/tests/validation/reference/LogSoftmaxLayer.h
+++ b/tests/validation/reference/LogSoftmaxLayer.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019 ARM Limited.
+ * Copyright (c) 2019-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -36,10 +36,10 @@ namespace validation
 namespace reference
 {
 template <typename T, typename std::enable_if<is_floating_point<T>::value, int>::type = 0>
-SimpleTensor<T> log_softmax_layer(const SimpleTensor<T> &src, float beta, size_t axis = 1);
+SimpleTensor<T> log_softmax_layer(const SimpleTensor<T> &src, float beta, int32_t axis = -1);
 
-template <typename T, typename std::enable_if<std::is_same<T, uint8_t>::value || std::is_same<T, int8_t>::value, int>::type = 0>
-SimpleTensor<T> log_softmax_layer(const SimpleTensor<T> &src, float beta, size_t axis = 1);
+template < typename T, typename std::enable_if < std::is_same<T, uint8_t>::value || std::is_same<T, int8_t>::value, int >::type = 0 >
+SimpleTensor<T> log_softmax_layer(const SimpleTensor<T> &src, float beta, int32_t axis = -1);
 } // namespace reference
 } // namespace validation
 } // namespace test
diff --git a/tests/validation/reference/SoftmaxLayer.cpp b/tests/validation/reference/SoftmaxLayer.cpp
index ee7a5f175..2fe1faef5 100644
--- a/tests/validation/reference/SoftmaxLayer.cpp
+++ b/tests/validation/reference/SoftmaxLayer.cpp
@@ -34,23 +34,29 @@ namespace validation
 namespace reference
 {
 template <typename T, typename std::enable_if<is_floating_point<T>::value, int>::type>
-SimpleTensor<T> softmax_layer_generic(const SimpleTensor<T> &src, float beta, size_t axis, bool is_log)
+SimpleTensor<T> softmax_layer_generic(const SimpleTensor<T> &src, float beta, int32_t axis, bool is_log)
 {
     // Create reference
     SimpleTensor<T> dst{ src.shape(), src.data_type(), 1 };
 
+    // Negative index is used to specify axis from the end (e.g. -1 for the last axis).
+    if(axis < 0)
+    {
+        axis += src.shape().num_dimensions();
+    }
+
     // Compute reference. Lower dims are the collapsing of the first axis
     // dimensions (i.e., the flattened dimension of each batch). The upper dims are
     // instead the batches we want to normalize
 
     int lower_dims = 1;
-    for(size_t i = 0; i < axis; i++)
+    for(size_t i = 0; i < static_cast<size_t>(axis); ++i)
     {
         lower_dims *= src.shape()[i];
     }
 
     int upper_dims = 1;
-    for(size_t i = axis; i < TensorShape::num_max_dimensions; i++)
+    for(size_t i = static_cast<size_t>(axis); i < TensorShape::num_max_dimensions; ++i)
     {
         upper_dims *= src.shape()[i];
     }
@@ -101,17 +107,17 @@ SimpleTensor<T> softmax_layer_generic(const SimpleTensor<T> &src, float beta, si
     return dst;
 }
 
-template SimpleTensor<float> softmax_layer_generic(const SimpleTensor<float> &src, float beta, size_t axis, bool is_log);
-template SimpleTensor<half> softmax_layer_generic(const SimpleTensor<half> &src, float beta, size_t axis, bool is_log);
+template SimpleTensor<float> softmax_layer_generic(const SimpleTensor<float> &src, float beta, int32_t axis, bool is_log);
+template SimpleTensor<half> softmax_layer_generic(const SimpleTensor<half> &src, float beta, int32_t axis, bool is_log);
 
 template <typename T, typename std::enable_if<is_floating_point<T>::value, int>::type>
-SimpleTensor<T> softmax_layer(const SimpleTensor<T> &src, float beta, size_t axis)
+SimpleTensor<T> softmax_layer(const SimpleTensor<T> &src, float beta, int32_t axis)
 {
     return softmax_layer_generic<T>(src, beta, axis, false);
 }
 
 template < typename T, typename std::enable_if < std::is_same<T, uint8_t>::value || std::is_same<T, int8_t>::value, int >::type >
-SimpleTensor<T> softmax_layer(const SimpleTensor<T> &src, float beta, size_t axis)
+SimpleTensor<T> softmax_layer(const SimpleTensor<T> &src, float beta, int32_t axis)
 {
     const QuantizationInfo output_quantization_info = arm_compute::get_softmax_output_quantization_info(src.data_type(), false);
 
@@ -121,10 +127,10 @@ SimpleTensor<T> softmax_layer(const SimpleTensor<T> &src, float beta, size_t axi
     return dst;
 }
 
-template SimpleTensor<float> softmax_layer(const SimpleTensor<float> &src, float beta, size_t axis);
-template SimpleTensor<half> softmax_layer(const SimpleTensor<half> &src, float beta, size_t axis);
-template SimpleTensor<uint8_t> softmax_layer(const SimpleTensor<uint8_t> &src, float beta, size_t axis);
-template SimpleTensor<int8_t> softmax_layer(const SimpleTensor<int8_t> &src, float beta, size_t axis);
+template SimpleTensor<float> softmax_layer(const SimpleTensor<float> &src, float beta, int32_t axis);
+template SimpleTensor<half> softmax_layer(const SimpleTensor<half> &src, float beta, int32_t axis);
+template SimpleTensor<uint8_t> softmax_layer(const SimpleTensor<uint8_t> &src, float beta, int32_t axis);
+template SimpleTensor<int8_t> softmax_layer(const SimpleTensor<int8_t> &src, float beta, int32_t axis);
 } // namespace reference
 } // namespace validation
 } // namespace test
diff --git a/tests/validation/reference/SoftmaxLayer.h b/tests/validation/reference/SoftmaxLayer.h
index 2be575c2a..f819853d9 100644
--- a/tests/validation/reference/SoftmaxLayer.h
+++ b/tests/validation/reference/SoftmaxLayer.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2019 ARM Limited.
+ * Copyright (c) 2017-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -36,13 +36,13 @@ namespace validation
 namespace reference
 {
 template <typename T, typename std::enable_if<is_floating_point<T>::value, int>::type = 0>
-SimpleTensor<T> softmax_layer_generic(const SimpleTensor<T> &src, float beta, size_t axis, bool is_log = false);
+SimpleTensor<T> softmax_layer_generic(const SimpleTensor<T> &src, float beta, int32_t axis, bool is_log = false);
 
 template <typename T, typename std::enable_if<is_floating_point<T>::value, int>::type = 0>
-SimpleTensor<T> softmax_layer(const SimpleTensor<T> &src, float beta, size_t axis = 1);
+SimpleTensor<T> softmax_layer(const SimpleTensor<T> &src, float beta, int32_t axis = -1);
 
-template <typename T, typename std::enable_if<std::is_same<T, uint8_t>::value || std::is_same<T, int8_t>::value, int>::type = 0>
-SimpleTensor<T> softmax_layer(const SimpleTensor<T> &src, float beta, size_t axis = 1);
+template < typename T, typename std::enable_if < std::is_same<T, uint8_t>::value || std::is_same<T, int8_t>::value, int >::type = 0 >
+SimpleTensor<T> softmax_layer(const SimpleTensor<T> &src, float beta, int32_t axis = -1);
 } // namespace reference
 } // namespace validation
 } // namespace test