and also change CONV to DLT_CONV2D for better description
Showing the changes to libavfilter (3 of 3 files). Show all
Unified · Split
| @@ -188,8 +188,9 @@ DNNModel *ff_dnn_load_model_native(const char *model_filename) · 187 unmodified lines | |||
| 188 | for (layer = 0; layer < network->layers_num; ++layer){ | 188 | for (layer = 0; layer < network->layers_num; ++layer){ |
| 189 | layer_type = (int32_t)avio_rl32(model_file_context); | 189 | layer_type = (int32_t)avio_rl32(model_file_context); |
| 190 | dnn_size += 4; | 190 | dnn_size += 4; |
| 191 | network->layers[layer].type = layer_type; | ||
| 191 | switch (layer_type){ | 192 | switch (layer_type){ |
| 192 | case CONV: | 193 | case DLT_CONV2D: |
| 193 | conv_params = av_malloc(sizeof(ConvolutionalParams)); | 194 | conv_params = av_malloc(sizeof(ConvolutionalParams)); |
| 194 | if (!conv_params){ | 195 | if (!conv_params){ |
| 195 | avio_closep(&model_file_context); | 196 | avio_closep(&model_file_context); |
| @@ -231,10 +232,9 @@ DNNModel *ff_dnn_load_model_native(const char *model_filename) · 35 unmodified lines | |||
| 231 | network->layers[layer].input_operand_indexes[0] = (int32_t)avio_rl32(model_file_context); | 232 | network->layers[layer].input_operand_indexes[0] = (int32_t)avio_rl32(model_file_context); |
| 232 | network->layers[layer].output_operand_index = (int32_t)avio_rl32(model_file_context); | 233 | network->layers[layer].output_operand_index = (int32_t)avio_rl32(model_file_context); |
| 233 | dnn_size += 8; | 234 | dnn_size += 8; |
| 234 | network->layers[layer].type = CONV; | ||
| 235 | network->layers[layer].params = conv_params; | 235 | network->layers[layer].params = conv_params; |
| 236 | break; | 236 | break; |
| 237 | case DEPTH_TO_SPACE: | 237 | case DLT_DEPTH_TO_SPACE: |
| 238 | depth_to_space_params = av_malloc(sizeof(DepthToSpaceParams)); | 238 | depth_to_space_params = av_malloc(sizeof(DepthToSpaceParams)); |
| 239 | if (!depth_to_space_params){ | 239 | if (!depth_to_space_params){ |
| 240 | avio_closep(&model_file_context); | 240 | avio_closep(&model_file_context); |
| @@ -246,10 +246,9 @@ DNNModel *ff_dnn_load_model_native(const char *model_filename) · 5 unmodified lines | |||
| 246 | network->layers[layer].input_operand_indexes[0] = (int32_t)avio_rl32(model_file_context); | 246 | network->layers[layer].input_operand_indexes[0] = (int32_t)avio_rl32(model_file_context); |
| 247 | network->layers[layer].output_operand_index = (int32_t)avio_rl32(model_file_context); | 247 | network->layers[layer].output_operand_index = (int32_t)avio_rl32(model_file_context); |
| 248 | dnn_size += 8; | 248 | dnn_size += 8; |
| 249 | network->layers[layer].type = DEPTH_TO_SPACE; | ||
| 250 | network->layers[layer].params = depth_to_space_params; | 249 | network->layers[layer].params = depth_to_space_params; |
| 251 | break; | 250 | break; |
| 252 | case MIRROR_PAD: | 251 | case DLT_MIRROR_PAD: |
| 253 | pad_params = av_malloc(sizeof(LayerPadParams)); | 252 | pad_params = av_malloc(sizeof(LayerPadParams)); |
| 254 | if (!pad_params){ | 253 | if (!pad_params){ |
| 255 | avio_closep(&model_file_context); | 254 | avio_closep(&model_file_context); |
| @@ -266,10 +265,9 @@ DNNModel *ff_dnn_load_model_native(const char *model_filename) · 10 unmodified lines | |||
| 266 | network->layers[layer].input_operand_indexes[0] = (int32_t)avio_rl32(model_file_context); | 265 | network->layers[layer].input_operand_indexes[0] = (int32_t)avio_rl32(model_file_context); |
| 267 | network->layers[layer].output_operand_index = (int32_t)avio_rl32(model_file_context); | 266 | network->layers[layer].output_operand_index = (int32_t)avio_rl32(model_file_context); |
| 268 | dnn_size += 8; | 267 | dnn_size += 8; |
| 269 | network->layers[layer].type = MIRROR_PAD; | ||
| 270 | network->layers[layer].params = pad_params; | 268 | network->layers[layer].params = pad_params; |
| 271 | break; | 269 | break; |
| 272 | case MAXIMUM: | 270 | case DLT_MAXIMUM: |
| 273 | maximum_params = av_malloc(sizeof(*maximum_params)); | 271 | maximum_params = av_malloc(sizeof(*maximum_params)); |
| 274 | if (!maximum_params){ | 272 | if (!maximum_params){ |
| 275 | avio_closep(&model_file_context); | 273 | avio_closep(&model_file_context); |
| @@ -278,7 +276,6 @@ DNNModel *ff_dnn_load_model_native(const char *model_filename) · 2 unmodified lines | |||
| 278 | } | 276 | } |
| 279 | maximum_params->val.u32 = avio_rl32(model_file_context); | 277 | maximum_params->val.u32 = avio_rl32(model_file_context); |
| 280 | dnn_size += 4; | 278 | dnn_size += 4; |
| 281 | network->layers[layer].type = MAXIMUM; | ||
| 282 | network->layers[layer].params = maximum_params; | 279 | network->layers[layer].params = maximum_params; |
| 283 | network->layers[layer].input_operand_indexes[0] = (int32_t)avio_rl32(model_file_context); | 280 | network->layers[layer].input_operand_indexes[0] = (int32_t)avio_rl32(model_file_context); |
| 284 | network->layers[layer].output_operand_index = (int32_t)avio_rl32(model_file_context); | 281 | network->layers[layer].output_operand_index = (int32_t)avio_rl32(model_file_context); |
| @@ -347,27 +344,27 @@ DNNReturnType ff_dnn_execute_model_native(const DNNModel *model, DNNData *output · 62 unmodified lines | |||
| 347 | 344 | ||
| 348 | for (layer = 0; layer < network->layers_num; ++layer){ | 345 | for (layer = 0; layer < network->layers_num; ++layer){ |
| 349 | switch (network->layers[layer].type){ | 346 | switch (network->layers[layer].type){ |
| 350 | case CONV: | 347 | case DLT_CONV2D: |
| 351 | conv_params = (ConvolutionalParams *)network->layers[layer].params; | 348 | conv_params = (ConvolutionalParams *)network->layers[layer].params; |
| 352 | convolve(network->operands, network->layers[layer].input_operand_indexes, | 349 | convolve(network->operands, network->layers[layer].input_operand_indexes, |
| 353 | network->layers[layer].output_operand_index, conv_params); | 350 | network->layers[layer].output_operand_index, conv_params); |
| 354 | break; | 351 | break; |
| 355 | case DEPTH_TO_SPACE: | 352 | case DLT_DEPTH_TO_SPACE: |
| 356 | depth_to_space_params = (DepthToSpaceParams *)network->layers[layer].params; | 353 | depth_to_space_params = (DepthToSpaceParams *)network->layers[layer].params; |
| 357 | depth_to_space(network->operands, network->layers[layer].input_operand_indexes, | 354 | depth_to_space(network->operands, network->layers[layer].input_operand_indexes, |
| 358 | network->layers[layer].output_operand_index, depth_to_space_params->block_size); | 355 | network->layers[layer].output_operand_index, depth_to_space_params->block_size); |
| 359 | break; | 356 | break; |
| 360 | case MIRROR_PAD: | 357 | case DLT_MIRROR_PAD: |
| 361 | pad_params = (LayerPadParams *)network->layers[layer].params; | 358 | pad_params = (LayerPadParams *)network->layers[layer].params; |
| 362 | dnn_execute_layer_pad(network->operands, network->layers[layer].input_operand_indexes, | 359 | dnn_execute_layer_pad(network->operands, network->layers[layer].input_operand_indexes, |
| 363 | network->layers[layer].output_operand_index, pad_params); | 360 | network->layers[layer].output_operand_index, pad_params); |
| 364 | break; | 361 | break; |
| 365 | case MAXIMUM: | 362 | case DLT_MAXIMUM: |
| 366 | maximum_params = (DnnLayerMaximumParams *)network->layers[layer].params; | 363 | maximum_params = (DnnLayerMaximumParams *)network->layers[layer].params; |
| 367 | dnn_execute_layer_maximum(network->operands, network->layers[layer].input_operand_indexes, | 364 | dnn_execute_layer_maximum(network->operands, network->layers[layer].input_operand_indexes, |
| 368 | network->layers[layer].output_operand_index, maximum_params); | 365 | network->layers[layer].output_operand_index, maximum_params); |
| 369 | break; | 366 | break; |
| 370 | case INPUT: | 367 | case DLT_INPUT: |
| 371 | return DNN_ERROR; | 368 | return DNN_ERROR; |
| 372 | } | 369 | } |
| 373 | } | 370 | } |
| @@ -408,7 +405,7 @@ void ff_dnn_free_model_native(DNNModel **model) · 34 unmodified lines | |||
| 408 | { | 405 | { |
| 409 | network = (ConvolutionalNetwork *)(*model)->model; | 406 | network = (ConvolutionalNetwork *)(*model)->model; |
| 410 | for (layer = 0; layer < network->layers_num; ++layer){ | 407 | for (layer = 0; layer < network->layers_num; ++layer){ |
| 411 | if (network->layers[layer].type == CONV){ | 408 | if (network->layers[layer].type == DLT_CONV2D){ |
| 412 | conv_params = (ConvolutionalParams *)network->layers[layer].params; | 409 | conv_params = (ConvolutionalParams *)network->layers[layer].params; |
| 413 | av_freep(&conv_params->kernel); | 410 | av_freep(&conv_params->kernel); |
| 414 | av_freep(&conv_params->biases); | 411 | av_freep(&conv_params->biases); |
| @@ -30,7 +30,17 @@ · 29 unmodified lines | |||
| 30 | #include "../dnn_interface.h" | 30 | #include "../dnn_interface.h" |
| 31 | #include "libavformat/avio.h" | 31 | #include "libavformat/avio.h" |
| 32 | 32 | ||
| 33 | typedef enum {INPUT = 0, CONV = 1, DEPTH_TO_SPACE = 2, MIRROR_PAD = 3, MAXIMUM = 4} DNNLayerType; | 33 | /** |
| 34 | * the enum value of DNNLayerType should not be changed, | ||
| 35 | * the same values are used in convert_from_tensorflow.py | ||
| 36 | */ | ||
| 37 | typedef enum { | ||
| 38 | DLT_INPUT = 0, | ||
| 39 | DLT_CONV2D = 1, | ||
| 40 | DLT_DEPTH_TO_SPACE = 2, | ||
| 41 | DLT_MIRROR_PAD = 3, | ||
| 42 | DLT_MAXIMUM = 4 | ||
| 43 | } DNNLayerType; | ||
| 34 | 44 | ||
| 35 | typedef enum {DOT_INPUT = 1, DOT_OUTPUT = 2, DOT_INTERMEDIATE = DOT_INPUT | DOT_INPUT} DNNOperandType; | 45 | typedef enum {DOT_INPUT = 1, DOT_OUTPUT = 2, DOT_INTERMEDIATE = DOT_INPUT | DOT_INPUT} DNNOperandType; |
| 36 | 46 | ||
| @@ -499,22 +499,22 @@ static DNNReturnType load_native_model(TFModel *tf_model, const char *model_file · 498 unmodified lines | |||
| 499 | 499 | ||
| 500 | for (layer = 0; layer < conv_network->layers_num; ++layer){ | 500 | for (layer = 0; layer < conv_network->layers_num; ++layer){ |
| 501 | switch (conv_network->layers[layer].type){ | 501 | switch (conv_network->layers[layer].type){ |
| 502 | case INPUT: | 502 | case DLT_INPUT: |
| 503 | layer_add_res = DNN_SUCCESS; | 503 | layer_add_res = DNN_SUCCESS; |
| 504 | break; | 504 | break; |
| 505 | case CONV: | 505 | case DLT_CONV2D: |
| 506 | layer_add_res = add_conv_layer(tf_model, transpose_op, &op, | 506 | layer_add_res = add_conv_layer(tf_model, transpose_op, &op, |
| 507 | (ConvolutionalParams *)conv_network->layers[layer].params, layer); | 507 | (ConvolutionalParams *)conv_network->layers[layer].params, layer); |
| 508 | break; | 508 | break; |
| 509 | case DEPTH_TO_SPACE: | 509 | case DLT_DEPTH_TO_SPACE: |
| 510 | layer_add_res = add_depth_to_space_layer(tf_model, &op, | 510 | layer_add_res = add_depth_to_space_layer(tf_model, &op, |
| 511 | (DepthToSpaceParams *)conv_network->layers[layer].params, layer); | 511 | (DepthToSpaceParams *)conv_network->layers[layer].params, layer); |
| 512 | break; | 512 | break; |
| 513 | case MIRROR_PAD: | 513 | case DLT_MIRROR_PAD: |
| 514 | layer_add_res = add_pad_layer(tf_model, &op, | 514 | layer_add_res = add_pad_layer(tf_model, &op, |
| 515 | (LayerPadParams *)conv_network->layers[layer].params, layer); | 515 | (LayerPadParams *)conv_network->layers[layer].params, layer); |
| 516 | break; | 516 | break; |
| 517 | case MAXIMUM: | 517 | case DLT_MAXIMUM: |
| 518 | layer_add_res = add_maximum_layer(tf_model, &op, | 518 | layer_add_res = add_maximum_layer(tf_model, &op, |
| 519 | (DnnLayerMaximumParams *)conv_network->layers[layer].params, layer); | 519 | (DnnLayerMaximumParams *)conv_network->layers[layer].params, layer); |
| 520 | break; | 520 | break; |