libavfilter/dnn: support multiple outputs for tensorflow model

some models such as ssd, yolo have more than one output.

the clean up code in this patch is a little complex, it is because
that set_input_output_tf could be called for many times together
with ff_dnn_execute_model_tf, we have to clean resources for the
case that the two interfaces are called interleaved.
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Pedro Arthur <bygrandao@gmail.com>

libavfilter/dnn_backend_native.c

@@ -25,7 +25,7 @@
 
 #include "dnn_backend_native.h"
 
-static DNNReturnType set_input_output_native(void *model, DNNData *input, const char *input_name, const char *output_name)
+static DNNReturnType set_input_output_native(void *model, DNNData *input, const char *input_name, const char **output_names, uint32_t nb_output)
 {
     ConvolutionalNetwork *network = (ConvolutionalNetwork *)model;
     InputParams *input_params;
@@ -275,7 +275,7 @@ static void depth_to_space(const float *input, float *output, int block_size, in
     }
 }
 
-DNNReturnType ff_dnn_execute_model_native(const DNNModel *model, DNNData *output)
+DNNReturnType ff_dnn_execute_model_native(const DNNModel *model, DNNData *outputs, uint32_t nb_output)
 {
     ConvolutionalNetwork *network = (ConvolutionalNetwork *)model->model;
     int cur_width, cur_height, cur_channels;
@@ -317,10 +317,13 @@ DNNReturnType ff_dnn_execute_model_native(const DNNModel *model, DNNData *output
         }
     }
 
-    output->data = network->layers[network->layers_num - 1].output;
-    output->height = cur_height;
-    output->width = cur_width;
-    output->channels = cur_channels;
+    // native mode does not support multiple outputs yet
+    if (nb_output > 1)
+        return DNN_ERROR;
+    outputs[0].data = network->layers[network->layers_num - 1].output;
+    outputs[0].height = cur_height;
+    outputs[0].width = cur_width;
+    outputs[0].channels = cur_channels;
 
     return DNN_SUCCESS;
 }

libavfilter/dnn_backend_native.h

@@ -63,7 +63,7 @@ typedef struct ConvolutionalNetwork{
 
 DNNModel *ff_dnn_load_model_native(const char *model_filename);
 
-DNNReturnType ff_dnn_execute_model_native(const DNNModel *model, DNNData *output);
+DNNReturnType ff_dnn_execute_model_native(const DNNModel *model, DNNData *outputs, uint32_t nb_output);
 
 void ff_dnn_free_model_native(DNNModel **model);
 

libavfilter/dnn_backend_tf.c

@@ -26,6 +26,7 @@
 #include "dnn_backend_tf.h"
 #include "dnn_backend_native.h"
 #include "libavformat/avio.h"
+#include "libavutil/avassert.h"
 
 #include <tensorflow/c/c_api.h>
 
@@ -33,9 +34,11 @@ typedef struct TFModel{
     TF_Graph *graph;
     TF_Session *session;
     TF_Status *status;
-    TF_Output input, output;
+    TF_Output input;
     TF_Tensor *input_tensor;
-    TF_Tensor *output_tensor;
+    TF_Output *outputs;
+    TF_Tensor **output_tensors;
+    uint32_t nb_output;
 } TFModel;
 
 static void free_buffer(void *data, size_t length)
@@ -76,7 +79,7 @@ static TF_Buffer *read_graph(const char *model_filename)
     return graph_buf;
 }
 
-static DNNReturnType set_input_output_tf(void *model, DNNData *input, const char *input_name, const char *output_name)
+static DNNReturnType set_input_output_tf(void *model, DNNData *input, const char *input_name, const char **output_names, uint32_t nb_output)
 {
     TFModel *tf_model = (TFModel *)model;
     int64_t input_dims[] = {1, input->height, input->width, input->channels};
@@ -100,11 +103,38 @@ static DNNReturnType set_input_output_tf(void *model, DNNData *input, const char
     input->data = (float *)TF_TensorData(tf_model->input_tensor);
 
     // Output operation
-    tf_model->output.oper = TF_GraphOperationByName(tf_model->graph, output_name);
-    if (!tf_model->output.oper){
+    if (nb_output == 0)
+        return DNN_ERROR;
+
+    av_freep(&tf_model->outputs);
+    tf_model->outputs = av_malloc_array(nb_output, sizeof(*tf_model->outputs));
+    if (!tf_model->outputs)
+        return DNN_ERROR;
+    for (int i = 0; i < nb_output; ++i) {
+        tf_model->outputs[i].oper = TF_GraphOperationByName(tf_model->graph, output_names[i]);
+        if (!tf_model->outputs[i].oper){
+            av_freep(&tf_model->outputs);
+            return DNN_ERROR;
+        }
+        tf_model->outputs[i].index = 0;
+    }
+
+    if (tf_model->output_tensors) {
+        for (uint32_t i = 0; i < tf_model->nb_output; ++i) {
+            if (tf_model->output_tensors[i]) {
+                TF_DeleteTensor(tf_model->output_tensors[i]);
+                tf_model->output_tensors[i] = NULL;
+            }
+        }
+    }
+    av_freep(&tf_model->output_tensors);
+    tf_model->output_tensors = av_mallocz_array(nb_output, sizeof(*tf_model->output_tensors));
+    if (!tf_model->output_tensors) {
+        av_freep(&tf_model->outputs);
         return DNN_ERROR;
     }
-    tf_model->output.index = 0;
+
+    tf_model->nb_output = nb_output;
 
     if (tf_model->session){
         TF_CloseSession(tf_model->session, tf_model->status);
@@ -484,25 +514,36 @@ DNNModel *ff_dnn_load_model_tf(const char *model_filename)
 
 
 
-DNNReturnType ff_dnn_execute_model_tf(const DNNModel *model, DNNData *output)
+DNNReturnType ff_dnn_execute_model_tf(const DNNModel *model, DNNData *outputs, uint32_t nb_output)
 {
     TFModel *tf_model = (TFModel *)model->model;
-    if (tf_model->output_tensor)
-        TF_DeleteTensor(tf_model->output_tensor);
+    uint32_t nb = FFMIN(nb_output, tf_model->nb_output);
+    if (nb == 0)
+        return DNN_ERROR;
+
+    av_assert0(tf_model->output_tensors);
+    for (uint32_t i = 0; i < tf_model->nb_output; ++i) {
+        if (tf_model->output_tensors[i]) {
+            TF_DeleteTensor(tf_model->output_tensors[i]);
+            tf_model->output_tensors[i] = NULL;
+        }
+    }
 
     TF_SessionRun(tf_model->session, NULL,
                   &tf_model->input, &tf_model->input_tensor, 1,
-                  &tf_model->output, &tf_model->output_tensor, 1,
+                  tf_model->outputs, tf_model->output_tensors, nb,
                   NULL, 0, NULL, tf_model->status);
 
     if (TF_GetCode(tf_model->status) != TF_OK){
         return DNN_ERROR;
     }
 
-    output->height = TF_Dim(tf_model->output_tensor, 1);
-    output->width = TF_Dim(tf_model->output_tensor, 2);
-    output->channels = TF_Dim(tf_model->output_tensor, 3);
-    output->data = TF_TensorData(tf_model->output_tensor);
+    for (uint32_t i = 0; i < nb; ++i) {
+        outputs[i].height = TF_Dim(tf_model->output_tensors[i], 1);
+        outputs[i].width = TF_Dim(tf_model->output_tensors[i], 2);
+        outputs[i].channels = TF_Dim(tf_model->output_tensors[i], 3);
+        outputs[i].data = TF_TensorData(tf_model->output_tensors[i]);
+    }
 
     return DNN_SUCCESS;
 }
@@ -526,9 +567,16 @@ void ff_dnn_free_model_tf(DNNModel **model)
         if (tf_model->input_tensor){
             TF_DeleteTensor(tf_model->input_tensor);
         }
-        if (tf_model->output_tensor){
-            TF_DeleteTensor(tf_model->output_tensor);
+        if (tf_model->output_tensors) {
+            for (uint32_t i = 0; i < tf_model->nb_output; ++i) {
+                if (tf_model->output_tensors[i]) {
+                    TF_DeleteTensor(tf_model->output_tensors[i]);
+                    tf_model->output_tensors[i] = NULL;
+                }
+            }
         }
+        av_freep(&tf_model->outputs);
+        av_freep(&tf_model->output_tensors);
         av_freep(&tf_model);
         av_freep(model);
     }

libavfilter/dnn_backend_tf.h

@@ -31,7 +31,7 @@
 
 DNNModel *ff_dnn_load_model_tf(const char *model_filename);
 
-DNNReturnType ff_dnn_execute_model_tf(const DNNModel *model, DNNData *output);
+DNNReturnType ff_dnn_execute_model_tf(const DNNModel *model, DNNData *outputs, uint32_t nb_output);
 
 void ff_dnn_free_model_tf(DNNModel **model);
 

libavfilter/dnn_interface.h

@@ -26,6 +26,8 @@
 #ifndef AVFILTER_DNN_INTERFACE_H
 #define AVFILTER_DNN_INTERFACE_H
 
+#include <stdint.h>
+
 typedef enum {DNN_SUCCESS, DNN_ERROR} DNNReturnType;
 
 typedef enum {DNN_NATIVE, DNN_TF} DNNBackendType;
@@ -40,7 +42,7 @@ typedef struct DNNModel{
     void *model;
     // Sets model input and output.
     // Should be called at least once before model execution.
-    DNNReturnType (*set_input_output)(void *model, DNNData *input, const char *input_name, const char *output_name);
+    DNNReturnType (*set_input_output)(void *model, DNNData *input, const char *input_name, const char **output_names, uint32_t nb_output);
 } DNNModel;
 
 // Stores pointers to functions for loading, executing, freeing DNN models for one of the backends.
@@ -48,7 +50,7 @@ typedef struct DNNModule{
     // Loads model and parameters from given file. Returns NULL if it is not possible.
     DNNModel *(*load_model)(const char *model_filename);
     // Executes model with specified input and output. Returns DNN_ERROR otherwise.
-    DNNReturnType (*execute_model)(const DNNModel *model, DNNData *output);
+    DNNReturnType (*execute_model)(const DNNModel *model, DNNData *outputs, uint32_t nb_output);
     // Frees memory allocated for model.
     void (*free_model)(DNNModel **model);
 } DNNModule;

libavfilter/vf_sr.c

@@ -116,18 +116,19 @@ static int config_props(AVFilterLink *inlink)
     AVFilterLink *outlink = context->outputs[0];
     DNNReturnType result;
     int sws_src_h, sws_src_w, sws_dst_h, sws_dst_w;
+    const char *model_output_name = "y";
 
     sr_context->input.width = inlink->w * sr_context->scale_factor;
     sr_context->input.height = inlink->h * sr_context->scale_factor;
     sr_context->input.channels = 1;
 
-    result = (sr_context->model->set_input_output)(sr_context->model->model, &sr_context->input, "x", "y");
+    result = (sr_context->model->set_input_output)(sr_context->model->model, &sr_context->input, "x", &model_output_name, 1);
     if (result != DNN_SUCCESS){
         av_log(context, AV_LOG_ERROR, "could not set input and output for the model\n");
         return AVERROR(EIO);
     }
 
-    result = (sr_context->dnn_module->execute_model)(sr_context->model, &sr_context->output);
+    result = (sr_context->dnn_module->execute_model)(sr_context->model, &sr_context->output, 1);
     if (result != DNN_SUCCESS){
         av_log(context, AV_LOG_ERROR, "failed to execute loaded model\n");
         return AVERROR(EIO);
@@ -136,12 +137,12 @@ static int config_props(AVFilterLink *inlink)
     if (sr_context->input.height != sr_context->output.height || sr_context->input.width != sr_context->output.width){
         sr_context->input.width = inlink->w;
         sr_context->input.height = inlink->h;
-        result = (sr_context->model->set_input_output)(sr_context->model->model, &sr_context->input, "x", "y");
+        result = (sr_context->model->set_input_output)(sr_context->model->model, &sr_context->input, "x", &model_output_name, 1);
         if (result != DNN_SUCCESS){
             av_log(context, AV_LOG_ERROR, "could not set input and output for the model\n");
             return AVERROR(EIO);
         }
-        result = (sr_context->dnn_module->execute_model)(sr_context->model, &sr_context->output);
+        result = (sr_context->dnn_module->execute_model)(sr_context->model, &sr_context->output, 1);
         if (result != DNN_SUCCESS){
             av_log(context, AV_LOG_ERROR, "failed to execute loaded model\n");
             return AVERROR(EIO);
@@ -256,7 +257,7 @@ static int filter_frame(AVFilterLink *inlink, AVFrame *in)
     }
     av_frame_free(&in);
 
-    dnn_result = (sr_context->dnn_module->execute_model)(sr_context->model, &sr_context->output);
+    dnn_result = (sr_context->dnn_module->execute_model)(sr_context->model, &sr_context->output, 1);
     if (dnn_result != DNN_SUCCESS){
         av_log(context, AV_LOG_ERROR, "failed to execute loaded model\n");
         return AVERROR(EIO);