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 *kernel_tensor = NULL, *biases_tensor = NULL;
 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     kernel_tensor = TF_AllocateTensor(TF_FLOAT, dims, dims_len, size * sizeof(float));
 351     memcpy(TF_TensorData(kernel_tensor), params->kernel, size * sizeof(float));
 352     TF_SetAttrTensor(op_desc, "value", kernel_tensor, tf_model->status);
 353     if (TF_GetCode(tf_model->status) != TF_OK){
 354         goto err;
 355     }
 356     op = TF_FinishOperation(op_desc, tf_model->status);
 357     if (TF_GetCode(tf_model->status) != TF_OK){
 358         goto err;
 359     }
 360
 361     snprintf(name_buffer, NAME_BUFFER_SIZE, "transpose%d", layer);
 362     op_desc = TF_NewOperation(tf_model->graph, "Transpose", name_buffer);
 363     input.oper = op;
 364     TF_AddInput(op_desc, input);
 365     input.oper = transpose_op;
 366     TF_AddInput(op_desc, input);
 367     TF_SetAttrType(op_desc, "T", TF_FLOAT);
 368     TF_SetAttrType(op_desc, "Tperm", TF_INT32);
 369     op = TF_FinishOperation(op_desc, tf_model->status);
 370     if (TF_GetCode(tf_model->status) != TF_OK){
 371         goto err;
 372     }
 373
 374     snprintf(name_buffer, NAME_BUFFER_SIZE, "conv2d%d", layer);
 375     op_desc = TF_NewOperation(tf_model->graph, "Conv2D", name_buffer);
 376     input.oper = *cur_op;
 377     TF_AddInput(op_desc, input);
 378     input.oper = op;
 379     TF_AddInput(op_desc, input);
 380     TF_SetAttrType(op_desc, "T", TF_FLOAT);
 381     TF_SetAttrIntList(op_desc, "strides", strides, 4);
 382     TF_SetAttrString(op_desc, "padding", "VALID", 5);
 383     *cur_op = TF_FinishOperation(op_desc, tf_model->status);
 384     if (TF_GetCode(tf_model->status) != TF_OK){
 385         goto err;
 386     }
 387
 388     snprintf(name_buffer, NAME_BUFFER_SIZE, "conv_biases%d", layer);
 389     op_desc = TF_NewOperation(tf_model->graph, "Const", name_buffer);
 390     TF_SetAttrType(op_desc, "dtype", TF_FLOAT);
 391     dims[0] = params->output_num;
 392     dims_len = 1;
 393     biases_tensor = TF_AllocateTensor(TF_FLOAT, dims, dims_len, params->output_num * sizeof(float));
 394     memcpy(TF_TensorData(biases_tensor), params->biases, params->output_num * sizeof(float));
 395     TF_SetAttrTensor(op_desc, "value", biases_tensor, tf_model->status);
 396     if (TF_GetCode(tf_model->status) != TF_OK){
 397         goto err;
 398     }
 399     op = TF_FinishOperation(op_desc, tf_model->status);
 400     if (TF_GetCode(tf_model->status) != TF_OK){
 401         goto err;
 402     }
 403
 404     snprintf(name_buffer, NAME_BUFFER_SIZE, "bias_add%d", layer);
 405     op_desc = TF_NewOperation(tf_model->graph, "BiasAdd", name_buffer);
 406     input.oper = *cur_op;
 407     TF_AddInput(op_desc, input);
 408     input.oper = op;
 409     TF_AddInput(op_desc, input);
 410     TF_SetAttrType(op_desc, "T", TF_FLOAT);
 411     *cur_op = TF_FinishOperation(op_desc, tf_model->status);
 412     if (TF_GetCode(tf_model->status) != TF_OK){
 413         goto err;
 414     }
 415
 416     snprintf(name_buffer, NAME_BUFFER_SIZE, "activation%d", layer);
 417     switch (params->activation){
 418     case RELU:
 419         op_desc = TF_NewOperation(tf_model->graph, "Relu", name_buffer);
 420         break;
 421     case TANH:
 422         op_desc = TF_NewOperation(tf_model->graph, "Tanh", name_buffer);
 423         break;
 424     case SIGMOID:
 425         op_desc = TF_NewOperation(tf_model->graph, "Sigmoid", name_buffer);
 426         break;
 427     default:
 428         avpriv_report_missing_feature(ctx, "convolutional activation function %d", params->activation);
 429         return DNN_ERROR;
 430     }
 431     input.oper = *cur_op;
 432     TF_AddInput(op_desc, input);
 433     TF_SetAttrType(op_desc, "T", TF_FLOAT);
 434     *cur_op = TF_FinishOperation(op_desc, tf_model->status);
 435     if (TF_GetCode(tf_model->status) != TF_OK){
 436         goto err;
 437     }
 438
 439     return DNN_SUCCESS;
 440 err:
 441     TF_DeleteTensor(kernel_tensor);
 442     TF_DeleteTensor(biases_tensor);
 443     av_log(ctx, AV_LOG_ERROR, "Failed to add conv layer %d\n", layer);
 444     return DNN_ERROR;
 445 }
 446
 447 static DNNReturnType add_depth_to_space_layer(TFModel *tf_model, TF_Operation **cur_op,
 448                                               DepthToSpaceParams *params, const int layer)
 449 {
 450     TFContext *ctx = &tf_model->ctx;
 451     TF_OperationDescription *op_desc;
 452     TF_Output input;
 453     char name_buffer[NAME_BUFFER_SIZE];
 454
 455     snprintf(name_buffer, NAME_BUFFER_SIZE, "depth_to_space%d", layer);
 456     op_desc = TF_NewOperation(tf_model->graph, "DepthToSpace", name_buffer);
 457     input.oper = *cur_op;
 458     input.index = 0;
 459     TF_AddInput(op_desc, input);
 460     TF_SetAttrType(op_desc, "T", TF_FLOAT);
 461     TF_SetAttrInt(op_desc, "block_size", params->block_size);
 462     *cur_op = TF_FinishOperation(op_desc, tf_model->status);
 463     if (TF_GetCode(tf_model->status) != TF_OK){
 464         av_log(ctx, AV_LOG_ERROR, "Failed to add depth_to_space to layer %d\n", layer);
 465         return DNN_ERROR;
 466     }
 467
 468     return DNN_SUCCESS;
 469 }
 470
 471 static DNNReturnType add_pad_layer(TFModel *tf_model, TF_Operation **cur_op,
 472                                               LayerPadParams *params, const int layer)
 473 {
 474     TFContext *ctx = &tf_model->ctx;
 475     TF_Operation *op;
 476     TF_Tensor *tensor;
 477     TF_OperationDescription *op_desc;
 478     TF_Output input;
 479     int32_t *pads;
 480     int64_t pads_shape[] = {4, 2};
 481
 482     char name_buffer[NAME_BUFFER_SIZE];
 483     snprintf(name_buffer, NAME_BUFFER_SIZE, "pad%d", layer);
 484
 485     op_desc = TF_NewOperation(tf_model->graph, "Const", name_buffer);
 486     TF_SetAttrType(op_desc, "dtype", TF_INT32);
 487     tensor = TF_AllocateTensor(TF_INT32, pads_shape, 2, 4 * 2 * sizeof(int32_t));
 488     pads = (int32_t *)TF_TensorData(tensor);
 489     pads[0] = params->paddings[0][0];
 490     pads[1] = params->paddings[0][1];
 491     pads[2] = params->paddings[1][0];
 492     pads[3] = params->paddings[1][1];
 493     pads[4] = params->paddings[2][0];
 494     pads[5] = params->paddings[2][1];
 495     pads[6] = params->paddings[3][0];
 496     pads[7] = params->paddings[3][1];
 497     TF_SetAttrTensor(op_desc, "value", tensor, tf_model->status);
 498     if (TF_GetCode(tf_model->status) != TF_OK){
 499         TF_DeleteTensor(tensor);
 500         av_log(ctx, AV_LOG_ERROR, "Failed to set value for pad of layer %d\n", layer);
 501         return DNN_ERROR;
 502     }
 503     op = TF_FinishOperation(op_desc, tf_model->status);
 504     if (TF_GetCode(tf_model->status) != TF_OK){
 505         TF_DeleteTensor(tensor);
 506         av_log(ctx, AV_LOG_ERROR, "Failed to add pad to layer %d\n", layer);
 507         return DNN_ERROR;
 508     }
 509
 510     op_desc = TF_NewOperation(tf_model->graph, "MirrorPad", "mirror_pad");
 511     input.oper = *cur_op;
 512     input.index = 0;
 513     TF_AddInput(op_desc, input);
 514     input.oper = op;
 515     TF_AddInput(op_desc, input);
 516     TF_SetAttrType(op_desc, "T", TF_FLOAT);
 517     TF_SetAttrType(op_desc, "Tpaddings", TF_INT32);
 518     TF_SetAttrString(op_desc, "mode", "SYMMETRIC", 9);
 519     *cur_op = TF_FinishOperation(op_desc, tf_model->status);
 520     if (TF_GetCode(tf_model->status) != TF_OK){
 521         TF_DeleteTensor(tensor);
 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         TF_DeleteTensor(tensor);
 550         av_log(ctx, AV_LOG_ERROR, "Failed to set value for maximum/y of layer %d", layer);
 551         return DNN_ERROR;
 552     }
 553     op = TF_FinishOperation(op_desc, tf_model->status);
 554     if (TF_GetCode(tf_model->status) != TF_OK){
 555         TF_DeleteTensor(tensor);
 556         av_log(ctx, AV_LOG_ERROR, "Failed to add maximum/y to layer %d\n", layer);
 557         return DNN_ERROR;
 558     }
 559
 560     snprintf(name_buffer, NAME_BUFFER_SIZE, "maximum%d", layer);
 561     op_desc = TF_NewOperation(tf_model->graph, "Maximum", name_buffer);
 562     input.oper = *cur_op;
 563     input.index = 0;
 564     TF_AddInput(op_desc, input);
 565     input.oper = op;
 566     TF_AddInput(op_desc, input);
 567     TF_SetAttrType(op_desc, "T", TF_FLOAT);
 568     *cur_op = TF_FinishOperation(op_desc, tf_model->status);
 569     if (TF_GetCode(tf_model->status) != TF_OK){
 570         TF_DeleteTensor(tensor);
 571         av_log(ctx, AV_LOG_ERROR, "Failed to add maximum to layer %d\n", layer);
 572         return DNN_ERROR;
 573     }
 574
 575     return DNN_SUCCESS;
 576 }
 577
 578 static DNNReturnType load_native_model(TFModel *tf_model, const char *model_filename)
 579 {
 580     TFContext *ctx = &tf_model->ctx;
 581     int32_t layer;
 582     TF_OperationDescription *op_desc;
 583     TF_Operation *op;
 584     TF_Operation *transpose_op;
 585     TF_Tensor *tensor = NULL;
 586     TF_Output input;
 587     int32_t *transpose_perm;
 588     int64_t transpose_perm_shape[] = {4};
 589     int64_t input_shape[] = {1, -1, -1, -1};
 590     DNNReturnType layer_add_res;
 591     DNNModel *model = NULL;
 592     NativeModel *native_model;
 593
 594     model = ff_dnn_load_model_native(model_filename, DFT_PROCESS_FRAME, NULL, NULL);
 595     if (!model){
 596         av_log(ctx, AV_LOG_ERROR, "Failed to load native model\n");
 597         return DNN_ERROR;
 598     }
 599
 600     native_model = model->model;
 601     tf_model->graph = TF_NewGraph();
 602     tf_model->status = TF_NewStatus();
 603
 604 #define CLEANUP_ON_ERROR(tf_model) \
 605     { \
 606         TF_DeleteTensor(tensor); \
 607         TF_DeleteGraph(tf_model->graph); \
 608         TF_DeleteStatus(tf_model->status); \
 609         av_log(ctx, AV_LOG_ERROR, "Failed to set value or add operator to layer\n"); \
 610         return DNN_ERROR; \
 611     }
 612
 613     op_desc = TF_NewOperation(tf_model->graph, "Placeholder", "x");
 614     TF_SetAttrType(op_desc, "dtype", TF_FLOAT);
 615     TF_SetAttrShape(op_desc, "shape", input_shape, 4);
 616     op = TF_FinishOperation(op_desc, tf_model->status);
 617     if (TF_GetCode(tf_model->status) != TF_OK){
 618         CLEANUP_ON_ERROR(tf_model);
 619     }
 620
 621     op_desc = TF_NewOperation(tf_model->graph, "Const", "transpose_perm");
 622     TF_SetAttrType(op_desc, "dtype", TF_INT32);
 623     tensor = TF_AllocateTensor(TF_INT32, transpose_perm_shape, 1, 4 * sizeof(int32_t));
 624     transpose_perm = (int32_t *)TF_TensorData(tensor);
 625     transpose_perm[0] = 1;
 626     transpose_perm[1] = 2;
 627     transpose_perm[2] = 3;
 628     transpose_perm[3] = 0;
 629     TF_SetAttrTensor(op_desc, "value", tensor, tf_model->status);
 630     if (TF_GetCode(tf_model->status) != TF_OK){
 631         CLEANUP_ON_ERROR(tf_model);
 632     }
 633     transpose_op = TF_FinishOperation(op_desc, tf_model->status);
 634     if (TF_GetCode(tf_model->status) != TF_OK){
 635         CLEANUP_ON_ERROR(tf_model);
 636     }
 637
 638     for (layer = 0; layer < native_model->layers_num; ++layer){
 639         switch (native_model->layers[layer].type){
 640         case DLT_INPUT:
 641             layer_add_res = DNN_SUCCESS;
 642             break;
 643         case DLT_CONV2D:
 644             layer_add_res = add_conv_layer(tf_model, transpose_op, &op,
 645                                            (ConvolutionalParams *)native_model->layers[layer].params, layer);
 646             break;
 647         case DLT_DEPTH_TO_SPACE:
 648             layer_add_res = add_depth_to_space_layer(tf_model, &op,
 649                                                      (DepthToSpaceParams *)native_model->layers[layer].params, layer);
 650             break;
 651         case DLT_MIRROR_PAD:
 652             layer_add_res = add_pad_layer(tf_model, &op,
 653                                           (LayerPadParams *)native_model->layers[layer].params, layer);
 654             break;
 655         case DLT_MAXIMUM:
 656             layer_add_res = add_maximum_layer(tf_model, &op,
 657                                           (DnnLayerMaximumParams *)native_model->layers[layer].params, layer);
 658             break;
 659         default:
 660             CLEANUP_ON_ERROR(tf_model);
 661         }
 662
 663         if (layer_add_res != DNN_SUCCESS){
 664             CLEANUP_ON_ERROR(tf_model);
 665         }
 666     }
 667
 668     op_desc = TF_NewOperation(tf_model->graph, "Identity", "y");
 669     input.oper = op;
 670     input.index = 0;
 671     TF_AddInput(op_desc, input);
 672     TF_FinishOperation(op_desc, tf_model->status);
 673     if (TF_GetCode(tf_model->status) != TF_OK){
 674         CLEANUP_ON_ERROR(tf_model);
 675     }
 676
 677     ff_dnn_free_model_native(&model);
 678
 679     return DNN_SUCCESS;
 680 }
 681
 682 DNNModel *ff_dnn_load_model_tf(const char *model_filename, DNNFunctionType func_type, const char *options, AVFilterContext *filter_ctx)
 683 {
 684     DNNModel *model = NULL;
 685     TFModel *tf_model = NULL;
 686
 687     model = av_mallocz(sizeof(DNNModel));
 688     if (!model){
 689         return NULL;
 690     }
 691
 692     tf_model = av_mallocz(sizeof(TFModel));
 693     if (!tf_model){
 694         av_freep(&model);
 695         return NULL;
 696     }
 697     tf_model->ctx.class = &dnn_tensorflow_class;
 698     tf_model->model = model;
 699
 700     //parse options
 701     av_opt_set_defaults(&tf_model->ctx);
 702     if (av_opt_set_from_string(&tf_model->ctx, options, NULL, "=", "&") < 0) {
 703         av_log(&tf_model->ctx, AV_LOG_ERROR, "Failed to parse options \"%s\"\n", options);
 704         av_freep(&tf_model);
 705         av_freep(&model);
 706         return NULL;
 707     }
 708
 709     if (load_tf_model(tf_model, model_filename) != DNN_SUCCESS){
 710         if (load_native_model(tf_model, model_filename) != DNN_SUCCESS){
 711             av_freep(&tf_model);
 712             av_freep(&model);
 713
 714             return NULL;
 715         }
 716     }
 717
 718     model->model = tf_model;
 719     model->get_input = &get_input_tf;
 720     model->get_output = &get_output_tf;
 721     model->options = options;
 722     model->filter_ctx = filter_ctx;
 723     model->func_type = func_type;
 724
 725     return model;
 726 }
 727
 728 static DNNReturnType execute_model_tf(const DNNModel *model, const char *input_name, AVFrame *in_frame,
 729                                       const char **output_names, uint32_t nb_output, AVFrame *out_frame,
 730                                       int do_ioproc)
 731 {
 732     TF_Output *tf_outputs;
 733     TFModel *tf_model = model->model;
 734     TFContext *ctx = &tf_model->ctx;
 735     DNNData input, output;
 736     TF_Tensor **output_tensors;
 737     TF_Output tf_input;
 738     TF_Tensor *input_tensor;
 739
 740     if (get_input_tf(tf_model, &input, input_name) != DNN_SUCCESS)
 741         return DNN_ERROR;
 742     input.height = in_frame->height;
 743     input.width = in_frame->width;
 744
 745     tf_input.oper = TF_GraphOperationByName(tf_model->graph, input_name);
 746     if (!tf_input.oper){
 747         av_log(ctx, AV_LOG_ERROR, "Could not find \"%s\" in model\n", input_name);
 748         return DNN_ERROR;
 749     }
 750     tf_input.index = 0;
 751     input_tensor = allocate_input_tensor(&input);
 752     if (!input_tensor){
 753         av_log(ctx, AV_LOG_ERROR, "Failed to allocate memory for input tensor\n");
 754         return DNN_ERROR;
 755     }
 756     input.data = (float *)TF_TensorData(input_tensor);
 757
 758     if (do_ioproc) {
 759         if (tf_model->model->frame_pre_proc != NULL) {
 760             tf_model->model->frame_pre_proc(in_frame, &input, tf_model->model->filter_ctx);
 761         } else {
 762             ff_proc_from_frame_to_dnn(in_frame, &input, tf_model->model->func_type, ctx);
 763         }
 764     }
 765
 766     if (nb_output != 1) {
 767         // currently, the filter does not need multiple outputs,
 768         // so we just pending the support until we really need it.
 769         TF_DeleteTensor(input_tensor);
 770         avpriv_report_missing_feature(ctx, "multiple outputs");
 771         return DNN_ERROR;
 772     }
 773
 774     tf_outputs = av_malloc_array(nb_output, sizeof(*tf_outputs));
 775     if (tf_outputs == NULL) {
 776         TF_DeleteTensor(input_tensor);
 777         av_log(ctx, AV_LOG_ERROR, "Failed to allocate memory for *tf_outputs\n"); \
 778         return DNN_ERROR;
 779     }
 780
 781     output_tensors = av_mallocz_array(nb_output, sizeof(*output_tensors));
 782     if (!output_tensors) {
 783         TF_DeleteTensor(input_tensor);
 784         av_freep(&tf_outputs);
 785         av_log(ctx, AV_LOG_ERROR, "Failed to allocate memory for output tensor\n"); \
 786         return DNN_ERROR;
 787     }
 788
 789     for (int i = 0; i < nb_output; ++i) {
 790         tf_outputs[i].oper = TF_GraphOperationByName(tf_model->graph, output_names[i]);
 791         if (!tf_outputs[i].oper) {
 792             TF_DeleteTensor(input_tensor);
 793             av_freep(&tf_outputs);
 794             av_freep(&output_tensors);
 795             av_log(ctx, AV_LOG_ERROR, "Could not find output \"%s\" in model\n", output_names[i]); \
 796             return DNN_ERROR;
 797         }
 798         tf_outputs[i].index = 0;
 799     }
 800
 801     TF_SessionRun(tf_model->session, NULL,
 802                   &tf_input, &input_tensor, 1,
 803                   tf_outputs, output_tensors, nb_output,
 804                   NULL, 0, NULL, tf_model->status);
 805     if (TF_GetCode(tf_model->status) != TF_OK) {
 806         TF_DeleteTensor(input_tensor);
 807         av_freep(&tf_outputs);
 808         av_freep(&output_tensors);
 809         av_log(ctx, AV_LOG_ERROR, "Failed to run session when executing model\n");
 810         return DNN_ERROR;
 811     }
 812
 813     for (uint32_t i = 0; i < nb_output; ++i) {
 814         output.height = TF_Dim(output_tensors[i], 1);
 815         output.width = TF_Dim(output_tensors[i], 2);
 816         output.channels = TF_Dim(output_tensors[i], 3);
 817         output.data = TF_TensorData(output_tensors[i]);
 818         output.dt = TF_TensorType(output_tensors[i]);
 819
 820         if (do_ioproc) {
 821             if (tf_model->model->frame_post_proc != NULL) {
 822                 tf_model->model->frame_post_proc(out_frame, &output, tf_model->model->filter_ctx);
 823             } else {
 824                 ff_proc_from_dnn_to_frame(out_frame, &output, ctx);
 825             }
 826         } else {
 827             out_frame->width = output.width;
 828             out_frame->height = output.height;
 829         }
 830     }
 831
 832     for (uint32_t i = 0; i < nb_output; ++i) {
 833         if (output_tensors[i]) {
 834             TF_DeleteTensor(output_tensors[i]);
 835         }
 836     }
 837     TF_DeleteTensor(input_tensor);
 838     av_freep(&output_tensors);
 839     av_freep(&tf_outputs);
 840     return DNN_SUCCESS;
 841 }
 842
 843 DNNReturnType ff_dnn_execute_model_tf(const DNNModel *model, const char *input_name, AVFrame *in_frame,
 844                                       const char **output_names, uint32_t nb_output, AVFrame *out_frame)
 845 {
 846     TFModel *tf_model = model->model;
 847     TFContext *ctx = &tf_model->ctx;
 848
 849     if (!in_frame) {
 850         av_log(ctx, AV_LOG_ERROR, "in frame is NULL when execute model.\n");
 851         return DNN_ERROR;
 852     }
 853
 854     if (!out_frame) {
 855         av_log(ctx, AV_LOG_ERROR, "out frame is NULL when execute model.\n");
 856         return DNN_ERROR;
 857     }
 858
 859     return execute_model_tf(model, input_name, in_frame, output_names, nb_output, out_frame, 1);
 860 }
 861
 862 void ff_dnn_free_model_tf(DNNModel **model)
 863 {
 864     TFModel *tf_model;
 865
 866     if (*model){
 867         tf_model = (*model)->model;
 868         if (tf_model->graph){
 869             TF_DeleteGraph(tf_model->graph);
 870         }
 871         if (tf_model->session){
 872             TF_CloseSession(tf_model->session, tf_model->status);
 873             TF_DeleteSession(tf_model->session, tf_model->status);
 874         }
 875         if (tf_model->status){
 876             TF_DeleteStatus(tf_model->status);
 877         }
 878         av_freep(&tf_model);
 879         av_freep(model);
 880     }
 881 }