Adds TensorFlow backend for dnn inference module.

Signed-off-by: Pedro Arthur <bygrandao@gmail.com>

Changelog

@@ -9,10 +9,9 @@ version <next>:
 - aderivative and aintegral audio filters
 - pal75bars and pal100bars video filter sources
 - support mbedTLS based TLS
-- DNN inference interface
-- Reimplemented SRCNN filter using DNN inference interface
 - adeclick filter
 - adeclip filter
+- libtensorflow backend for DNN based filters like srcnn
 
 
 version 4.0:

configure

@@ -259,6 +259,8 @@ External library support:
   --enable-libspeex        enable Speex de/encoding via libspeex [no]
   --enable-libsrt          enable Haivision SRT protocol via libsrt [no]
   --enable-libssh          enable SFTP protocol via libssh [no]
+  --enable-libtensorflow   enable TensorFlow as a DNN module backend
+                           for DNN based filters like srcnn [no]
   --enable-libtesseract    enable Tesseract, needed for ocr filter [no]
   --enable-libtheora       enable Theora encoding via libtheora [no]
   --enable-libtls          enable LibreSSL (via libtls), needed for https support
@@ -1713,6 +1715,7 @@ EXTERNAL_LIBRARY_LIST="
     libspeex
     libsrt
     libssh
+    libtensorflow
     libtesseract
     libtheora
     libtwolame
@@ -2244,6 +2247,7 @@ CONFIG_EXTRA="
     cbs_mpeg2
     cbs_vp9
     dirac_parse
+    dnn
     dvprofile
     exif
     faandct
@@ -2507,6 +2511,7 @@ cbs_mpeg2_select="cbs"
 cbs_vp9_select="cbs"
 dct_select="rdft"
 dirac_parse_select="golomb"
+dnn_suggest="libtensorflow"
 error_resilience_select="me_cmp"
 faandct_deps="faan"
 faandct_select="fdctdsp"
@@ -3396,6 +3401,7 @@ spectrumsynth_filter_select="fft"
 spp_filter_deps="gpl avcodec"
 spp_filter_select="fft idctdsp fdctdsp me_cmp pixblockdsp"
 srcnn_filter_deps="avformat"
+srcnn_filter_select="dnn"
 stereo3d_filter_deps="gpl"
 subtitles_filter_deps="avformat avcodec libass"
 super2xsai_filter_deps="gpl"
@@ -6055,6 +6061,7 @@ enabled libsoxr           && require libsoxr soxr.h soxr_create -lsoxr
 enabled libssh            && require_pkg_config libssh libssh libssh/sftp.h sftp_init
 enabled libspeex          && require_pkg_config libspeex speex speex/speex.h speex_decoder_init
 enabled libsrt            && require_pkg_config libsrt "srt >= 1.2.0" srt/srt.h srt_socket
+enabled libtensorflow     && require libtensorflow tensorflow/c/c_api.h TF_Version -ltensorflow
 enabled libtesseract      && require_pkg_config libtesseract tesseract tesseract/capi.h TessBaseAPICreate
 enabled libtheora         && require libtheora theora/theoraenc.h th_info_init -ltheoraenc -ltheoradec -logg
 enabled libtls            && require_pkg_config libtls libtls tls.h tls_configure

libavfilter/Makefile

@@ -26,6 +26,8 @@ OBJS-$(HAVE_THREADS)                         += pthread.o
 
 # subsystems
 OBJS-$(CONFIG_QSVVPP)                        += qsvvpp.o
+DNN-OBJS-$(CONFIG_LIBTENSORFLOW)             += dnn_backend_tf.o
+OBJS-$(CONFIG_DNN)                           += dnn_interface.o dnn_backend_native.o $(DNN-OBJS-yes)
 
 # audio filters
 OBJS-$(CONFIG_ABENCH_FILTER)                 += f_bench.o
@@ -336,7 +338,7 @@ OBJS-$(CONFIG_SMARTBLUR_FILTER)              += vf_smartblur.o
 OBJS-$(CONFIG_SOBEL_FILTER)                  += vf_convolution.o
 OBJS-$(CONFIG_SPLIT_FILTER)                  += split.o
 OBJS-$(CONFIG_SPP_FILTER)                    += vf_spp.o
-OBJS-$(CONFIG_SRCNN_FILTER)                  += vf_srcnn.o dnn_interface.o dnn_backend_native.o
+OBJS-$(CONFIG_SRCNN_FILTER)                  += vf_srcnn.o
 OBJS-$(CONFIG_SSIM_FILTER)                   += vf_ssim.o framesync.o
 OBJS-$(CONFIG_STEREO3D_FILTER)               += vf_stereo3d.o
 OBJS-$(CONFIG_STREAMSELECT_FILTER)           += f_streamselect.o framesync.o

libavfilter/dnn_backend_tf.c

@@ -0,0 +1,309 @@
+/*
+ * Copyright (c) 2018 Sergey Lavrushkin
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+/**
+ * @file
+ * DNN tensorflow backend implementation.
+ */
+
+#include "dnn_backend_tf.h"
+#include "dnn_srcnn.h"
+#include "libavformat/avio.h"
+
+#include <tensorflow/c/c_api.h>
+
+typedef struct TFModel{
+    TF_Graph* graph;
+    TF_Session* session;
+    TF_Status* status;
+    TF_Output input, output;
+    TF_Tensor* input_tensor;
+    TF_Tensor* output_tensor;
+    const DNNData* input_data;
+    const DNNData* output_data;
+} TFModel;
+
+static void free_buffer(void* data, size_t length)
+{
+    av_freep(&data);
+}
+
+static TF_Buffer* read_graph(const char* model_filename)
+{
+    TF_Buffer* graph_buf;
+    unsigned char* graph_data = NULL;
+    AVIOContext* model_file_context;
+    long size, bytes_read;
+
+    if (avio_open(&model_file_context, model_filename, AVIO_FLAG_READ) < 0){
+        return NULL;
+    }
+
+    size = avio_size(model_file_context);
+
+    graph_data = av_malloc(size);
+    if (!graph_data){
+        avio_closep(&model_file_context);
+        return NULL;
+    }
+    bytes_read = avio_read(model_file_context, graph_data, size);
+    avio_closep(&model_file_context);
+    if (bytes_read != size){
+        av_freep(&graph_data);
+        return NULL;
+    }
+
+    graph_buf = TF_NewBuffer();
+    graph_buf->data = (void*)graph_data;
+    graph_buf->length = size;
+    graph_buf->data_deallocator = free_buffer;
+
+    return graph_buf;
+}
+
+static DNNReturnType set_input_output_tf(void* model, const DNNData* input, const DNNData* output)
+{
+    TFModel* tf_model = (TFModel*)model;
+    int64_t input_dims[] = {1, input->height, input->width, input->channels};
+    int64_t output_dims[] = {1, output->height, output->width, output->channels};
+    TF_SessionOptions* sess_opts;
+    const TF_Operation* init_op = TF_GraphOperationByName(tf_model->graph, "init");
+
+    // Input operation should be named 'x'
+    tf_model->input.oper = TF_GraphOperationByName(tf_model->graph, "x");
+    if (!tf_model->input.oper){
+        return DNN_ERROR;
+    }
+    tf_model->input.index = 0;
+    if (tf_model->input_tensor){
+        TF_DeleteTensor(tf_model->input_tensor);
+    }
+    tf_model->input_tensor = TF_AllocateTensor(TF_FLOAT, input_dims, 4,
+                                               input_dims[1] * input_dims[2] * input_dims[3] * sizeof(float));
+    if (!tf_model->input_tensor){
+        return DNN_ERROR;
+    }
+
+    // Output operation should be named 'y'
+    tf_model->output.oper = TF_GraphOperationByName(tf_model->graph, "y");
+    if (!tf_model->output.oper){
+        return DNN_ERROR;
+    }
+    tf_model->output.index = 0;
+    if (tf_model->output_tensor){
+        TF_DeleteTensor(tf_model->output_tensor);
+    }
+    tf_model->output_tensor = TF_AllocateTensor(TF_FLOAT, output_dims, 4,
+                                                output_dims[1] * output_dims[2] * output_dims[3] * sizeof(float));
+    if (!tf_model->output_tensor){
+        return DNN_ERROR;
+    }
+
+    tf_model->input_data = input;
+    tf_model->output_data = output;
+
+    if (tf_model->session){
+        TF_CloseSession(tf_model->session, tf_model->status);
+        TF_DeleteSession(tf_model->session, tf_model->status);
+    }
+
+    sess_opts = TF_NewSessionOptions();
+    tf_model->session = TF_NewSession(tf_model->graph, sess_opts, tf_model->status);
+    TF_DeleteSessionOptions(sess_opts);
+    if (TF_GetCode(tf_model->status) != TF_OK)
+    {
+        return DNN_ERROR;
+    }
+
+    // Run initialization operation with name "init" if it is present in graph
+    if (init_op){
+        TF_SessionRun(tf_model->session, NULL,
+                      NULL, NULL, 0,
+                      NULL, NULL, 0,
+                      &init_op, 1, NULL, tf_model->status);
+        if (TF_GetCode(tf_model->status) != TF_OK)
+        {
+            return DNN_ERROR;
+        }
+    }
+
+    return DNN_SUCCESS;
+}
+
+DNNModel* ff_dnn_load_model_tf(const char* model_filename)
+{
+    DNNModel* model = NULL;
+    TFModel* tf_model = NULL;
+    TF_Buffer* graph_def;
+    TF_ImportGraphDefOptions* graph_opts;
+
+    model = av_malloc(sizeof(DNNModel));
+    if (!model){
+        return NULL;
+    }
+
+    tf_model = av_malloc(sizeof(TFModel));
+    if (!tf_model){
+        av_freep(&model);
+        return NULL;
+    }
+    tf_model->session = NULL;
+    tf_model->input_tensor = NULL;
+    tf_model->output_tensor = NULL;
+
+    graph_def = read_graph(model_filename);
+    if (!graph_def){
+        av_freep(&tf_model);
+        av_freep(&model);
+        return NULL;
+    }
+    tf_model->graph = TF_NewGraph();
+    tf_model->status = TF_NewStatus();
+    graph_opts = TF_NewImportGraphDefOptions();
+    TF_GraphImportGraphDef(tf_model->graph, graph_def, graph_opts, tf_model->status);
+    TF_DeleteImportGraphDefOptions(graph_opts);
+    TF_DeleteBuffer(graph_def);
+    if (TF_GetCode(tf_model->status) != TF_OK){
+        TF_DeleteGraph(tf_model->graph);
+        TF_DeleteStatus(tf_model->status);
+        av_freep(&tf_model);
+        av_freep(&model);
+        return NULL;
+    }
+
+    model->model = (void*)tf_model;
+    model->set_input_output = &set_input_output_tf;
+
+    return model;
+}
+
+DNNModel* ff_dnn_load_default_model_tf(DNNDefaultModel model_type)
+{
+    DNNModel* model = NULL;
+    TFModel* tf_model = NULL;
+    TF_Buffer* graph_def;
+    unsigned char* graph_data = NULL;
+    TF_ImportGraphDefOptions* graph_opts;
+
+    graph_def = TF_NewBuffer();
+    switch (model_type){
+    case DNN_SRCNN:
+        graph_data = av_malloc(srcnn_tf_size);
+        if (!graph_data){
+            TF_DeleteBuffer(graph_def);
+            return NULL;
+        }
+        memcpy(graph_data, srcnn_tf_model, srcnn_tf_size);
+        graph_def->data = (void*)graph_data;
+        graph_def->length = srcnn_tf_size;
+        graph_def->data_deallocator = free_buffer;
+        break;
+    default:
+        TF_DeleteBuffer(graph_def);
+        return NULL;
+    }
+
+    model = av_malloc(sizeof(DNNModel));
+    if (!model){
+        TF_DeleteBuffer(graph_def);
+        return NULL;
+    }
+
+    tf_model = av_malloc(sizeof(TFModel));
+    if (!tf_model){
+        TF_DeleteBuffer(graph_def);
+        av_freep(&model);
+        return NULL;
+    }
+    tf_model->session = NULL;
+    tf_model->input_tensor = NULL;
+    tf_model->output_tensor = NULL;
+
+    tf_model->graph = TF_NewGraph();
+    tf_model->status = TF_NewStatus();
+    graph_opts = TF_NewImportGraphDefOptions();
+    TF_GraphImportGraphDef(tf_model->graph, graph_def, graph_opts, tf_model->status);
+    TF_DeleteImportGraphDefOptions(graph_opts);
+    TF_DeleteBuffer(graph_def);
+    if (TF_GetCode(tf_model->status) != TF_OK){
+        TF_DeleteGraph(tf_model->graph);
+        TF_DeleteStatus(tf_model->status);
+        av_freep(&tf_model);
+        av_freep(&model);
+        return NULL;
+    }
+
+    model->model = (void*)tf_model;
+    model->set_input_output = &set_input_output_tf;
+
+    return model;
+}
+
+DNNReturnType ff_dnn_execute_model_tf(const DNNModel* model)
+{
+    TFModel* tf_model = (TFModel*)model->model;
+
+    memcpy(TF_TensorData(tf_model->input_tensor), tf_model->input_data->data,
+           tf_model->input_data->height * tf_model->input_data->width *
+           tf_model->input_data->channels * sizeof(float));
+
+    TF_SessionRun(tf_model->session, NULL,
+                  &tf_model->input, &tf_model->input_tensor, 1,
+                  &tf_model->output, &tf_model->output_tensor, 1,
+                  NULL, 0, NULL, tf_model->status);
+
+    if (TF_GetCode(tf_model->status) != TF_OK){
+        return DNN_ERROR;
+    }
+    else{
+        memcpy(tf_model->output_data->data, TF_TensorData(tf_model->output_tensor),
+           tf_model->output_data->height * tf_model->output_data->width *
+           tf_model->output_data->channels * sizeof(float));
+
+        return DNN_SUCCESS;
+    }
+}
+
+void ff_dnn_free_model_tf(DNNModel** model)
+{
+    TFModel* tf_model;
+
+    if (*model){
+        tf_model = (TFModel*)(*model)->model;
+        if (tf_model->graph){
+            TF_DeleteGraph(tf_model->graph);
+        }
+        if (tf_model->session){
+            TF_CloseSession(tf_model->session, tf_model->status);
+            TF_DeleteSession(tf_model->session, tf_model->status);
+        }
+        if (tf_model->status){
+            TF_DeleteStatus(tf_model->status);
+        }
+        if (tf_model->input_tensor){
+            TF_DeleteTensor(tf_model->input_tensor);
+        }
+        if (tf_model->output_tensor){
+            TF_DeleteTensor(tf_model->output_tensor);
+        }
+        av_freep(&tf_model);
+        av_freep(model);
+    }
+}

libavfilter/dnn_backend_tf.h

@@ -0,0 +1,40 @@
+/*
+ * Copyright (c) 2018 Sergey Lavrushkin
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+/**
+ * @file
+ * DNN inference functions interface for TensorFlow backend.
+ */
+
+
+#ifndef AVFILTER_DNN_BACKEND_TF_H
+#define AVFILTER_DNN_BACKEND_TF_H
+
+#include "dnn_interface.h"
+
+DNNModel* ff_dnn_load_model_tf(const char* model_filename);
+
+DNNModel* ff_dnn_load_default_model_tf(DNNDefaultModel model_type);
+
+DNNReturnType ff_dnn_execute_model_tf(const DNNModel* model);
+
+void ff_dnn_free_model_tf(DNNModel** model);
+
+#endif

libavfilter/dnn_interface.c

@@ -25,6 +25,7 @@
 
 #include "dnn_interface.h"
 #include "dnn_backend_native.h"
+#include "dnn_backend_tf.h"
 #include "libavutil/mem.h"
 
 DNNModule* ff_get_dnn_module(DNNBackendType backend_type)
@@ -42,6 +43,17 @@ DNNModule* ff_get_dnn_module(DNNBackendType backend_type)
         dnn_module->load_default_model = &ff_dnn_load_default_model_native;
         dnn_module->execute_model = &ff_dnn_execute_model_native;
         dnn_module->free_model = &ff_dnn_free_model_native;
+        break;
+    case DNN_TF:
+    #if (CONFIG_LIBTENSORFLOW == 1)
+        dnn_module->load_model = &ff_dnn_load_model_tf;
+        dnn_module->load_default_model = &ff_dnn_load_default_model_tf;
+        dnn_module->execute_model = &ff_dnn_execute_model_tf;
+        dnn_module->free_model = &ff_dnn_free_model_tf;
+    #else
+        av_freep(dnn_module);
+        return NULL;
+    #endif
     }
 
     return dnn_module;

libavfilter/dnn_interface.h

@@ -28,7 +28,7 @@
 
 typedef enum {DNN_SUCCESS, DNN_ERROR} DNNReturnType;
 
-typedef enum {DNN_NATIVE} DNNBackendType;
+typedef enum {DNN_NATIVE, DNN_TF} DNNBackendType;
 
 typedef enum {DNN_SRCNN} DNNDefaultModel;
 
@@ -37,7 +37,6 @@ typedef struct DNNData{
     int width, height, channels;
 } DNNData;
 
-
 typedef struct DNNModel{
     // Stores model that can be different for different backends.
     void* model;

libavfilter/vf_srcnn.c

@@ -41,7 +41,6 @@ typedef struct SRCNNContext {
     DNNData input_output;
 } SRCNNContext;
 
-
 #define OFFSET(x) offsetof(SRCNNContext, x)
 #define FLAGS AV_OPT_FLAG_FILTERING_PARAM | AV_OPT_FLAG_VIDEO_PARAM
 static const AVOption srcnn_options[] = {
@@ -55,11 +54,20 @@ static av_cold int init(AVFilterContext* context)
 {
     SRCNNContext* srcnn_context = context->priv;
 
+    srcnn_context->dnn_module = ff_get_dnn_module(DNN_TF);
+    if (!srcnn_context->dnn_module){
         srcnn_context->dnn_module = ff_get_dnn_module(DNN_NATIVE);
         if (!srcnn_context->dnn_module){
             av_log(context, AV_LOG_ERROR, "could not create dnn module\n");
             return AVERROR(ENOMEM);
         }
+        else{
+            av_log(context, AV_LOG_INFO, "using native backend for DNN inference\n");
+        }
+    }
+    else{
+        av_log(context, AV_LOG_INFO, "using tensorflow backend for DNN inference\n");
+    }
     if (!srcnn_context->model_filename){
         av_log(context, AV_LOG_INFO, "model file for network was not specified, using default network for x2 upsampling\n");
         srcnn_context->model = (srcnn_context->dnn_module->load_default_model)(DNN_SRCNN);