2 * Copyright (c) 2018 Sergey Lavrushkin
4 * This file is part of FFmpeg.
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.
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.
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
23 * DNN tensorflow backend implementation.
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 "dnn_backend_native_layer_pad.h"
34 #include <tensorflow/c/c_api.h>
36 typedef struct TFModel{
41 TF_Tensor *input_tensor;
43 TF_Tensor **output_tensors;
47 static void free_buffer(void *data, size_t length)
52 static TF_Buffer *read_graph(const char *model_filename)
55 unsigned char *graph_data = NULL;
56 AVIOContext *model_file_context;
57 long size, bytes_read;
59 if (avio_open(&model_file_context, model_filename, AVIO_FLAG_READ) < 0){
63 size = avio_size(model_file_context);
65 graph_data = av_malloc(size);
67 avio_closep(&model_file_context);
70 bytes_read = avio_read(model_file_context, graph_data, size);
71 avio_closep(&model_file_context);
72 if (bytes_read != size){
73 av_freep(&graph_data);
77 graph_buf = TF_NewBuffer();
78 graph_buf->data = (void *)graph_data;
79 graph_buf->length = size;
80 graph_buf->data_deallocator = free_buffer;
85 static TF_Tensor *allocate_input_tensor(const DNNInputData *input)
89 int64_t input_dims[] = {1, input->height, input->width, input->channels};
100 av_assert0(!"should not reach here");
103 return TF_AllocateTensor(dt, input_dims, 4,
104 input_dims[1] * input_dims[2] * input_dims[3] * size);
107 static DNNReturnType set_input_output_tf(void *model, DNNInputData *input, const char *input_name, const char **output_names, uint32_t nb_output)
109 TFModel *tf_model = (TFModel *)model;
110 TF_SessionOptions *sess_opts;
111 const TF_Operation *init_op = TF_GraphOperationByName(tf_model->graph, "init");
114 tf_model->input.oper = TF_GraphOperationByName(tf_model->graph, input_name);
115 if (!tf_model->input.oper){
118 tf_model->input.index = 0;
119 if (tf_model->input_tensor){
120 TF_DeleteTensor(tf_model->input_tensor);
122 tf_model->input_tensor = allocate_input_tensor(input);
123 if (!tf_model->input_tensor){
126 input->data = (float *)TF_TensorData(tf_model->input_tensor);
132 av_freep(&tf_model->outputs);
133 tf_model->outputs = av_malloc_array(nb_output, sizeof(*tf_model->outputs));
134 if (!tf_model->outputs)
136 for (int i = 0; i < nb_output; ++i) {
137 tf_model->outputs[i].oper = TF_GraphOperationByName(tf_model->graph, output_names[i]);
138 if (!tf_model->outputs[i].oper){
139 av_freep(&tf_model->outputs);
142 tf_model->outputs[i].index = 0;
145 if (tf_model->output_tensors) {
146 for (uint32_t i = 0; i < tf_model->nb_output; ++i) {
147 if (tf_model->output_tensors[i]) {
148 TF_DeleteTensor(tf_model->output_tensors[i]);
149 tf_model->output_tensors[i] = NULL;
153 av_freep(&tf_model->output_tensors);
154 tf_model->output_tensors = av_mallocz_array(nb_output, sizeof(*tf_model->output_tensors));
155 if (!tf_model->output_tensors) {
156 av_freep(&tf_model->outputs);
160 tf_model->nb_output = nb_output;
162 if (tf_model->session){
163 TF_CloseSession(tf_model->session, tf_model->status);
164 TF_DeleteSession(tf_model->session, tf_model->status);
167 sess_opts = TF_NewSessionOptions();
168 tf_model->session = TF_NewSession(tf_model->graph, sess_opts, tf_model->status);
169 TF_DeleteSessionOptions(sess_opts);
170 if (TF_GetCode(tf_model->status) != TF_OK)
175 // Run initialization operation with name "init" if it is present in graph
177 TF_SessionRun(tf_model->session, NULL,
180 &init_op, 1, NULL, tf_model->status);
181 if (TF_GetCode(tf_model->status) != TF_OK)
190 static DNNReturnType load_tf_model(TFModel *tf_model, const char *model_filename)
192 TF_Buffer *graph_def;
193 TF_ImportGraphDefOptions *graph_opts;
195 graph_def = read_graph(model_filename);
199 tf_model->graph = TF_NewGraph();
200 tf_model->status = TF_NewStatus();
201 graph_opts = TF_NewImportGraphDefOptions();
202 TF_GraphImportGraphDef(tf_model->graph, graph_def, graph_opts, tf_model->status);
203 TF_DeleteImportGraphDefOptions(graph_opts);
204 TF_DeleteBuffer(graph_def);
205 if (TF_GetCode(tf_model->status) != TF_OK){
206 TF_DeleteGraph(tf_model->graph);
207 TF_DeleteStatus(tf_model->status);
214 #define NAME_BUFFER_SIZE 256
216 static DNNReturnType add_conv_layer(TFModel *tf_model, TF_Operation *transpose_op, TF_Operation **cur_op,
217 ConvolutionalParams* params, const int layer)
220 TF_OperationDescription *op_desc;
222 int64_t strides[] = {1, 1, 1, 1};
226 char name_buffer[NAME_BUFFER_SIZE];
229 size = params->input_num * params->output_num * params->kernel_size * params->kernel_size;
232 snprintf(name_buffer, NAME_BUFFER_SIZE, "conv_kernel%d", layer);
233 op_desc = TF_NewOperation(tf_model->graph, "Const", name_buffer);
234 TF_SetAttrType(op_desc, "dtype", TF_FLOAT);
235 dims[0] = params->output_num;
236 dims[1] = params->kernel_size;
237 dims[2] = params->kernel_size;
238 dims[3] = params->input_num;
240 tensor = TF_AllocateTensor(TF_FLOAT, dims, dims_len, size * sizeof(float));
241 memcpy(TF_TensorData(tensor), params->kernel, size * sizeof(float));
242 TF_SetAttrTensor(op_desc, "value", tensor, tf_model->status);
243 if (TF_GetCode(tf_model->status) != TF_OK){
246 op = TF_FinishOperation(op_desc, tf_model->status);
247 if (TF_GetCode(tf_model->status) != TF_OK){
251 snprintf(name_buffer, NAME_BUFFER_SIZE, "transpose%d", layer);
252 op_desc = TF_NewOperation(tf_model->graph, "Transpose", name_buffer);
254 TF_AddInput(op_desc, input);
255 input.oper = transpose_op;
256 TF_AddInput(op_desc, input);
257 TF_SetAttrType(op_desc, "T", TF_FLOAT);
258 TF_SetAttrType(op_desc, "Tperm", TF_INT32);
259 op = TF_FinishOperation(op_desc, tf_model->status);
260 if (TF_GetCode(tf_model->status) != TF_OK){
264 snprintf(name_buffer, NAME_BUFFER_SIZE, "conv2d%d", layer);
265 op_desc = TF_NewOperation(tf_model->graph, "Conv2D", name_buffer);
266 input.oper = *cur_op;
267 TF_AddInput(op_desc, input);
269 TF_AddInput(op_desc, input);
270 TF_SetAttrType(op_desc, "T", TF_FLOAT);
271 TF_SetAttrIntList(op_desc, "strides", strides, 4);
272 TF_SetAttrString(op_desc, "padding", "VALID", 5);
273 *cur_op = TF_FinishOperation(op_desc, tf_model->status);
274 if (TF_GetCode(tf_model->status) != TF_OK){
278 snprintf(name_buffer, NAME_BUFFER_SIZE, "conv_biases%d", layer);
279 op_desc = TF_NewOperation(tf_model->graph, "Const", name_buffer);
280 TF_SetAttrType(op_desc, "dtype", TF_FLOAT);
281 dims[0] = params->output_num;
283 tensor = TF_AllocateTensor(TF_FLOAT, dims, dims_len, params->output_num * sizeof(float));
284 memcpy(TF_TensorData(tensor), params->biases, params->output_num * sizeof(float));
285 TF_SetAttrTensor(op_desc, "value", tensor, tf_model->status);
286 if (TF_GetCode(tf_model->status) != TF_OK){
289 op = TF_FinishOperation(op_desc, tf_model->status);
290 if (TF_GetCode(tf_model->status) != TF_OK){
294 snprintf(name_buffer, NAME_BUFFER_SIZE, "bias_add%d", layer);
295 op_desc = TF_NewOperation(tf_model->graph, "BiasAdd", name_buffer);
296 input.oper = *cur_op;
297 TF_AddInput(op_desc, input);
299 TF_AddInput(op_desc, input);
300 TF_SetAttrType(op_desc, "T", TF_FLOAT);
301 *cur_op = TF_FinishOperation(op_desc, tf_model->status);
302 if (TF_GetCode(tf_model->status) != TF_OK){
306 snprintf(name_buffer, NAME_BUFFER_SIZE, "activation%d", layer);
307 switch (params->activation){
309 op_desc = TF_NewOperation(tf_model->graph, "Relu", name_buffer);
312 op_desc = TF_NewOperation(tf_model->graph, "Tanh", name_buffer);
315 op_desc = TF_NewOperation(tf_model->graph, "Sigmoid", name_buffer);
320 input.oper = *cur_op;
321 TF_AddInput(op_desc, input);
322 TF_SetAttrType(op_desc, "T", TF_FLOAT);
323 *cur_op = TF_FinishOperation(op_desc, tf_model->status);
324 if (TF_GetCode(tf_model->status) != TF_OK){
331 static DNNReturnType add_depth_to_space_layer(TFModel *tf_model, TF_Operation **cur_op,
332 DepthToSpaceParams *params, const int layer)
334 TF_OperationDescription *op_desc;
336 char name_buffer[NAME_BUFFER_SIZE];
338 snprintf(name_buffer, NAME_BUFFER_SIZE, "depth_to_space%d", layer);
339 op_desc = TF_NewOperation(tf_model->graph, "DepthToSpace", name_buffer);
340 input.oper = *cur_op;
342 TF_AddInput(op_desc, input);
343 TF_SetAttrType(op_desc, "T", TF_FLOAT);
344 TF_SetAttrInt(op_desc, "block_size", params->block_size);
345 *cur_op = TF_FinishOperation(op_desc, tf_model->status);
346 if (TF_GetCode(tf_model->status) != TF_OK){
353 static DNNReturnType add_pad_layer(TFModel *tf_model, TF_Operation **cur_op,
354 LayerPadParams *params, const int layer)
358 TF_OperationDescription *op_desc;
361 int64_t pads_shape[] = {4, 2};
363 char name_buffer[NAME_BUFFER_SIZE];
364 snprintf(name_buffer, NAME_BUFFER_SIZE, "pad%d", layer);
366 op_desc = TF_NewOperation(tf_model->graph, "Const", name_buffer);
367 TF_SetAttrType(op_desc, "dtype", TF_INT32);
368 tensor = TF_AllocateTensor(TF_INT32, pads_shape, 2, 4 * 2 * sizeof(int32_t));
369 pads = (int32_t *)TF_TensorData(tensor);
370 pads[0] = params->paddings[0][0];
371 pads[1] = params->paddings[0][1];
372 pads[2] = params->paddings[1][0];
373 pads[3] = params->paddings[1][1];
374 pads[4] = params->paddings[2][0];
375 pads[5] = params->paddings[2][1];
376 pads[6] = params->paddings[3][0];
377 pads[7] = params->paddings[3][1];
378 TF_SetAttrTensor(op_desc, "value", tensor, tf_model->status);
379 if (TF_GetCode(tf_model->status) != TF_OK){
382 op = TF_FinishOperation(op_desc, tf_model->status);
383 if (TF_GetCode(tf_model->status) != TF_OK){
387 op_desc = TF_NewOperation(tf_model->graph, "MirrorPad", "mirror_pad");
388 input.oper = *cur_op;
390 TF_AddInput(op_desc, input);
392 TF_AddInput(op_desc, input);
393 TF_SetAttrType(op_desc, "T", TF_FLOAT);
394 TF_SetAttrType(op_desc, "Tpaddings", TF_INT32);
395 TF_SetAttrString(op_desc, "mode", "SYMMETRIC", 9);
396 *cur_op = TF_FinishOperation(op_desc, tf_model->status);
397 if (TF_GetCode(tf_model->status) != TF_OK){
404 static DNNReturnType load_native_model(TFModel *tf_model, const char *model_filename)
407 TF_OperationDescription *op_desc;
409 TF_Operation *transpose_op;
412 int32_t *transpose_perm;
413 int64_t transpose_perm_shape[] = {4};
414 int64_t input_shape[] = {1, -1, -1, -1};
415 DNNReturnType layer_add_res;
416 DNNModel *native_model = NULL;
417 ConvolutionalNetwork *conv_network;
419 native_model = ff_dnn_load_model_native(model_filename);
424 conv_network = (ConvolutionalNetwork *)native_model->model;
425 tf_model->graph = TF_NewGraph();
426 tf_model->status = TF_NewStatus();
428 #define CLEANUP_ON_ERROR(tf_model) \
430 TF_DeleteGraph(tf_model->graph); \
431 TF_DeleteStatus(tf_model->status); \
435 op_desc = TF_NewOperation(tf_model->graph, "Placeholder", "x");
436 TF_SetAttrType(op_desc, "dtype", TF_FLOAT);
437 TF_SetAttrShape(op_desc, "shape", input_shape, 4);
438 op = TF_FinishOperation(op_desc, tf_model->status);
439 if (TF_GetCode(tf_model->status) != TF_OK){
440 CLEANUP_ON_ERROR(tf_model);
443 op_desc = TF_NewOperation(tf_model->graph, "Const", "transpose_perm");
444 TF_SetAttrType(op_desc, "dtype", TF_INT32);
445 tensor = TF_AllocateTensor(TF_INT32, transpose_perm_shape, 1, 4 * sizeof(int32_t));
446 transpose_perm = (int32_t *)TF_TensorData(tensor);
447 transpose_perm[0] = 1;
448 transpose_perm[1] = 2;
449 transpose_perm[2] = 3;
450 transpose_perm[3] = 0;
451 TF_SetAttrTensor(op_desc, "value", tensor, tf_model->status);
452 if (TF_GetCode(tf_model->status) != TF_OK){
453 CLEANUP_ON_ERROR(tf_model);
455 transpose_op = TF_FinishOperation(op_desc, tf_model->status);
457 for (layer = 0; layer < conv_network->layers_num; ++layer){
458 switch (conv_network->layers[layer].type){
460 layer_add_res = DNN_SUCCESS;
463 layer_add_res = add_conv_layer(tf_model, transpose_op, &op,
464 (ConvolutionalParams *)conv_network->layers[layer].params, layer);
467 layer_add_res = add_depth_to_space_layer(tf_model, &op,
468 (DepthToSpaceParams *)conv_network->layers[layer].params, layer);
471 layer_add_res = add_pad_layer(tf_model, &op,
472 (LayerPadParams *)conv_network->layers[layer].params, layer);
475 CLEANUP_ON_ERROR(tf_model);
478 if (layer_add_res != DNN_SUCCESS){
479 CLEANUP_ON_ERROR(tf_model);
483 op_desc = TF_NewOperation(tf_model->graph, "Identity", "y");
486 TF_AddInput(op_desc, input);
487 TF_FinishOperation(op_desc, tf_model->status);
488 if (TF_GetCode(tf_model->status) != TF_OK){
489 CLEANUP_ON_ERROR(tf_model);
492 ff_dnn_free_model_native(&native_model);
497 DNNModel *ff_dnn_load_model_tf(const char *model_filename)
499 DNNModel *model = NULL;
500 TFModel *tf_model = NULL;
502 model = av_malloc(sizeof(DNNModel));
507 tf_model = av_mallocz(sizeof(TFModel));
513 if (load_tf_model(tf_model, model_filename) != DNN_SUCCESS){
514 if (load_native_model(tf_model, model_filename) != DNN_SUCCESS){
522 model->model = (void *)tf_model;
523 model->set_input_output = &set_input_output_tf;
530 DNNReturnType ff_dnn_execute_model_tf(const DNNModel *model, DNNData *outputs, uint32_t nb_output)
532 TFModel *tf_model = (TFModel *)model->model;
533 uint32_t nb = FFMIN(nb_output, tf_model->nb_output);
537 av_assert0(tf_model->output_tensors);
538 for (uint32_t i = 0; i < tf_model->nb_output; ++i) {
539 if (tf_model->output_tensors[i]) {
540 TF_DeleteTensor(tf_model->output_tensors[i]);
541 tf_model->output_tensors[i] = NULL;
545 TF_SessionRun(tf_model->session, NULL,
546 &tf_model->input, &tf_model->input_tensor, 1,
547 tf_model->outputs, tf_model->output_tensors, nb,
548 NULL, 0, NULL, tf_model->status);
550 if (TF_GetCode(tf_model->status) != TF_OK){
554 for (uint32_t i = 0; i < nb; ++i) {
555 outputs[i].height = TF_Dim(tf_model->output_tensors[i], 1);
556 outputs[i].width = TF_Dim(tf_model->output_tensors[i], 2);
557 outputs[i].channels = TF_Dim(tf_model->output_tensors[i], 3);
558 outputs[i].data = TF_TensorData(tf_model->output_tensors[i]);
564 void ff_dnn_free_model_tf(DNNModel **model)
569 tf_model = (TFModel *)(*model)->model;
570 if (tf_model->graph){
571 TF_DeleteGraph(tf_model->graph);
573 if (tf_model->session){
574 TF_CloseSession(tf_model->session, tf_model->status);
575 TF_DeleteSession(tf_model->session, tf_model->status);
577 if (tf_model->status){
578 TF_DeleteStatus(tf_model->status);
580 if (tf_model->input_tensor){
581 TF_DeleteTensor(tf_model->input_tensor);
583 if (tf_model->output_tensors) {
584 for (uint32_t i = 0; i < tf_model->nb_output; ++i) {
585 if (tf_model->output_tensors[i]) {
586 TF_DeleteTensor(tf_model->output_tensors[i]);
587 tf_model->output_tensors[i] = NULL;
591 av_freep(&tf_model->outputs);
592 av_freep(&tf_model->output_tensors);