Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1 | /* |
Sheri Zhang | 1f567af | 2020-05-05 11:47:36 +0100 | [diff] [blame] | 2 | * Copyright (c) 2017-2020 ARM Limited. |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
Michalis Spyrou | f464337 | 2019-11-29 16:17:13 +0000 | [diff] [blame] | 24 | #ifndef ARM_COMPUTE_NESOFTMAXLAYER_H |
| 25 | #define ARM_COMPUTE_NESOFTMAXLAYER_H |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 26 | |
| 27 | #include "arm_compute/core/NEON/kernels/NEFillBorderKernel.h" |
Manuel Bottini | 678d83a | 2019-01-07 16:05:36 +0000 | [diff] [blame] | 28 | #include "arm_compute/core/NEON/kernels/NEFlattenLayerKernel.h" |
| 29 | #include "arm_compute/core/NEON/kernels/NEReshapeLayerKernel.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 30 | #include "arm_compute/core/NEON/kernels/NESoftmaxLayerKernel.h" |
| 31 | #include "arm_compute/runtime/IFunction.h" |
Georgios Pinitas | baf174e | 2017-09-08 19:47:30 +0100 | [diff] [blame] | 32 | #include "arm_compute/runtime/MemoryGroup.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 33 | #include "arm_compute/runtime/Tensor.h" |
| 34 | |
| 35 | namespace arm_compute |
| 36 | { |
| 37 | class ITensor; |
| 38 | |
Sang-Hoon Park | d24affe | 2019-10-08 18:07:23 +0100 | [diff] [blame] | 39 | /** Basic function to compute a SoftmaxLayer and a Log SoftmaxLayer. |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 40 | * |
| 41 | * Softmax is calculated by : |
| 42 | * @f[ out = \frac{e^{x - max(x)}}{\sum{e^{x - max(x)}}} @f] |
| 43 | * |
Sang-Hoon Park | d24affe | 2019-10-08 18:07:23 +0100 | [diff] [blame] | 44 | * Log Softmax is calculated by : |
| 45 | * @f[ out = (x - max(x)) - \sum{e^{x - max(x)}} @f] |
| 46 | * |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 47 | * This function runs the following kernels: |
Diego Lopez Recas | 35ceeb2 | 2017-12-04 18:56:10 +0000 | [diff] [blame] | 48 | * -# @ref NEFillBorderKernel |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 49 | * -# @ref NELogits1DMaxKernel |
Diego Lopez Recas | 35ceeb2 | 2017-12-04 18:56:10 +0000 | [diff] [blame] | 50 | * -# @ref NELogits1DSoftmaxKernel |
SiCong Li | d004a7a | 2020-05-28 15:26:41 +0100 | [diff] [blame] | 51 | * And if the reduce_end_axis is not 0 or -input_num_dimensions, the function will use one of the the following kernels |
| 52 | * to reshape the input and perform softmax on the reshaped input: |
| 53 | * -# @ref NEFlattenLayerKernel |
| 54 | * -# @ref NEReshapeLayerKernel |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 55 | */ |
Sang-Hoon Park | d24affe | 2019-10-08 18:07:23 +0100 | [diff] [blame] | 56 | template <bool IS_LOG = false> |
| 57 | class NESoftmaxLayerGeneric : public IFunction |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 58 | { |
| 59 | public: |
| 60 | /** Constructor */ |
Sang-Hoon Park | d24affe | 2019-10-08 18:07:23 +0100 | [diff] [blame] | 61 | NESoftmaxLayerGeneric(std::shared_ptr<IMemoryManager> memory_manager = nullptr); |
Manuel Bottini | 678d83a | 2019-01-07 16:05:36 +0000 | [diff] [blame] | 62 | /** Prevent instances of this class from being copied (As this class contains pointers) */ |
Sang-Hoon Park | d24affe | 2019-10-08 18:07:23 +0100 | [diff] [blame] | 63 | NESoftmaxLayerGeneric(const NESoftmaxLayerGeneric &) = delete; |
Manuel Bottini | 678d83a | 2019-01-07 16:05:36 +0000 | [diff] [blame] | 64 | /** Default move constructor */ |
Sang-Hoon Park | d24affe | 2019-10-08 18:07:23 +0100 | [diff] [blame] | 65 | NESoftmaxLayerGeneric(NESoftmaxLayerGeneric &&) = default; |
Manuel Bottini | 678d83a | 2019-01-07 16:05:36 +0000 | [diff] [blame] | 66 | /** Prevent instances of this class from being copied (As this class contains pointers) */ |
Sang-Hoon Park | d24affe | 2019-10-08 18:07:23 +0100 | [diff] [blame] | 67 | NESoftmaxLayerGeneric &operator=(const NESoftmaxLayerGeneric &) = delete; |
Manuel Bottini | 678d83a | 2019-01-07 16:05:36 +0000 | [diff] [blame] | 68 | /** Default move assignment operator */ |
Sang-Hoon Park | d24affe | 2019-10-08 18:07:23 +0100 | [diff] [blame] | 69 | NESoftmaxLayerGeneric &operator=(NESoftmaxLayerGeneric &&) = default; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 70 | /** Set the input and output tensors. |
| 71 | * |
SiCong Li | d004a7a | 2020-05-28 15:26:41 +0100 | [diff] [blame] | 72 | * @param[in,out] input Source tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. If the width is not a |
| 73 | * multiple of the internal processing block size, @ref NEFillBorderKernel replicates the |
| 74 | * last value of each row to the nearest multiple. |
| 75 | * @param[out] output Destination tensor. Data types supported: same as @p input. |
| 76 | * @param[in] beta (Optional) A scaling factor for the exponent. |
| 77 | * @param[in] reduce_end_axis (Optional) The last axis of the first n dimensions (inclusive)to reduce. Defaults to 0. |
| 78 | * It has the purpose of squashing together the first n dimensions till (including) the @p reduce_end_axis. For instance, given a [2x3x4x5] image, |
| 79 | * when @p reduce_end_axis is 1, the reduction will be applied to axes 0 and 1, and the Softmax op will be applied on each of the [2x3] planes of the input image. |
| 80 | * Negative index is used to specify axis from the end (e.g. -1 for the last axis). |
| 81 | * Must be in range [-input_num_dimensions, input_num_dimensions). |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 82 | */ |
SiCong Li | d004a7a | 2020-05-28 15:26:41 +0100 | [diff] [blame] | 83 | void configure(ITensor *input, ITensor *output, float beta = 1.0f, int32_t reduce_end_axis = 0); |
Michalis Spyrou | afa5d81 | 2017-11-30 14:25:57 +0000 | [diff] [blame] | 84 | /** Static function to check if given info will lead to a valid configuration of @ref NESoftmaxLayer |
| 85 | * |
SiCong Li | d004a7a | 2020-05-28 15:26:41 +0100 | [diff] [blame] | 86 | * @param[in] input Source tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. |
| 87 | * @param[in] output Destination tensor info. Data types supported: same as @p input |
| 88 | * @param[in] beta (Optional) A scaling factor for the exponent. |
| 89 | * @param[in] reduce_end_axis (Optional) The last axis of the first n dimensions (inclusive)to reduce. Defaults to 0. |
| 90 | * It has the purpose of squashing together the first n dimensions till (including) the @p reduce_end_axis. For instance, given a [2x3x4x5] image, |
| 91 | * when @p reduce_end_axis is 1, the reduction will be applied to axes 0 and 1, and the Softmax op will be applied on each of the [2x3] planes of the input image. |
| 92 | * Negative index is used to specify axis from the end (e.g. -1 for the last axis). |
| 93 | * Must be in range [-input_num_dimensions, input_num_dimensions). |
Michalis Spyrou | afa5d81 | 2017-11-30 14:25:57 +0000 | [diff] [blame] | 94 | * @return a status |
| 95 | */ |
SiCong Li | d004a7a | 2020-05-28 15:26:41 +0100 | [diff] [blame] | 96 | static Status validate(const ITensorInfo *input, const ITensorInfo *output, float beta = 1.0f, int32_t reduce_end_axis = 0); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 97 | |
| 98 | // Inherited methods overridden: |
| 99 | void run() override; |
| 100 | |
| 101 | private: |
Manuel Bottini | 678d83a | 2019-01-07 16:05:36 +0000 | [diff] [blame] | 102 | /** Utility method to configure the kernels needed to flatten the input |
| 103 | * tensor. |
| 104 | * |
| 105 | * @note This function changes the internal state of this class. In particular, |
| 106 | * it initializes the kernel @p _flatten_kernel and the tensors @p _input_flat and |
| 107 | * @p _output_flat |
| 108 | * |
SiCong Li | d004a7a | 2020-05-28 15:26:41 +0100 | [diff] [blame] | 109 | * @param[in] input Original source tensor. |
| 110 | * @param[in] output Original destination tensor. |
| 111 | * @param[in] reduce_end_axis (Optional) The last axis of the first n dimensions (inclusive)to reduce. Defaults to 0. |
| 112 | * It has the purpose of squashing together the first n dimensions till (including) the @p reduce_end_axis. For instance, given a [2x3x4x5] image, |
| 113 | * when @p reduce_end_axis is 1, the reduction will be applied to axes 0 and 1, and the Softmax op will be applied on each of the [2x3] planes of the input image. |
| 114 | * Negative index is used to specify axis from the end (e.g. -1 for the last axis). |
| 115 | * Must be in range [-input_num_dimensions, input_num_dimensions). |
Manuel Bottini | 678d83a | 2019-01-07 16:05:36 +0000 | [diff] [blame] | 116 | */ |
SiCong Li | d004a7a | 2020-05-28 15:26:41 +0100 | [diff] [blame] | 117 | void configure_reshape_input_kernel(const ITensor *input, const ITensor *output, int32_t reduce_end_axis); |
Manuel Bottini | 678d83a | 2019-01-07 16:05:36 +0000 | [diff] [blame] | 118 | |
Sang-Hoon Park | d24affe | 2019-10-08 18:07:23 +0100 | [diff] [blame] | 119 | MemoryGroup _memory_group; |
| 120 | NELogits1DMaxKernel _max_kernel; |
| 121 | NELogits1DSoftmaxKernel<IS_LOG> _softmax_kernel; |
| 122 | std::unique_ptr<INEKernel> _flat_or_reshape_kernel_ptr; |
| 123 | NEFillBorderKernel _fill_border_kernel; |
| 124 | NEReshapeLayerKernel _reshape_kernel; |
| 125 | Tensor _max; |
| 126 | Tensor _tmp; |
| 127 | Tensor _input_flattened; |
| 128 | Tensor _output_flattened; |
| 129 | bool _needs_flattening; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 130 | }; |
Sang-Hoon Park | d24affe | 2019-10-08 18:07:23 +0100 | [diff] [blame] | 131 | |
| 132 | using NESoftmaxLayer = NESoftmaxLayerGeneric<false>; |
| 133 | using NELogSoftmaxLayer = NESoftmaxLayerGeneric<true>; |
| 134 | |
Manuel Bottini | 678d83a | 2019-01-07 16:05:36 +0000 | [diff] [blame] | 135 | } // namespace arm_compute |
Michalis Spyrou | f464337 | 2019-11-29 16:17:13 +0000 | [diff] [blame] | 136 | #endif /* ARM_COMPUTE_NESOFTMAXLAYER_H */ |