1 /* 2 * Copyright (c) 2018-2020 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 "src/core/NEON/INEKernel.h" 28 29 namespace arm_compute 30 { 31 // Forward declarations 32 class ITensor; 33 34 /** Interface to convert the 2D Fully Connected weights from NCHW to NHWC or vice versa. 35 * 36 * @note This function can be applied to the 2D weights used by a Fully Connected layer if: 37 * - It follows a Convolution layer 38 * - The data layout used by the network does not match the one the model has been trained in. 39 * 40 * @note This function assumes the weights are already reshaped (transposed) 41 */ 42 class NEConvertFullyConnectedWeightsKernel : public INEKernel 43 { 44 public: name()45 const char *name() const override 46 { 47 return "NEConvertFullyConnectedWeightsKernel"; 48 } 49 /** Default constructor */ 50 NEConvertFullyConnectedWeightsKernel(); 51 /** Prevent instances of this class from being copied (As this class contains pointers) */ 52 NEConvertFullyConnectedWeightsKernel(const NEConvertFullyConnectedWeightsKernel &) = delete; 53 /** Prevent instances of this class from being copied (As this class contains pointers) */ 54 NEConvertFullyConnectedWeightsKernel &operator=(const NEConvertFullyConnectedWeightsKernel &) = delete; 55 /** Allow instances of this class to be moved */ 56 NEConvertFullyConnectedWeightsKernel(NEConvertFullyConnectedWeightsKernel &&) = default; 57 /** Allow instances of this class to be moved */ 58 NEConvertFullyConnectedWeightsKernel &operator=(NEConvertFullyConnectedWeightsKernel &&) = default; 59 /** Default destructor */ 60 ~NEConvertFullyConnectedWeightsKernel() = default; 61 /** Set the input and output tensor. 62 * 63 * @param[in] input Source weights tensor to convert. Must be 2 dimensional. Data types supported: All. 64 * @param[out] output The converted weights tensor. Shape and Data Type: Same as @p input. 65 * @param[in] original_input_shape Shape of the original input tensor (the one entering fully connected layer). 66 * @param[in] data_layout The data layout the weights have been trained in. 67 */ 68 void configure(const ITensor *input, ITensor *output, const TensorShape &original_input_shape, DataLayout data_layout); 69 /** Static function to check if given info will lead to a valid configuration of @ref NEConvertFullyConnectedWeightsKernel 70 * 71 * @param[in] input Source weights tensor info to convert. Must be 2 dimensional. Data types supported: All. 72 * @param[in] output The converted weights tensor info. Shape and Data Type: Same as @p input. 73 * @param[in] original_input_shape Shape of the original input tensor (the one entering fully connected layer). 74 * @param[in] data_layout The data layout the weights have been trained in. 75 */ 76 static Status validate(const ITensorInfo *input, const ITensorInfo *output, const TensorShape &original_input_shape, DataLayout data_layout); 77 78 // Inherited methods overridden: 79 void run(const Window &window, const ThreadInfo &info) override; 80 81 private: 82 /** Template function to run the permute 83 * 84 * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()). 85 */ 86 template <typename T> 87 void run_convert_fc_weights(const Window &window); 88 89 const ITensor *_input; 90 ITensor *_output; 91 unsigned int _factor1; /* equals to the number of elements per original input plane if @p data_layout == NCHW; its number of channels otherwise */ 92 unsigned int _factor2; /* equals to the number of elements per original input plane if @p data_layout == NHWC; its number of channels otherwise */ 93 }; 94 } // namespace arm_compute 95 #endif /*ARM_COMPUTE_NECONVERTFULLYCONNECTEDWEIGHTSKERNEL_H */ 96