git.sesse.net Git - ffmpeg/blob - libavfilter/dnn_backend_native.c

   1 /*
   2  * Copyright (c) 2018 Sergey Lavrushkin
   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 /**
  22  * @file
  23  * DNN native backend implementation.
  24  */
  25
  26 #include "dnn_backend_native.h"
  27 #include "dnn_srcnn.h"
  28 #include "libavformat/avio.h"
  29
  30 typedef enum {INPUT, CONV} LayerType;
  31
  32 typedef struct Layer{
  33     LayerType type;
  34     float* output;
  35     void* params;
  36 } Layer;
  37
  38 typedef struct ConvolutionalParams{
  39     int32_t input_num, output_num, kernel_size;
  40     float* kernel;
  41     float* biases;
  42 } ConvolutionalParams;
  43
  44 typedef struct InputParams{
  45     int height, width, channels;
  46 } InputParams;
  47
  48 // Represents simple feed-forward convolutional network.
  49 typedef struct ConvolutionalNetwork{
  50     Layer* layers;
  51     int32_t layers_num;
  52 } ConvolutionalNetwork;
  53
  54 static DNNReturnType set_input_output_native(void* model, const DNNData* input, const DNNData* output)
  55 {
  56     ConvolutionalNetwork* network = (ConvolutionalNetwork*)model;
  57     InputParams* input_params;
  58     ConvolutionalParams* conv_params;
  59     int cur_width, cur_height, cur_channels;
  60     int32_t layer;
  61
  62     if (network->layers_num <= 0 || network->layers[0].type != INPUT){
  63         return DNN_ERROR;
  64     }
  65     else{
  66         network->layers[0].output = input->data;
  67         input_params = (InputParams*)network->layers[0].params;
  68         input_params->width = cur_width = input->width;
  69         input_params->height = cur_height = input->height;
  70         input_params->channels = cur_channels = input->channels;
  71     }
  72
  73     for (layer = 1; layer < network->layers_num; ++layer){
  74         switch (network->layers[layer].type){
  75         case CONV:
  76             conv_params = (ConvolutionalParams*)network->layers[layer].params;
  77             if (conv_params->input_num != cur_channels){
  78                 return DNN_ERROR;
  79             }
  80             cur_channels = conv_params->output_num;
  81             if (layer < network->layers_num - 1){
  82                 if (!network->layers[layer].output){
  83                     av_freep(&network->layers[layer].output);
  84                 }
  85                 network->layers[layer].output = av_malloc(cur_height * cur_width * cur_channels * sizeof(float));
  86                 if (!network->layers[layer].output){
  87                     return DNN_ERROR;
  88                 }
  89             }
  90             else{
  91                 network->layers[layer].output = output->data;
  92                 if (output->width != cur_width || output->height != cur_height || output->channels != cur_channels){
  93                     return DNN_ERROR;
  94                 }
  95             }
  96             break;
  97         default:
  98             return DNN_ERROR;
  99         }
 100     }
 101
 102     return DNN_SUCCESS;
 103 }
 104
 105 // Loads model and its parameters that are stored in a binary file with following structure:
 106 // layers_num,conv_input_num,conv_output_num,conv_kernel_size,conv_kernel,conv_biases,conv_input_num...
 107 DNNModel* ff_dnn_load_model_native(const char* model_filename)
 108 {
 109     DNNModel* model = NULL;
 110     ConvolutionalNetwork* network = NULL;
 111     AVIOContext* model_file_context;
 112     int file_size, dnn_size, kernel_size, i;
 113     int32_t layer;
 114     ConvolutionalParams* conv_params;
 115
 116     model = av_malloc(sizeof(DNNModel));
 117     if (!model){
 118         return NULL;
 119     }
 120
 121     if (avio_open(&model_file_context, model_filename, AVIO_FLAG_READ) < 0){
 122         av_freep(&model);
 123         return NULL;
 124     }
 125     file_size = avio_size(model_file_context);
 126
 127     network = av_malloc(sizeof(ConvolutionalNetwork));
 128     if (!network){
 129         avio_closep(&model_file_context);
 130         av_freep(&model);
 131         return NULL;
 132     }
 133     model->model = (void*)network;
 134
 135     network->layers_num = 1 + (int32_t)avio_rl32(model_file_context);
 136     dnn_size = 4;
 137
 138     network->layers = av_malloc(network->layers_num * sizeof(Layer));
 139     if (!network->layers){
 140         av_freep(&network);
 141         avio_closep(&model_file_context);
 142         av_freep(&model);
 143         return NULL;
 144     }
 145
 146     for (layer = 0; layer < network->layers_num; ++layer){
 147         network->layers[layer].output = NULL;
 148         network->layers[layer].params = NULL;
 149     }
 150     network->layers[0].type = INPUT;
 151     network->layers[0].params = av_malloc(sizeof(InputParams));
 152     if (!network->layers[0].params){
 153         avio_closep(&model_file_context);
 154         ff_dnn_free_model_native(&model);
 155         return NULL;
 156     }
 157
 158     for (layer = 1; layer < network->layers_num; ++layer){
 159         conv_params = av_malloc(sizeof(ConvolutionalParams));
 160         if (!conv_params){
 161             avio_closep(&model_file_context);
 162             ff_dnn_free_model_native(&model);
 163             return NULL;
 164         }
 165         conv_params->input_num = (int32_t)avio_rl32(model_file_context);
 166         conv_params->output_num = (int32_t)avio_rl32(model_file_context);
 167         conv_params->kernel_size = (int32_t)avio_rl32(model_file_context);
 168         kernel_size = conv_params->input_num * conv_params->output_num *
 169                       conv_params->kernel_size * conv_params->kernel_size;
 170         dnn_size += 12 + (kernel_size + conv_params->output_num << 2);
 171         if (dnn_size > file_size || conv_params->input_num <= 0 ||
 172             conv_params->output_num <= 0 || conv_params->kernel_size <= 0){
 173             avio_closep(&model_file_context);
 174             ff_dnn_free_model_native(&model);
 175             return NULL;
 176         }
 177         conv_params->kernel = av_malloc(kernel_size * sizeof(float));
 178         conv_params->biases = av_malloc(conv_params->output_num * sizeof(float));
 179         if (!conv_params->kernel || !conv_params->biases){
 180             avio_closep(&model_file_context);
 181             ff_dnn_free_model_native(&model);
 182             return NULL;
 183         }
 184         for (i = 0; i < kernel_size; ++i){
 185             conv_params->kernel[i] = av_int2float(avio_rl32(model_file_context));
 186         }
 187         for (i = 0; i < conv_params->output_num; ++i){
 188             conv_params->biases[i] = av_int2float(avio_rl32(model_file_context));
 189         }
 190         network->layers[layer].type = CONV;
 191         network->layers[layer].params = conv_params;
 192     }
 193
 194     avio_closep(&model_file_context);
 195
 196     if (dnn_size != file_size){
 197         ff_dnn_free_model_native(&model);
 198         return NULL;
 199     }
 200
 201     model->set_input_output = &set_input_output_native;
 202
 203     return model;
 204 }
 205
 206 static int set_up_conv_layer(Layer* layer, const float* kernel, const float* biases, int32_t input_num, int32_t output_num, int32_t size)
 207 {
 208     ConvolutionalParams* conv_params;
 209     int kernel_size;
 210
 211     conv_params = av_malloc(sizeof(ConvolutionalParams));
 212     if (!conv_params){
 213         return DNN_ERROR;
 214     }
 215     conv_params->input_num = input_num;
 216     conv_params->output_num = output_num;
 217     conv_params->kernel_size = size;
 218     kernel_size = input_num * output_num * size * size;
 219     conv_params->kernel = av_malloc(kernel_size * sizeof(float));
 220     conv_params->biases = av_malloc(conv_params->output_num * sizeof(float));
 221     if (!conv_params->kernel || !conv_params->biases){
 222         av_freep(&conv_params->kernel);
 223         av_freep(&conv_params->biases);
 224         av_freep(&conv_params);
 225         return DNN_ERROR;
 226     }
 227     memcpy(conv_params->kernel, kernel, kernel_size * sizeof(float));
 228     memcpy(conv_params->biases, biases, output_num * sizeof(float));
 229     layer->type = CONV;
 230     layer->params = conv_params;
 231
 232     return DNN_SUCCESS;
 233 }
 234
 235 DNNModel* ff_dnn_load_default_model_native(DNNDefaultModel model_type)
 236 {
 237     DNNModel* model = NULL;
 238     ConvolutionalNetwork* network = NULL;
 239     int32_t layer;
 240
 241     model = av_malloc(sizeof(DNNModel));
 242     if (!model){
 243         return NULL;
 244     }
 245
 246     network = av_malloc(sizeof(ConvolutionalNetwork));
 247     if (!network){
 248         av_freep(&model);
 249         return NULL;
 250     }
 251     model->model = (void*)network;
 252
 253     switch (model_type){
 254     case DNN_SRCNN:
 255         network->layers_num = 4;
 256
 257         network->layers = av_malloc(network->layers_num * sizeof(Layer));
 258         if (!network->layers){
 259             av_freep(&network);
 260             av_freep(&model);
 261             return NULL;
 262         }
 263
 264         for (layer = 0; layer < network->layers_num; ++layer){
 265             network->layers[layer].output = NULL;
 266             network->layers[layer].params = NULL;
 267         }
 268         network->layers[0].type = INPUT;
 269         network->layers[0].params = av_malloc(sizeof(InputParams));
 270         if (!network->layers[0].params){
 271             ff_dnn_free_model_native(&model);
 272             return NULL;
 273         }
 274
 275         if (set_up_conv_layer(network->layers + 1, conv1_kernel, conv1_biases, 1, 64, 9) != DNN_SUCCESS ||
 276             set_up_conv_layer(network->layers + 2, conv2_kernel, conv2_biases, 64, 32, 1) != DNN_SUCCESS ||
 277             set_up_conv_layer(network->layers + 3, conv3_kernel, conv3_biases, 32, 1, 5) != DNN_SUCCESS){
 278             ff_dnn_free_model_native(&model);
 279             return NULL;
 280         }
 281
 282         model->set_input_output = &set_input_output_native;
 283
 284         return model;
 285     default:
 286         av_freep(&network);
 287         av_freep(&model);
 288         return NULL;
 289     }
 290 }
 291
 292 #define CLAMP_TO_EDGE(x, w) ((x) < 0 ? 0 : ((x) >= (w) ? (w - 1) : (x)))
 293
 294 static void convolve(const float* input, float* output, const ConvolutionalParams* conv_params, int32_t width, int32_t height)
 295 {
 296     int y, x, n_filter, ch, kernel_y, kernel_x;
 297     int radius = conv_params->kernel_size >> 1;
 298     int src_linesize = width * conv_params->input_num;
 299     int filter_linesize = conv_params->kernel_size * conv_params->input_num;
 300     int filter_size = conv_params->kernel_size * filter_linesize;
 301
 302     for (y = 0; y < height; ++y){
 303         for (x = 0; x < width; ++x){
 304             for (n_filter = 0; n_filter < conv_params->output_num; ++n_filter){
 305                 output[n_filter] = conv_params->biases[n_filter];
 306                 for (ch = 0; ch < conv_params->input_num; ++ch){
 307                     for (kernel_y = 0; kernel_y < conv_params->kernel_size; ++kernel_y){
 308                         for (kernel_x = 0; kernel_x < conv_params->kernel_size; ++kernel_x){
 309                             output[n_filter] += input[CLAMP_TO_EDGE(y + kernel_y - radius, height) * src_linesize +
 310                                                       CLAMP_TO_EDGE(x + kernel_x - radius, width) * conv_params->input_num + ch] *
 311                                                 conv_params->kernel[n_filter * filter_size + kernel_y * filter_linesize +
 312                                                                     kernel_x * conv_params->input_num + ch];
 313                         }
 314                     }
 315                 }
 316                 output[n_filter] = FFMAX(output[n_filter], 0.0);
 317             }
 318             output += conv_params->output_num;
 319         }
 320     }
 321 }
 322
 323 DNNReturnType ff_dnn_execute_model_native(const DNNModel* model)
 324 {
 325     ConvolutionalNetwork* network = (ConvolutionalNetwork*)model->model;
 326     InputParams* input_params;
 327     int cur_width, cur_height;
 328     int32_t layer;
 329
 330     if (network->layers_num <= 0 || network->layers[0].type != INPUT || !network->layers[0].output){
 331         return DNN_ERROR;
 332     }
 333     else{
 334         input_params = (InputParams*)network->layers[0].params;
 335         cur_width = input_params->width;
 336         cur_height = input_params->height;
 337     }
 338
 339     for (layer = 1; layer < network->layers_num; ++layer){
 340         if (!network->layers[layer].output){
 341             return DNN_ERROR;
 342         }
 343         switch (network->layers[layer].type){
 344         case CONV:
 345             convolve(network->layers[layer - 1].output, network->layers[layer].output, (ConvolutionalParams*)network->layers[layer].params, cur_width, cur_height);
 346             break;
 347         case INPUT:
 348             return DNN_ERROR;
 349         }
 350     }
 351
 352     return DNN_SUCCESS;
 353 }
 354
 355 void ff_dnn_free_model_native(DNNModel** model)
 356 {
 357     ConvolutionalNetwork* network;
 358     ConvolutionalParams* conv_params;
 359     int32_t layer;
 360
 361     if (*model)
 362     {
 363         network = (ConvolutionalNetwork*)(*model)->model;
 364         for (layer = 0; layer < network->layers_num; ++layer){
 365             switch (network->layers[layer].type){
 366             case CONV:
 367                 if (layer < network->layers_num - 1){
 368                     av_freep(&network->layers[layer].output);
 369                 }
 370                 conv_params = (ConvolutionalParams*)network->layers[layer].params;
 371                 av_freep(&conv_params->kernel);
 372                 av_freep(&conv_params->biases);
 373                 av_freep(&conv_params);
 374                 break;
 375             case INPUT:
 376                 av_freep(&network->layers[layer].params);
 377             }
 378         }
 379         av_freep(network);
 380         av_freep(model);
 381     }
 382 }