inference-engine/thirdparty/clDNN/api/CPP/fully_connected.hpp

   1 /*
   2 // Copyright (c) 2016 Intel Corporation
   3 //
   4 // Licensed under the Apache License, Version 2.0 (the "License");
   5 // you may not use this file except in compliance with the License.
   6 // You may obtain a copy of the License at
   7 //
   8 //      http://www.apache.org/licenses/LICENSE-2.0
   9 //
  10 // Unless required by applicable law or agreed to in writing, software
  11 // distributed under the License is distributed on an "AS IS" BASIS,
  12 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  13 // See the License for the specific language governing permissions and
  14 // limitations under the License.
  15 */
  16
  17 ///////////////////////////////////////////////////////////////////////////////////////////////////
  18 #pragma once
  19 #include "../C/fully_connected.h"
  20 #include "primitive.hpp"
  21 #include <vector>
  22
  23 namespace cldnn {
  24 /// @addtogroup cpp_api C++ API
  25 /// @{
  26 /// @addtogroup cpp_topology Network Topology
  27 /// @{
  28 /// @addtogroup cpp_primitives Primitives
  29 /// @{
  30
  31 /// @brief Performs forward fully connected layer (inner product).
  32 /// Also supports built-in Relu @CLDNN_PRIMITIVE_DESC{activation} available by setting it in arguments.
  33 /// @notes
  34 /// - Equation: Input[F x Y x F] x Output(X) == Weights(B x F x X x F) has to be fulfilled
  35 /// - Bias has to be linear data [1,1,1,X], where X is equal to number of outputs.
  36
  37 /// <table>
  38 /// <caption id = "multi_row">Format support</caption>
  39 ///        <tr><th>Data type               <th>activation format       <th>weights format
  40 ///        <tr><td rowspan="7">F32         <td rowspan="4">bfyx        <td>yxfb
  41 ///        <tr>                                                        <td>fyxb
  42 ///        <tr>                                                        <td>bs_xs_xsv8_bsv8
  43 ///        <tr>                                                        <td>bs_x_bsv16
  44 ///        <tr>                            <td rowspan="3">yxfb        <td>bfyx
  45 ///        <tr>                                                        <td>yxfb
  46 ///        <tr>                                                        <td>bs_xs_xsv8_bsv8
  47 ///        <tr><td rowspan="4">F16         <td rowspan="3">bfyx        <td>yxfb
  48 ///        <tr>                                                        <td>fyxb
  49 ///        <tr>                                                        <td>bs_x_bsv16
  50 ///        <tr>                            <td >yxfb                   <td>bfyx
  51 /// </table>
  52
  53 struct fully_connected : public primitive_base<fully_connected, CLDNN_PRIMITIVE_DESC(fully_connected)> {
  54     CLDNN_DECLARE_PRIMITIVE(fully_connected)
  55
  56     /// @brief Constructs fully connected layer.
  57     /// @param id This primitive id.
  58     /// @param input Input primitive id.
  59     /// @param weights Primitive id containing weights data.
  60     /// @param bias Primitive id containing bias data. Provide empty string if using Relu without bias.
  61     /// @param with_activation Enable Relu activation.
  62     /// @param activation_slp Relu activation slope.
  63     fully_connected(const primitive_id& id,
  64                     const primitive_id& input,
  65                     const primitive_id& weights,
  66                     const primitive_id& bias = "",
  67                     bool with_activation = false,
  68                     float activation_slp = 0.0f,
  69                     const padding& output_padding = padding())
  70         : primitive_base(id, {input}, output_padding),
  71           weights(weights),
  72           bias(bias),
  73           weights_quantization_factors(""),
  74           output_calibration_factors(""),
  75           input_quantization_factor(1.0f),
  76           output_quantization_factor(1.0f),
  77           with_activation(with_activation),
  78           activation_negative_slope(activation_slp) {}
  79
  80     /// @brief Constructs fully connected layer.
  81     /// @param id This primitive id.
  82     /// @param input Input primitive id.
  83     /// @param weights Primitive id containing weights data.
  84     /// @param bias Primitive id containing bias data. Provide empty string if using Relu without bias.
  85     /// @param w_quantization_factor Primitive id containing weights quanitization factors per output feature map.
  86     /// @param i_quantization_factor Input quantization factor
  87     /// @param o_quantization_factor Output quantization factor
  88     /// @param with_activation Enable Relu activation.
  89     /// @param activation_slp Relu activation slope.
  90     fully_connected(const primitive_id& id,
  91                     const primitive_id& input,
  92                     const primitive_id& weights,
  93                     const primitive_id& bias,
  94                     const primitive_id& w_quantization_factor,
  95                     const float i_quantization_factor,
  96                     const float o_quantization_factor,
  97                     bool with_activation = false,
  98                     float activation_slp = 0.0f,
  99                     const padding& output_padding = padding())
 100         : primitive_base(id, {input}, output_padding),
 101           weights(weights),
 102           bias(bias),
 103           weights_quantization_factors(w_quantization_factor),
 104           output_calibration_factors(""),
 105           input_quantization_factor(i_quantization_factor),
 106           output_quantization_factor(o_quantization_factor),
 107           with_activation(with_activation),
 108           activation_negative_slope(activation_slp) {}
 109
 110     /// @brief Constructs fully connected layer.
 111     /// @param id This primitive id.
 112     /// @param input Input primitive id.
 113     /// @param weights Primitive id containing weights data.
 114     /// @param bias Primitive id containing bias data. Provide empty string if using Relu without bias.
 115     /// @param w_quantization_factor Primitive id containing weights quanitization factors per output feature map.
 116     /// @param output_calibration_factors Primitive id containing output calibration factors per output feature map.
 117     /// @param i_quantization_factor Input quantization factor
 118     /// @param with_activation Enable Relu activation.
 119     /// @param activation_slp Relu activation slope.
 120     fully_connected(const primitive_id& id,
 121                     const primitive_id& input,
 122                     const primitive_id& weights,
 123                     const primitive_id& bias,
 124                     const primitive_id& w_quantization_factor,
 125                     const primitive_id& output_calibration_factors,
 126                     const float i_quantization_factor,
 127                     bool with_activation = false,
 128                     float activation_slp = 0.0f,
 129                     const padding& output_padding = padding())
 130         : primitive_base(id, {input}, output_padding),
 131           weights(weights),
 132           bias(bias),
 133           weights_quantization_factors(w_quantization_factor),
 134           output_calibration_factors(output_calibration_factors),
 135           input_quantization_factor(i_quantization_factor),
 136           output_quantization_factor(1.0f),
 137           with_activation(with_activation),
 138           activation_negative_slope(activation_slp) {}
 139
 140     /// @brief Constructs a copy from basic C API @CLDNN_PRIMITIVE_DESC{fully_connected}
 141     fully_connected(const dto* dto)
 142         : primitive_base(dto),
 143           weights(dto->weights),
 144           bias(dto->bias),
 145           weights_quantization_factors(dto->weights_quantization_factors),
 146           output_calibration_factors(dto->output_calibration_factors),
 147           input_quantization_factor(dto->input_quantization_factor),
 148           output_quantization_factor(dto->output_quantization_factor),
 149           with_activation(dto->with_activation != 0),
 150           activation_negative_slope(dto->activation_negative_slope) {}
 151
 152     /// @brief Primitive id containing weights data.
 153     primitive_id weights;
 154     /// @brief Primitive id containing bias data.
 155     primitive_id bias;
 156     /// @brief Primitive id containing weights quanitization factors per output feature map.
 157     primitive_id weights_quantization_factors;
 158     /// @brief Primitive id containing output quanitization factors per output feature map.
 159     primitive_id output_calibration_factors;
 160     /// @brief Input quantization factor
 161     float input_quantization_factor;
 162     /// @brief Output quantization factor
 163     float output_quantization_factor;
 164     /// @brief Enable Relu activation.
 165     bool with_activation;
 166     /// @brief Relu activation slope.
 167     float activation_negative_slope;
 168
 169 protected:
 170     std::vector<std::reference_wrapper<const primitive_id>> get_dependencies() const override {
 171         std::vector<std::reference_wrapper<const primitive_id>> ret;
 172         ret.push_back(weights);
 173
 174         if (!bias.empty())
 175             ret.push_back(bias);
 176
 177         if (!weights_quantization_factors.empty())
 178             ret.push_back(weights_quantization_factors);
 179
 180         if (!output_calibration_factors.empty())
 181             ret.push_back(output_calibration_factors);
 182
 183         return ret;
 184     }
 185
 186     void update_dto(dto& dto) const override {
 187         dto.weights = weights.c_str();
 188         dto.bias = bias.c_str();
 189         dto.weights_quantization_factors = weights_quantization_factors.c_str();
 190         dto.output_calibration_factors = output_calibration_factors.c_str();
 191         dto.input_quantization_factor = input_quantization_factor;
 192         dto.output_quantization_factor = output_quantization_factor;
 193         dto.with_activation = with_activation;
 194         dto.activation_negative_slope = activation_negative_slope;
 195     }
 196 };
 197 /// @}
 198 /// @}
 199 /// @}
 200 }  // namespace cldnn