git.sesse.net Git - ffmpeg/blob - libavfilter/dnn/dnn_backend_native_layer_dense.c

   1 /*
   2  * Copyright (c) 2020
   3  *
   4  * This file is part of FFmpeg.
   5  *
   6  * FFmpeg is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2.1 of the License, or (at your option) any later version.
  10  *
  11  * FFmpeg is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * Lesser General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU Lesser General Public
  17  * License along with FFmpeg; if not, write to the Free Software
  18  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  19  */
  20
  21 #include "libavutil/avassert.h"
  22 #include "dnn_backend_native_layer_dense.h"
  23
  24 int ff_dnn_load_layer_dense(Layer *layer, AVIOContext *model_file_context, int file_size, int operands_num)
  25 {
  26     DenseParams *dense_params;
  27     int kernel_size;
  28     int dnn_size = 0;
  29     dense_params = av_malloc(sizeof(*dense_params));
  30     if (!dense_params)
  31         return 0;
  32
  33     dense_params->activation = (int32_t)avio_rl32(model_file_context);
  34     dense_params->input_num = (int32_t)avio_rl32(model_file_context);
  35     dense_params->output_num = (int32_t)avio_rl32(model_file_context);
  36     dense_params->has_bias = (int32_t)avio_rl32(model_file_context);
  37     dnn_size += 16;
  38
  39     kernel_size = dense_params->input_num * dense_params->output_num;
  40     dnn_size += kernel_size * 4;
  41     if (dense_params->has_bias)
  42         dnn_size += dense_params->output_num * 4;
  43
  44     if (dnn_size > file_size || dense_params->input_num <= 0 ||
  45         dense_params->output_num <= 0){
  46         av_freep(&dense_params);
  47         return 0;
  48     }
  49
  50     dense_params->kernel = av_malloc(kernel_size * sizeof(float));
  51     if (!dense_params->kernel) {
  52         av_freep(&dense_params);
  53         return 0;
  54     }
  55     for (int i = 0; i < kernel_size; ++i) {
  56         dense_params->kernel[i] = av_int2float(avio_rl32(model_file_context));
  57     }
  58
  59     dense_params->biases = NULL;
  60     if (dense_params->has_bias) {
  61         dense_params->biases = av_malloc(dense_params->output_num * sizeof(float));
  62         if (!dense_params->biases){
  63             av_freep(&dense_params->kernel);
  64             av_freep(&dense_params);
  65             return 0;
  66         }
  67         for (int i = 0; i < dense_params->output_num; ++i){
  68             dense_params->biases[i] = av_int2float(avio_rl32(model_file_context));
  69         }
  70     }
  71
  72     layer->params = dense_params;
  73
  74     layer->input_operand_indexes[0] = (int32_t)avio_rl32(model_file_context);
  75     layer->output_operand_index = (int32_t)avio_rl32(model_file_context);
  76     dnn_size += 8;
  77
  78     if (layer->input_operand_indexes[0] >= operands_num || layer->output_operand_index >= operands_num) {
  79         return 0;
  80     }
  81
  82     return dnn_size;
  83 }
  84
  85 int ff_dnn_execute_layer_dense(DnnOperand *operands, const int32_t *input_operand_indexes,
  86                                int32_t output_operand_index, const void *parameters, NativeContext *ctx)
  87 {
  88     float *output;
  89     int32_t input_operand_index = input_operand_indexes[0];
  90     int number = operands[input_operand_index].dims[0];
  91     int height = operands[input_operand_index].dims[1];
  92     int width = operands[input_operand_index].dims[2];
  93     int channel = operands[input_operand_index].dims[3];
  94     const float *input = operands[input_operand_index].data;
  95     const DenseParams *dense_params = parameters;
  96
  97     int src_linesize = width * channel;
  98     DnnOperand *output_operand = &operands[output_operand_index];
  99     output_operand->dims[0] = number;
 100     output_operand->dims[1] = height;
 101     output_operand->dims[2] = width;
 102     output_operand->dims[3] = dense_params->output_num;
 103     output_operand->data_type = operands[input_operand_index].data_type;
 104     output_operand->length = ff_calculate_operand_data_length(output_operand);
 105     if (output_operand->length <= 0) {
 106         av_log(ctx, AV_LOG_ERROR, "The output data length overflow\n");
 107         return DNN_ERROR;
 108     }
 109     output_operand->data = av_realloc(output_operand->data, output_operand->length);
 110     if (!output_operand->data) {
 111         av_log(ctx, AV_LOG_ERROR, "Failed to reallocate memory for output\n");
 112         return DNN_ERROR;
 113     }
 114     output = output_operand->data;
 115
 116     av_assert0(channel == dense_params->input_num);
 117
 118     for (int y = 0; y < height; ++y) {
 119         for (int x = 0; x < width; ++x) {
 120             for (int n_filter = 0; n_filter < dense_params->output_num; ++n_filter) {
 121                 if (dense_params->has_bias)
 122                     output[n_filter] = dense_params->biases[n_filter];
 123                 else
 124                     output[n_filter] = 0.f;
 125
 126                 for (int ch = 0; ch < dense_params->input_num; ++ch) {
 127                     float input_pel;
 128                     input_pel = input[y * src_linesize + x * dense_params->input_num + ch];
 129                     output[n_filter] += input_pel * dense_params->kernel[n_filter*dense_params->input_num + ch];
 130                 }
 131                 switch (dense_params->activation){
 132                 case RELU:
 133                     output[n_filter] = FFMAX(output[n_filter], 0.0);
 134                     break;
 135                 case TANH:
 136                     output[n_filter] = 2.0f  / (1.0f + exp(-2.0f * output[n_filter])) - 1.0f;
 137                     break;
 138                 case SIGMOID:
 139                     output[n_filter] = 1.0f / (1.0f + exp(-output[n_filter]));
 140                     break;
 141                 case NONE:
 142                     break;
 143                 case LEAKY_RELU:
 144                     output[n_filter] = FFMAX(output[n_filter], 0.0) + 0.2 * FFMIN(output[n_filter], 0.0);
 145                 }
 146             }
 147             output += dense_params->output_num;
 148         }
 149     }
 150     return 0;
 151 }