git.sesse.net Git - ffmpeg/blob - libavfilter/dnn/dnn_backend_tf.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 tensorflow backend implementation.
  24  */
  25
  26 #include "dnn_backend_tf.h"
  27 #include "dnn_backend_native.h"
  28 #include "dnn_backend_native_layer_conv2d.h"
  29 #include "dnn_backend_native_layer_depth2space.h"
  30 #include "libavformat/avio.h"
  31 #include "libavutil/avassert.h"
  32 #include "../internal.h"
  33 #include "dnn_backend_native_layer_pad.h"
  34 #include "dnn_backend_native_layer_maximum.h"
  35 #include "dnn_io_proc.h"
  36
  37 #include <tensorflow/c/c_api.h>
  38
  39 typedef struct TFOptions{
  40     char *sess_config;
  41 } TFOptions;
  42
  43 typedef struct TFContext {
  44     const AVClass *class;
  45     TFOptions options;
  46 } TFContext;
  47
  48 typedef struct TFModel{
  49     TFContext ctx;
  50     DNNModel *model;
  51     TF_Graph *graph;
  52     TF_Session *session;
  53     TF_Status *status;
  54 } TFModel;
  55
  56 #define OFFSET(x) offsetof(TFContext, x)
  57 #define FLAGS AV_OPT_FLAG_FILTERING_PARAM
  58 static const AVOption dnn_tensorflow_options[] = {
  59     { "sess_config", "config for SessionOptions", OFFSET(options.sess_config), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, FLAGS },
  60     { NULL }
  61 };
  62
  63 AVFILTER_DEFINE_CLASS(dnn_tensorflow);
  64
  65 static DNNReturnType execute_model_tf(const DNNModel *model, const char *input_name, AVFrame *in_frame,
  66                                       const char **output_names, uint32_t nb_output, AVFrame *out_frame,
  67                                       int do_ioproc);
  68
  69 static void free_buffer(void *data, size_t length)
  70 {
  71     av_freep(&data);
  72 }
  73
  74 static TF_Buffer *read_graph(const char *model_filename)
  75 {
  76     TF_Buffer *graph_buf;
  77     unsigned char *graph_data = NULL;
  78     AVIOContext *model_file_context;
  79     long size, bytes_read;
  80
  81     if (avio_open(&model_file_context, model_filename, AVIO_FLAG_READ) < 0){
  82         return NULL;
  83     }
  84
  85     size = avio_size(model_file_context);
  86
  87     graph_data = av_malloc(size);
  88     if (!graph_data){
  89         avio_closep(&model_file_context);
  90         return NULL;
  91     }
  92     bytes_read = avio_read(model_file_context, graph_data, size);
  93     avio_closep(&model_file_context);
  94     if (bytes_read != size){
  95         av_freep(&graph_data);
  96         return NULL;
  97     }
  98
  99     graph_buf = TF_NewBuffer();
 100     graph_buf->data = graph_data;
 101     graph_buf->length = size;
 102     graph_buf->data_deallocator = free_buffer;
 103
 104     return graph_buf;
 105 }
 106
 107 static TF_Tensor *allocate_input_tensor(const DNNData *input)
 108 {
 109     TF_DataType dt;
 110     size_t size;
 111     int64_t input_dims[] = {1, input->height, input->width, input->channels};
 112     switch (input->dt) {
 113     case DNN_FLOAT:
 114         dt = TF_FLOAT;
 115         size = sizeof(float);
 116         break;
 117     case DNN_UINT8:
 118         dt = TF_UINT8;
 119         size = 1;
 120         break;
 121     default:
 122         av_assert0(!"should not reach here");
 123     }
 124
 125     return TF_AllocateTensor(dt, input_dims, 4,
 126                              input_dims[1] * input_dims[2] * input_dims[3] * size);
 127 }
 128
 129 static DNNReturnType get_input_tf(void *model, DNNData *input, const char *input_name)
 130 {
 131     TFModel *tf_model = model;
 132     TFContext *ctx = &tf_model->ctx;
 133     TF_Status *status;
 134     int64_t dims[4];
 135
 136     TF_Output tf_output;
 137     tf_output.oper = TF_GraphOperationByName(tf_model->graph, input_name);
 138     if (!tf_output.oper) {
 139         av_log(ctx, AV_LOG_ERROR, "Could not find \"%s\" in model\n", input_name);
 140         return DNN_ERROR;
 141     }
 142
 143     tf_output.index = 0;
 144     input->dt = TF_OperationOutputType(tf_output);
 145
 146     status = TF_NewStatus();
 147     TF_GraphGetTensorShape(tf_model->graph, tf_output, dims, 4, status);
 148     if (TF_GetCode(status) != TF_OK){
 149         TF_DeleteStatus(status);
 150         av_log(ctx, AV_LOG_ERROR, "Failed to get input tensor shape: number of dimension incorrect\n");
 151         return DNN_ERROR;
 152     }
 153     TF_DeleteStatus(status);
 154
 155     // currently only NHWC is supported
 156     av_assert0(dims[0] == 1);
 157     input->height = dims[1];
 158     input->width = dims[2];
 159     input->channels = dims[3];
 160
 161     return DNN_SUCCESS;
 162 }
 163
 164 static DNNReturnType get_output_tf(void *model, const char *input_name, int input_width, int input_height,
 165                                    const char *output_name, int *output_width, int *output_height)
 166 {
 167     DNNReturnType ret;
 168     TFModel *tf_model = model;
 169     TFContext *ctx = &tf_model->ctx;
 170     AVFrame *in_frame = av_frame_alloc();
 171     AVFrame *out_frame = NULL;
 172
 173     if (!in_frame) {
 174         av_log(ctx, AV_LOG_ERROR, "Failed to allocate memory for input frame\n");
 175         return DNN_ERROR;
 176     }
 177
 178     out_frame = av_frame_alloc();
 179     if (!out_frame) {
 180         av_log(ctx, AV_LOG_ERROR, "Failed to allocate memory for output frame\n");
 181         av_frame_free(&in_frame);
 182         return DNN_ERROR;
 183     }
 184
 185     in_frame->width = input_width;
 186     in_frame->height = input_height;
 187
 188     ret = execute_model_tf(tf_model->model, input_name, in_frame, &output_name, 1, out_frame, 0);
 189     *output_width = out_frame->width;
 190     *output_height = out_frame->height;
 191
 192     av_frame_free(&out_frame);
 193     av_frame_free(&in_frame);
 194     return ret;
 195 }
 196
 197 static DNNReturnType load_tf_model(TFModel *tf_model, const char *model_filename)
 198 {
 199     TFContext *ctx = &tf_model->ctx;
 200     TF_Buffer *graph_def;
 201     TF_ImportGraphDefOptions *graph_opts;
 202     TF_SessionOptions *sess_opts;
 203     const TF_Operation *init_op;
 204     uint8_t *sess_config = NULL;
 205     int sess_config_length = 0;
 206
 207     // prepare the sess config data
 208     if (tf_model->ctx.options.sess_config != NULL) {
 209         /*
 210         tf_model->ctx.options.sess_config is hex to present the serialized proto
 211         required by TF_SetConfig below, so we need to first generate the serialized
 212         proto in a python script, the following is a script example to generate
 213         serialized proto which specifies one GPU, we can change the script to add
 214         more options.
 215
 216         import tensorflow as tf
 217         gpu_options = tf.GPUOptions(visible_device_list='0')
 218         config = tf.ConfigProto(gpu_options=gpu_options)
 219         s = config.SerializeToString()
 220         b = ''.join("%02x" % int(ord(b)) for b in s[::-1])
 221         print('0x%s' % b)
 222
 223         the script output looks like: 0xab...cd, and then pass 0xab...cd to sess_config.
 224         */
 225         char tmp[3];
 226         tmp[2] = '\0';
 227
 228         if (strncmp(tf_model->ctx.options.sess_config, "0x", 2) != 0) {
 229             av_log(ctx, AV_LOG_ERROR, "sess_config should start with '0x'\n");
 230             return DNN_ERROR;
 231         }
 232
 233         sess_config_length = strlen(tf_model->ctx.options.sess_config);
 234         if (sess_config_length % 2 != 0) {
 235             av_log(ctx, AV_LOG_ERROR, "the length of sess_config is not even (%s), "
 236                                       "please re-generate the config.\n",
 237                                       tf_model->ctx.options.sess_config);
 238             return DNN_ERROR;
 239         }
 240
 241         sess_config_length -= 2; //ignore the first '0x'
 242         sess_config_length /= 2; //get the data length in byte
 243
 244         sess_config = av_malloc(sess_config_length);
 245         if (!sess_config) {
 246             av_log(ctx, AV_LOG_ERROR, "failed to allocate memory\n");
 247             return DNN_ERROR;
 248         }
 249
 250         for (int i = 0; i < sess_config_length; i++) {
 251             int index = 2 + (sess_config_length - 1 - i) * 2;
 252             tmp[0] = tf_model->ctx.options.sess_config[index];
 253             tmp[1] = tf_model->ctx.options.sess_config[index + 1];
 254             sess_config[i] = strtol(tmp, NULL, 16);
 255         }
 256     }
 257
 258     graph_def = read_graph(model_filename);
 259     if (!graph_def){
 260         av_log(ctx, AV_LOG_ERROR, "Failed to read model \"%s\" graph\n", model_filename);
 261         av_freep(&sess_config);
 262         return DNN_ERROR;
 263     }
 264     tf_model->graph = TF_NewGraph();
 265     tf_model->status = TF_NewStatus();
 266     graph_opts = TF_NewImportGraphDefOptions();
 267     TF_GraphImportGraphDef(tf_model->graph, graph_def, graph_opts, tf_model->status);
 268     TF_DeleteImportGraphDefOptions(graph_opts);
 269     TF_DeleteBuffer(graph_def);
 270     if (TF_GetCode(tf_model->status) != TF_OK){
 271         TF_DeleteGraph(tf_model->graph);
 272         TF_DeleteStatus(tf_model->status);
 273         av_log(ctx, AV_LOG_ERROR, "Failed to import serialized graph to model graph\n");
 274         av_freep(&sess_config);
 275         return DNN_ERROR;
 276     }
 277
 278     init_op = TF_GraphOperationByName(tf_model->graph, "init");
 279     sess_opts = TF_NewSessionOptions();
 280
 281     if (sess_config) {
 282         TF_SetConfig(sess_opts, sess_config, sess_config_length,tf_model->status);
 283         av_freep(&sess_config);
 284         if (TF_GetCode(tf_model->status) != TF_OK) {
 285             TF_DeleteGraph(tf_model->graph);
 286             TF_DeleteStatus(tf_model->status);
 287             TF_DeleteSessionOptions(sess_opts);
 288             av_log(ctx, AV_LOG_ERROR, "Failed to set config for sess options with %s\n",
 289                                       tf_model->ctx.options.sess_config);
 290             return DNN_ERROR;
 291         }
 292     }
 293
 294     tf_model->session = TF_NewSession(tf_model->graph, sess_opts, tf_model->status);
 295     TF_DeleteSessionOptions(sess_opts);
 296     if (TF_GetCode(tf_model->status) != TF_OK)
 297     {
 298         TF_DeleteGraph(tf_model->graph);
 299         TF_DeleteStatus(tf_model->status);
 300         av_log(ctx, AV_LOG_ERROR, "Failed to create new session with model graph\n");
 301         return DNN_ERROR;
 302     }
 303
 304     // Run initialization operation with name "init" if it is present in graph
 305     if (init_op){
 306         TF_SessionRun(tf_model->session, NULL,
 307                       NULL, NULL, 0,
 308                       NULL, NULL, 0,
 309                       &init_op, 1, NULL, tf_model->status);
 310         if (TF_GetCode(tf_model->status) != TF_OK)
 311         {
 312             TF_DeleteSession(tf_model->session, tf_model->status);
 313             TF_DeleteGraph(tf_model->graph);
 314             TF_DeleteStatus(tf_model->status);
 315             av_log(ctx, AV_LOG_ERROR, "Failed to run session when initializing\n");
 316             return DNN_ERROR;
 317         }
 318     }
 319
 320     return DNN_SUCCESS;
 321 }
 322
 323 #define NAME_BUFFER_SIZE 256
 324
 325 static DNNReturnType add_conv_layer(TFModel *tf_model, TF_Operation *transpose_op, TF_Operation **cur_op,
 326                                     ConvolutionalParams* params, const int layer)
 327 {
 328     TFContext *ctx = &tf_model->ctx;
 329     TF_Operation *op;
 330     TF_OperationDescription *op_desc;
 331     TF_Output input;
 332     int64_t strides[] = {1, 1, 1, 1};
 333     TF_Tensor *tensor;
 334     int64_t dims[4];
 335     int dims_len;
 336     char name_buffer[NAME_BUFFER_SIZE];
 337     int32_t size;
 338
 339     size = params->input_num * params->output_num * params->kernel_size * params->kernel_size;
 340     input.index = 0;
 341
 342     snprintf(name_buffer, NAME_BUFFER_SIZE, "conv_kernel%d", layer);
 343     op_desc = TF_NewOperation(tf_model->graph, "Const", name_buffer);
 344     TF_SetAttrType(op_desc, "dtype", TF_FLOAT);
 345     dims[0] = params->output_num;
 346     dims[1] = params->kernel_size;
 347     dims[2] = params->kernel_size;
 348     dims[3] = params->input_num;
 349     dims_len = 4;
 350     tensor = TF_AllocateTensor(TF_FLOAT, dims, dims_len, size * sizeof(float));
 351     memcpy(TF_TensorData(tensor), params->kernel, size * sizeof(float));
 352     TF_SetAttrTensor(op_desc, "value", tensor, tf_model->status);
 353     if (TF_GetCode(tf_model->status) != TF_OK){
 354         av_log(ctx, AV_LOG_ERROR, "Failed to set value for kernel of conv layer %d\n", layer);
 355         return DNN_ERROR;
 356     }
 357     op = TF_FinishOperation(op_desc, tf_model->status);
 358     if (TF_GetCode(tf_model->status) != TF_OK){
 359         av_log(ctx, AV_LOG_ERROR, "Failed to add kernel to conv layer %d\n", layer);
 360         return DNN_ERROR;
 361     }
 362
 363     snprintf(name_buffer, NAME_BUFFER_SIZE, "transpose%d", layer);
 364     op_desc = TF_NewOperation(tf_model->graph, "Transpose", name_buffer);
 365     input.oper = op;
 366     TF_AddInput(op_desc, input);
 367     input.oper = transpose_op;
 368     TF_AddInput(op_desc, input);
 369     TF_SetAttrType(op_desc, "T", TF_FLOAT);
 370     TF_SetAttrType(op_desc, "Tperm", TF_INT32);
 371     op = TF_FinishOperation(op_desc, tf_model->status);
 372     if (TF_GetCode(tf_model->status) != TF_OK){
 373         av_log(ctx, AV_LOG_ERROR, "Failed to add transpose to conv layer %d\n", layer);
 374         return DNN_ERROR;
 375     }
 376
 377     snprintf(name_buffer, NAME_BUFFER_SIZE, "conv2d%d", layer);
 378     op_desc = TF_NewOperation(tf_model->graph, "Conv2D", name_buffer);
 379     input.oper = *cur_op;
 380     TF_AddInput(op_desc, input);
 381     input.oper = op;
 382     TF_AddInput(op_desc, input);
 383     TF_SetAttrType(op_desc, "T", TF_FLOAT);
 384     TF_SetAttrIntList(op_desc, "strides", strides, 4);
 385     TF_SetAttrString(op_desc, "padding", "VALID", 5);
 386     *cur_op = TF_FinishOperation(op_desc, tf_model->status);
 387     if (TF_GetCode(tf_model->status) != TF_OK){
 388         av_log(ctx, AV_LOG_ERROR, "Failed to add conv2d to conv layer %d\n", layer);
 389         return DNN_ERROR;
 390     }
 391
 392     snprintf(name_buffer, NAME_BUFFER_SIZE, "conv_biases%d", layer);
 393     op_desc = TF_NewOperation(tf_model->graph, "Const", name_buffer);
 394     TF_SetAttrType(op_desc, "dtype", TF_FLOAT);
 395     dims[0] = params->output_num;
 396     dims_len = 1;
 397     tensor = TF_AllocateTensor(TF_FLOAT, dims, dims_len, params->output_num * sizeof(float));
 398     memcpy(TF_TensorData(tensor), params->biases, params->output_num * sizeof(float));
 399     TF_SetAttrTensor(op_desc, "value", tensor, tf_model->status);
 400     if (TF_GetCode(tf_model->status) != TF_OK){
 401         av_log(ctx, AV_LOG_ERROR, "Failed to set value for conv_biases of conv layer %d\n", layer);
 402         return DNN_ERROR;
 403     }
 404     op = TF_FinishOperation(op_desc, tf_model->status);
 405     if (TF_GetCode(tf_model->status) != TF_OK){
 406         av_log(ctx, AV_LOG_ERROR, "Failed to add conv_biases to conv layer %d\n", layer);
 407         return DNN_ERROR;
 408     }
 409
 410     snprintf(name_buffer, NAME_BUFFER_SIZE, "bias_add%d", layer);
 411     op_desc = TF_NewOperation(tf_model->graph, "BiasAdd", name_buffer);
 412     input.oper = *cur_op;
 413     TF_AddInput(op_desc, input);
 414     input.oper = op;
 415     TF_AddInput(op_desc, input);
 416     TF_SetAttrType(op_desc, "T", TF_FLOAT);
 417     *cur_op = TF_FinishOperation(op_desc, tf_model->status);
 418     if (TF_GetCode(tf_model->status) != TF_OK){
 419         av_log(ctx, AV_LOG_ERROR, "Failed to add bias_add to conv layer %d\n", layer);
 420         return DNN_ERROR;
 421     }
 422
 423     snprintf(name_buffer, NAME_BUFFER_SIZE, "activation%d", layer);
 424     switch (params->activation){
 425     case RELU:
 426         op_desc = TF_NewOperation(tf_model->graph, "Relu", name_buffer);
 427         break;
 428     case TANH:
 429         op_desc = TF_NewOperation(tf_model->graph, "Tanh", name_buffer);
 430         break;
 431     case SIGMOID:
 432         op_desc = TF_NewOperation(tf_model->graph, "Sigmoid", name_buffer);
 433         break;
 434     default:
 435         avpriv_report_missing_feature(ctx, "convolutional activation function %d", params->activation);
 436         return DNN_ERROR;
 437     }
 438     input.oper = *cur_op;
 439     TF_AddInput(op_desc, input);
 440     TF_SetAttrType(op_desc, "T", TF_FLOAT);
 441     *cur_op = TF_FinishOperation(op_desc, tf_model->status);
 442     if (TF_GetCode(tf_model->status) != TF_OK){
 443         av_log(ctx, AV_LOG_ERROR, "Failed to add activation function to conv layer %d\n", layer);
 444         return DNN_ERROR;
 445     }
 446
 447     return DNN_SUCCESS;
 448 }
 449
 450 static DNNReturnType add_depth_to_space_layer(TFModel *tf_model, TF_Operation **cur_op,
 451                                               DepthToSpaceParams *params, const int layer)
 452 {
 453     TFContext *ctx = &tf_model->ctx;
 454     TF_OperationDescription *op_desc;
 455     TF_Output input;
 456     char name_buffer[NAME_BUFFER_SIZE];
 457
 458     snprintf(name_buffer, NAME_BUFFER_SIZE, "depth_to_space%d", layer);
 459     op_desc = TF_NewOperation(tf_model->graph, "DepthToSpace", name_buffer);
 460     input.oper = *cur_op;
 461     input.index = 0;
 462     TF_AddInput(op_desc, input);
 463     TF_SetAttrType(op_desc, "T", TF_FLOAT);
 464     TF_SetAttrInt(op_desc, "block_size", params->block_size);
 465     *cur_op = TF_FinishOperation(op_desc, tf_model->status);
 466     if (TF_GetCode(tf_model->status) != TF_OK){
 467         av_log(ctx, AV_LOG_ERROR, "Failed to add depth_to_space to layer %d\n", layer);
 468         return DNN_ERROR;
 469     }
 470
 471     return DNN_SUCCESS;
 472 }
 473
 474 static DNNReturnType add_pad_layer(TFModel *tf_model, TF_Operation **cur_op,
 475                                               LayerPadParams *params, const int layer)
 476 {
 477     TFContext *ctx = &tf_model->ctx;
 478     TF_Operation *op;
 479     TF_Tensor *tensor;
 480     TF_OperationDescription *op_desc;
 481     TF_Output input;
 482     int32_t *pads;
 483     int64_t pads_shape[] = {4, 2};
 484
 485     char name_buffer[NAME_BUFFER_SIZE];
 486     snprintf(name_buffer, NAME_BUFFER_SIZE, "pad%d", layer);
 487
 488     op_desc = TF_NewOperation(tf_model->graph, "Const", name_buffer);
 489     TF_SetAttrType(op_desc, "dtype", TF_INT32);
 490     tensor = TF_AllocateTensor(TF_INT32, pads_shape, 2, 4 * 2 * sizeof(int32_t));
 491     pads = (int32_t *)TF_TensorData(tensor);
 492     pads[0] = params->paddings[0][0];
 493     pads[1] = params->paddings[0][1];
 494     pads[2] = params->paddings[1][0];
 495     pads[3] = params->paddings[1][1];
 496     pads[4] = params->paddings[2][0];
 497     pads[5] = params->paddings[2][1];
 498     pads[6] = params->paddings[3][0];
 499     pads[7] = params->paddings[3][1];
 500     TF_SetAttrTensor(op_desc, "value", tensor, tf_model->status);
 501     if (TF_GetCode(tf_model->status) != TF_OK){
 502         av_log(ctx, AV_LOG_ERROR, "Failed to set value for pad of layer %d\n", layer);
 503         return DNN_ERROR;
 504     }
 505     op = TF_FinishOperation(op_desc, tf_model->status);
 506     if (TF_GetCode(tf_model->status) != TF_OK){
 507         av_log(ctx, AV_LOG_ERROR, "Failed to add pad to layer %d\n", layer);
 508         return DNN_ERROR;
 509     }
 510
 511     op_desc = TF_NewOperation(tf_model->graph, "MirrorPad", "mirror_pad");
 512     input.oper = *cur_op;
 513     input.index = 0;
 514     TF_AddInput(op_desc, input);
 515     input.oper = op;
 516     TF_AddInput(op_desc, input);
 517     TF_SetAttrType(op_desc, "T", TF_FLOAT);
 518     TF_SetAttrType(op_desc, "Tpaddings", TF_INT32);
 519     TF_SetAttrString(op_desc, "mode", "SYMMETRIC", 9);
 520     *cur_op = TF_FinishOperation(op_desc, tf_model->status);
 521     if (TF_GetCode(tf_model->status) != TF_OK){
 522         av_log(ctx, AV_LOG_ERROR, "Failed to add mirror_pad to layer %d\n", layer);
 523         return DNN_ERROR;
 524     }
 525
 526     return DNN_SUCCESS;
 527 }
 528
 529 static DNNReturnType add_maximum_layer(TFModel *tf_model, TF_Operation **cur_op,
 530                                        DnnLayerMaximumParams *params, const int layer)
 531 {
 532     TFContext *ctx = &tf_model->ctx;
 533     TF_Operation *op;
 534     TF_Tensor *tensor;
 535     TF_OperationDescription *op_desc;
 536     TF_Output input;
 537     float *y;
 538
 539     char name_buffer[NAME_BUFFER_SIZE];
 540     snprintf(name_buffer, NAME_BUFFER_SIZE, "maximum/y%d", layer);
 541
 542     op_desc = TF_NewOperation(tf_model->graph, "Const", name_buffer);
 543     TF_SetAttrType(op_desc, "dtype", TF_FLOAT);
 544     tensor = TF_AllocateTensor(TF_FLOAT, NULL, 0, TF_DataTypeSize(TF_FLOAT));
 545     y = (float *)TF_TensorData(tensor);
 546     *y = params->val.y;
 547     TF_SetAttrTensor(op_desc, "value", tensor, tf_model->status);
 548     if (TF_GetCode(tf_model->status) != TF_OK){
 549         av_log(ctx, AV_LOG_ERROR, "Failed to set value for maximum/y of layer %d", layer);
 550         return DNN_ERROR;
 551     }
 552     op = TF_FinishOperation(op_desc, tf_model->status);
 553     if (TF_GetCode(tf_model->status) != TF_OK){
 554         av_log(ctx, AV_LOG_ERROR, "Failed to add maximum/y to layer %d\n", layer);
 555         return DNN_ERROR;
 556     }
 557
 558     snprintf(name_buffer, NAME_BUFFER_SIZE, "maximum%d", layer);
 559     op_desc = TF_NewOperation(tf_model->graph, "Maximum", name_buffer);
 560     input.oper = *cur_op;
 561     input.index = 0;
 562     TF_AddInput(op_desc, input);
 563     input.oper = op;
 564     TF_AddInput(op_desc, input);
 565     TF_SetAttrType(op_desc, "T", TF_FLOAT);
 566     *cur_op = TF_FinishOperation(op_desc, tf_model->status);
 567     if (TF_GetCode(tf_model->status) != TF_OK){
 568         av_log(ctx, AV_LOG_ERROR, "Failed to add maximum to layer %d\n", layer);
 569         return DNN_ERROR;
 570     }
 571
 572     return DNN_SUCCESS;
 573 }
 574
 575 static DNNReturnType load_native_model(TFModel *tf_model, const char *model_filename)
 576 {
 577     TFContext *ctx = &tf_model->ctx;
 578     int32_t layer;
 579     TF_OperationDescription *op_desc;
 580     TF_Operation *op;
 581     TF_Operation *transpose_op;
 582     TF_Tensor *tensor;
 583     TF_Output input;
 584     int32_t *transpose_perm;
 585     int64_t transpose_perm_shape[] = {4};
 586     int64_t input_shape[] = {1, -1, -1, -1};
 587     DNNReturnType layer_add_res;
 588     DNNModel *model = NULL;
 589     NativeModel *native_model;
 590
 591     model = ff_dnn_load_model_native(model_filename, DFT_PROCESS_FRAME, NULL, NULL);
 592     if (!model){
 593         av_log(ctx, AV_LOG_ERROR, "Failed to load native model\n");
 594         return DNN_ERROR;
 595     }
 596
 597     native_model = model->model;
 598     tf_model->graph = TF_NewGraph();
 599     tf_model->status = TF_NewStatus();
 600
 601 #define CLEANUP_ON_ERROR(tf_model) \
 602     { \
 603         TF_DeleteGraph(tf_model->graph); \
 604         TF_DeleteStatus(tf_model->status); \
 605         av_log(ctx, AV_LOG_ERROR, "Failed to set value or add operator to layer\n"); \
 606         return DNN_ERROR; \
 607     }
 608
 609     op_desc = TF_NewOperation(tf_model->graph, "Placeholder", "x");
 610     TF_SetAttrType(op_desc, "dtype", TF_FLOAT);
 611     TF_SetAttrShape(op_desc, "shape", input_shape, 4);
 612     op = TF_FinishOperation(op_desc, tf_model->status);
 613     if (TF_GetCode(tf_model->status) != TF_OK){
 614         CLEANUP_ON_ERROR(tf_model);
 615     }
 616
 617     op_desc = TF_NewOperation(tf_model->graph, "Const", "transpose_perm");
 618     TF_SetAttrType(op_desc, "dtype", TF_INT32);
 619     tensor = TF_AllocateTensor(TF_INT32, transpose_perm_shape, 1, 4 * sizeof(int32_t));
 620     transpose_perm = (int32_t *)TF_TensorData(tensor);
 621     transpose_perm[0] = 1;
 622     transpose_perm[1] = 2;
 623     transpose_perm[2] = 3;
 624     transpose_perm[3] = 0;
 625     TF_SetAttrTensor(op_desc, "value", tensor, tf_model->status);
 626     if (TF_GetCode(tf_model->status) != TF_OK){
 627         CLEANUP_ON_ERROR(tf_model);
 628     }
 629     transpose_op = TF_FinishOperation(op_desc, tf_model->status);
 630
 631     for (layer = 0; layer < native_model->layers_num; ++layer){
 632         switch (native_model->layers[layer].type){
 633         case DLT_INPUT:
 634             layer_add_res = DNN_SUCCESS;
 635             break;
 636         case DLT_CONV2D:
 637             layer_add_res = add_conv_layer(tf_model, transpose_op, &op,
 638                                            (ConvolutionalParams *)native_model->layers[layer].params, layer);
 639             break;
 640         case DLT_DEPTH_TO_SPACE:
 641             layer_add_res = add_depth_to_space_layer(tf_model, &op,
 642                                                      (DepthToSpaceParams *)native_model->layers[layer].params, layer);
 643             break;
 644         case DLT_MIRROR_PAD:
 645             layer_add_res = add_pad_layer(tf_model, &op,
 646                                           (LayerPadParams *)native_model->layers[layer].params, layer);
 647             break;
 648         case DLT_MAXIMUM:
 649             layer_add_res = add_maximum_layer(tf_model, &op,
 650                                           (DnnLayerMaximumParams *)native_model->layers[layer].params, layer);
 651             break;
 652         default:
 653             CLEANUP_ON_ERROR(tf_model);
 654         }
 655
 656         if (layer_add_res != DNN_SUCCESS){
 657             CLEANUP_ON_ERROR(tf_model);
 658         }
 659     }
 660
 661     op_desc = TF_NewOperation(tf_model->graph, "Identity", "y");
 662     input.oper = op;
 663     input.index = 0;
 664     TF_AddInput(op_desc, input);
 665     TF_FinishOperation(op_desc, tf_model->status);
 666     if (TF_GetCode(tf_model->status) != TF_OK){
 667         CLEANUP_ON_ERROR(tf_model);
 668     }
 669
 670     ff_dnn_free_model_native(&model);
 671
 672     return DNN_SUCCESS;
 673 }
 674
 675 DNNModel *ff_dnn_load_model_tf(const char *model_filename, DNNFunctionType func_type, const char *options, AVFilterContext *filter_ctx)
 676 {
 677     DNNModel *model = NULL;
 678     TFModel *tf_model = NULL;
 679
 680     model = av_mallocz(sizeof(DNNModel));
 681     if (!model){
 682         return NULL;
 683     }
 684
 685     tf_model = av_mallocz(sizeof(TFModel));
 686     if (!tf_model){
 687         av_freep(&model);
 688         return NULL;
 689     }
 690     tf_model->ctx.class = &dnn_tensorflow_class;
 691     tf_model->model = model;
 692
 693     //parse options
 694     av_opt_set_defaults(&tf_model->ctx);
 695     if (av_opt_set_from_string(&tf_model->ctx, options, NULL, "=", "&") < 0) {
 696         av_log(&tf_model->ctx, AV_LOG_ERROR, "Failed to parse options \"%s\"\n", options);
 697         av_freep(&tf_model);
 698         av_freep(&model);
 699         return NULL;
 700     }
 701
 702     if (load_tf_model(tf_model, model_filename) != DNN_SUCCESS){
 703         if (load_native_model(tf_model, model_filename) != DNN_SUCCESS){
 704             av_freep(&tf_model);
 705             av_freep(&model);
 706
 707             return NULL;
 708         }
 709     }
 710
 711     model->model = tf_model;
 712     model->get_input = &get_input_tf;
 713     model->get_output = &get_output_tf;
 714     model->options = options;
 715     model->filter_ctx = filter_ctx;
 716     model->func_type = func_type;
 717
 718     return model;
 719 }
 720
 721 static DNNReturnType execute_model_tf(const DNNModel *model, const char *input_name, AVFrame *in_frame,
 722                                       const char **output_names, uint32_t nb_output, AVFrame *out_frame,
 723                                       int do_ioproc)
 724 {
 725     TF_Output *tf_outputs;
 726     TFModel *tf_model = model->model;
 727     TFContext *ctx = &tf_model->ctx;
 728     DNNData input, output;
 729     TF_Tensor **output_tensors;
 730     TF_Output tf_input;
 731     TF_Tensor *input_tensor;
 732
 733     if (get_input_tf(tf_model, &input, input_name) != DNN_SUCCESS)
 734         return DNN_ERROR;
 735     input.height = in_frame->height;
 736     input.width = in_frame->width;
 737
 738     tf_input.oper = TF_GraphOperationByName(tf_model->graph, input_name);
 739     if (!tf_input.oper){
 740         av_log(ctx, AV_LOG_ERROR, "Could not find \"%s\" in model\n", input_name);
 741         return DNN_ERROR;
 742     }
 743     tf_input.index = 0;
 744     input_tensor = allocate_input_tensor(&input);
 745     if (!input_tensor){
 746         av_log(ctx, AV_LOG_ERROR, "Failed to allocate memory for input tensor\n");
 747         return DNN_ERROR;
 748     }
 749     input.data = (float *)TF_TensorData(input_tensor);
 750
 751     if (do_ioproc) {
 752         if (tf_model->model->pre_proc != NULL) {
 753             tf_model->model->pre_proc(in_frame, &input, tf_model->model->filter_ctx);
 754         } else {
 755             ff_proc_from_frame_to_dnn(in_frame, &input, tf_model->model->func_type, ctx);
 756         }
 757     }
 758
 759     if (nb_output != 1) {
 760         // currently, the filter does not need multiple outputs,
 761         // so we just pending the support until we really need it.
 762         avpriv_report_missing_feature(ctx, "multiple outputs");
 763         return DNN_ERROR;
 764     }
 765
 766     tf_outputs = av_malloc_array(nb_output, sizeof(*tf_outputs));
 767     if (tf_outputs == NULL) {
 768         av_log(ctx, AV_LOG_ERROR, "Failed to allocate memory for *tf_outputs\n"); \
 769         return DNN_ERROR;
 770     }
 771
 772     output_tensors = av_mallocz_array(nb_output, sizeof(*output_tensors));
 773     if (!output_tensors) {
 774         av_freep(&tf_outputs);
 775         av_log(ctx, AV_LOG_ERROR, "Failed to allocate memory for output tensor\n"); \
 776         return DNN_ERROR;
 777     }
 778
 779     for (int i = 0; i < nb_output; ++i) {
 780         tf_outputs[i].oper = TF_GraphOperationByName(tf_model->graph, output_names[i]);
 781         if (!tf_outputs[i].oper) {
 782             av_freep(&tf_outputs);
 783             av_freep(&output_tensors);
 784             av_log(ctx, AV_LOG_ERROR, "Could not find output \"%s\" in model\n", output_names[i]); \
 785             return DNN_ERROR;
 786         }
 787         tf_outputs[i].index = 0;
 788     }
 789
 790     TF_SessionRun(tf_model->session, NULL,
 791                   &tf_input, &input_tensor, 1,
 792                   tf_outputs, output_tensors, nb_output,
 793                   NULL, 0, NULL, tf_model->status);
 794     if (TF_GetCode(tf_model->status) != TF_OK) {
 795         av_freep(&tf_outputs);
 796         av_freep(&output_tensors);
 797         av_log(ctx, AV_LOG_ERROR, "Failed to run session when executing model\n");
 798         return DNN_ERROR;
 799     }
 800
 801     for (uint32_t i = 0; i < nb_output; ++i) {
 802         output.height = TF_Dim(output_tensors[i], 1);
 803         output.width = TF_Dim(output_tensors[i], 2);
 804         output.channels = TF_Dim(output_tensors[i], 3);
 805         output.data = TF_TensorData(output_tensors[i]);
 806         output.dt = TF_TensorType(output_tensors[i]);
 807
 808         if (do_ioproc) {
 809             if (tf_model->model->post_proc != NULL) {
 810                 tf_model->model->post_proc(out_frame, &output, tf_model->model->filter_ctx);
 811             } else {
 812                 ff_proc_from_dnn_to_frame(out_frame, &output, ctx);
 813             }
 814         } else {
 815             out_frame->width = output.width;
 816             out_frame->height = output.height;
 817         }
 818     }
 819
 820     for (uint32_t i = 0; i < nb_output; ++i) {
 821         if (output_tensors[i]) {
 822             TF_DeleteTensor(output_tensors[i]);
 823         }
 824     }
 825     TF_DeleteTensor(input_tensor);
 826     av_freep(&output_tensors);
 827     av_freep(&tf_outputs);
 828     return DNN_SUCCESS;
 829 }
 830
 831 DNNReturnType ff_dnn_execute_model_tf(const DNNModel *model, const char *input_name, AVFrame *in_frame,
 832                                       const char **output_names, uint32_t nb_output, AVFrame *out_frame)
 833 {
 834     TFModel *tf_model = model->model;
 835     TFContext *ctx = &tf_model->ctx;
 836
 837     if (!in_frame) {
 838         av_log(ctx, AV_LOG_ERROR, "in frame is NULL when execute model.\n");
 839         return DNN_ERROR;
 840     }
 841
 842     if (!out_frame) {
 843         av_log(ctx, AV_LOG_ERROR, "out frame is NULL when execute model.\n");
 844         return DNN_ERROR;
 845     }
 846
 847     return execute_model_tf(model, input_name, in_frame, output_names, nb_output, out_frame, 1);
 848 }
 849
 850 void ff_dnn_free_model_tf(DNNModel **model)
 851 {
 852     TFModel *tf_model;
 853
 854     if (*model){
 855         tf_model = (*model)->model;
 856         if (tf_model->graph){
 857             TF_DeleteGraph(tf_model->graph);
 858         }
 859         if (tf_model->session){
 860             TF_CloseSession(tf_model->session, tf_model->status);
 861             TF_DeleteSession(tf_model->session, tf_model->status);
 862         }
 863         if (tf_model->status){
 864             TF_DeleteStatus(tf_model->status);
 865         }
 866         av_freep(&tf_model);
 867         av_freep(model);
 868     }
 869 }