Giorgio Arena | 657bdb3 | 2018-04-26 18:52:01 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2018 ARM Limited. |
| 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 | */ |
| 24 | #ifndef __ARM_COMPUTE_NECONVERTFULLYCONNECTEDWEIGHTSKERNEL_H__ |
| 25 | #define __ARM_COMPUTE_NECONVERTFULLYCONNECTEDWEIGHTSKERNEL_H__ |
| 26 | |
| 27 | #include "arm_compute/core/NEON/INEKernel.h" |
| 28 | |
| 29 | namespace arm_compute |
| 30 | { |
| 31 | class ITensor; |
| 32 | |
| 33 | /** Interface to convert the 2D Fully Connected weights from NCHW to NHWC or vice versa. |
| 34 | * |
| 35 | * @note This function can be applied to the 2D weights used by a Fully Connected layer if: |
| 36 | * - It follows a Convolution layer |
| 37 | * - The data layout used by the network does not match the one the model has been trained in. |
| 38 | * |
| 39 | * @note This function assumes the weights are already reshaped (transposed) |
| 40 | */ |
| 41 | class NEConvertFullyConnectedWeightsKernel : public INEKernel |
| 42 | { |
| 43 | public: |
| 44 | const char *name() const override |
| 45 | { |
| 46 | return "NEConvertFullyConnectedWeightsKernel"; |
| 47 | } |
| 48 | /** Default constructor */ |
| 49 | NEConvertFullyConnectedWeightsKernel(); |
| 50 | /** Prevent instances of this class from being copied (As this class contains pointers) */ |
| 51 | NEConvertFullyConnectedWeightsKernel(const NEConvertFullyConnectedWeightsKernel &) = delete; |
| 52 | /** Prevent instances of this class from being copied (As this class contains pointers) */ |
| 53 | NEConvertFullyConnectedWeightsKernel &operator=(const NEConvertFullyConnectedWeightsKernel &) = delete; |
| 54 | /** Allow instances of this class to be moved */ |
| 55 | NEConvertFullyConnectedWeightsKernel(NEConvertFullyConnectedWeightsKernel &&) = default; |
| 56 | /** Allow instances of this class to be moved */ |
| 57 | NEConvertFullyConnectedWeightsKernel &operator=(NEConvertFullyConnectedWeightsKernel &&) = default; |
| 58 | /** Default destructor */ |
| 59 | ~NEConvertFullyConnectedWeightsKernel() = default; |
| 60 | /** Set the input and output tensor. |
| 61 | * |
Vidhya Sudhan Loganathan | f4cb81b | 2018-07-04 15:13:14 +0100 | [diff] [blame] | 62 | * @param[in] input Source weights tensor to convert. Must be 2 dimensional. Data types supported: U8/S8/QASYMM8/U16/S16/U32/S32/F16/F32. |
Giorgio Arena | 657bdb3 | 2018-04-26 18:52:01 +0100 | [diff] [blame] | 63 | * @param[out] output The converted weights tensor. Shape and Data Type: Same as @p input. |
Georgios Pinitas | 7d66a8e | 2018-07-17 12:28:42 +0100 | [diff] [blame] | 64 | * @param[in] original_input_shape Shape of the original input tensor (the one entering fully connected layer). |
Giorgio Arena | 657bdb3 | 2018-04-26 18:52:01 +0100 | [diff] [blame] | 65 | * @param[in] data_layout The data layout the weights have been trained in. |
| 66 | */ |
| 67 | void configure(const ITensor *input, ITensor *output, const TensorShape &original_input_shape, DataLayout data_layout); |
| 68 | /** Static function to check if given info will lead to a valid configuration of @ref NEConvertFullyConnectedWeightsKernel |
| 69 | * |
Vidhya Sudhan Loganathan | f4cb81b | 2018-07-04 15:13:14 +0100 | [diff] [blame] | 70 | * @param[in] input Source weights tensor info to convert. Must be 2 dimensional. Data types supported: U8/S8/QASYMM8/U16/S16/U32/S32/F16/F32. |
Giorgio Arena | 657bdb3 | 2018-04-26 18:52:01 +0100 | [diff] [blame] | 71 | * @param[in] output The converted weights tensor info. Shape and Data Type: Same as @p input. |
Georgios Pinitas | 7d66a8e | 2018-07-17 12:28:42 +0100 | [diff] [blame] | 72 | * @param[in] original_input_shape Shape of the original input tensor (the one entering fully connected layer). |
Giorgio Arena | 657bdb3 | 2018-04-26 18:52:01 +0100 | [diff] [blame] | 73 | * @param[in] data_layout The data layout the weights have been trained in. |
| 74 | */ |
| 75 | static Status validate(const ITensorInfo *input, const ITensorInfo *output, const TensorShape &original_input_shape, DataLayout data_layout); |
| 76 | |
| 77 | // Inherited methods overridden: |
| 78 | void run(const Window &window, const ThreadInfo &info) override; |
| 79 | |
| 80 | private: |
| 81 | /** Template function to run the permute |
| 82 | * |
| 83 | * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()). |
| 84 | */ |
| 85 | template <typename T> |
| 86 | void run_convert_fc_weights(const Window &window); |
| 87 | |
| 88 | const ITensor *_input; |
| 89 | ITensor *_output; |
| 90 | unsigned int _factor1; /* equals to the number of elements per original input plane if @p data_layout == NCHW; its number of channels otherwise */ |
| 91 | unsigned int _factor2; /* equals to the number of elements per original input plane if @p data_layout == NHWC; its number of channels otherwise */ |
| 92 | }; |
| 93 | } // namespace arm_compute |
| 94 | #endif /*__ARM_COMPUTE_NECONVERTFULLYCONNECTEDWEIGHTSKERNEL_H__ */ |