dnn/native: add native support for avg_pool

Not support pooling strides in channel dimension yet.

Signed-off-by: Ting Fu <ting.fu@intel.com>
Reviewed-by: Guo, Yejun <yejun.guo@intel.com>

libavfilter/dnn/Makefile

@@ -1,6 +1,7 @@
 OBJS-$(CONFIG_DNN)                           += dnn/dnn_interface.o
 OBJS-$(CONFIG_DNN)                           += dnn/dnn_backend_native.o
 OBJS-$(CONFIG_DNN)                           += dnn/dnn_backend_native_layers.o
+OBJS-$(CONFIG_DNN)                           += dnn/dnn_backend_native_layer_avgpool.o
 OBJS-$(CONFIG_DNN)                           += dnn/dnn_backend_native_layer_pad.o
 OBJS-$(CONFIG_DNN)                           += dnn/dnn_backend_native_layer_conv2d.o
 OBJS-$(CONFIG_DNN)                           += dnn/dnn_backend_native_layer_depth2space.o

libavfilter/dnn/dnn_backend_native.h

@@ -43,10 +43,12 @@ typedef enum {
     DLT_MAXIMUM = 4,
     DLT_MATH_BINARY = 5,
     DLT_MATH_UNARY = 6,
+    DLT_AVG_POOL = 7,
     DLT_COUNT
 } DNNLayerType;
 
 typedef enum {DOT_INPUT = 1, DOT_OUTPUT = 2, DOT_INTERMEDIATE = DOT_INPUT | DOT_OUTPUT} DNNOperandType;
+typedef enum {VALID, SAME, SAME_CLAMP_TO_EDGE} DNNPaddingParam;
 
 typedef struct Layer{
     DNNLayerType type;

libavfilter/dnn/dnn_backend_native_layer_avgpool.c

@@ -0,0 +1,141 @@
+/*
+ * Copyright (c) 2020
+ *
+ * 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 native backend implementation.
+ */
+
+#include "libavutil/avassert.h"
+#include "dnn_backend_native_layer_avgpool.h"
+
+int dnn_load_layer_avg_pool(Layer *layer, AVIOContext *model_file_context, int file_size, int operands_num)
+{
+    AvgPoolParams *avgpool_params;
+    int dnn_size = 0;
+    avgpool_params = av_malloc(sizeof(*avgpool_params));
+    if(!avgpool_params)
+        return 0;
+
+    avgpool_params->strides = (int32_t)avio_rl32(model_file_context);
+    avgpool_params->padding_method = (int32_t)avio_rl32(model_file_context);
+    avgpool_params->kernel_size = (int32_t)avio_rl32(model_file_context);
+    dnn_size += 12;
+
+    if (dnn_size > file_size || avgpool_params->kernel_size <= 0 || avgpool_params->strides <=0){
+        av_freep(&avgpool_params);
+        return 0;
+    }
+
+    layer->params = avgpool_params;
+    layer->input_operand_indexes[0] = (int32_t)avio_rl32(model_file_context);
+    layer->output_operand_index = (int32_t)avio_rl32(model_file_context);
+    dnn_size += 8;
+
+    if (layer->input_operand_indexes[0] >= operands_num || layer->output_operand_index >= operands_num) {
+        return 0;
+    }
+    return dnn_size;
+}
+
+int dnn_execute_layer_avg_pool(DnnOperand *operands, const int32_t *input_operand_indexes,
+                             int32_t output_operand_index, const void *parameters)
+{
+    float *output;
+    int height_end, width_end, height_radius, width_radius, output_height, output_width, kernel_area;
+    int32_t input_operand_index = input_operand_indexes[0];
+    int number = operands[input_operand_index].dims[0];
+    int height = operands[input_operand_index].dims[1];
+    int width = operands[input_operand_index].dims[2];
+    int channel = operands[input_operand_index].dims[3];
+    const float *input = operands[input_operand_index].data;
+    const AvgPoolParams *avgpool_params = (const AvgPoolParams *)parameters;
+
+    int kernel_strides = avgpool_params->strides;
+    int src_linesize = width * channel;
+    DnnOperand *output_operand = &operands[output_operand_index];
+
+    /**
+     * When padding_method = SAME, the tensorflow will only padding the hald number of 0 pxiels
+     * except the remainders.
+     * Eg: assuming the input height = 1080, the strides = 11, so the remainders = 1080 % 11 = 2
+     *     and if ksize = 5: it will fill (5 - 2) >> 1 = 1 line before the first line of input image,
+     *                       and 5 - 2 - 1 = 2 lines after the last line of input image.
+     *     and if ksize = 7: it will fill (7 - 2) >> 1 = 2 lines before the first line of input image,
+     *                       and 7 - 2 - 2 = 3 lines after the last line of input image.
+     */
+    if (avgpool_params->padding_method == SAME) {
+        height_end = height;
+        width_end = width;
+        height_radius = avgpool_params->kernel_size - ((height - 1) % kernel_strides + 1);
+        width_radius = avgpool_params->kernel_size - ((width - 1) % kernel_strides + 1);
+        height_radius = height_radius < 0 ? 0 : height_radius >> 1;
+        width_radius = width_radius < 0 ? 0 : width_radius >> 1;
+        output_height = ceil(height / (kernel_strides * 1.0));
+        output_width = ceil(width / (kernel_strides * 1.0));
+    } else {
+        assert(avgpool_params->padding_method = VALID);
+        height_end = height - avgpool_params->kernel_size + 1;
+        width_end = width - avgpool_params->kernel_size + 1;
+        height_radius = 0;
+        width_radius = 0;
+        output_height = ceil((height - avgpool_params->kernel_size + 1) / (kernel_strides * 1.0));
+        output_width = ceil((width - avgpool_params->kernel_size + 1) / (kernel_strides * 1.0));
+    }
+
+    output_operand->dims[0] = number;
+    output_operand->dims[1] = output_height;
+    output_operand->dims[2] = output_width;
+    // not support pooling in channel dimension now
+    output_operand->dims[3] = channel;
+    output_operand->data_type = operands[input_operand_index].data_type;
+    output_operand->length = calculate_operand_data_length(output_operand);
+    output_operand->data = av_realloc(output_operand->data, output_operand->length);
+    if (!output_operand->data)
+        return -1;
+    output = output_operand->data;
+
+    for (int y = 0; y < height_end; y += kernel_strides) {
+        for (int x = 0; x < width_end; x += kernel_strides) {
+            for (int n_channel = 0; n_channel < channel; ++n_channel) {
+                output[n_channel] = 0.0;
+                kernel_area = 0;
+                for (int kernel_y = 0; kernel_y < avgpool_params->kernel_size; ++kernel_y) {
+                    for (int kernel_x = 0; kernel_x < avgpool_params->kernel_size; ++kernel_x) {
+                        float input_pel;
+                        int y_pos = y + (kernel_y - height_radius);
+                        int x_pos = x + (kernel_x - width_radius);
+                        if (x_pos < 0 || x_pos >= width || y_pos < 0 || y_pos >= height) {
+                            input_pel = 0.0;
+                        } else {
+                            kernel_area++;
+                            input_pel = input[y_pos * src_linesize + x_pos * channel + n_channel];
+                        }
+                        output[n_channel] += input_pel;
+                    }
+                }
+                output[n_channel] /= kernel_area;
+            }
+            output += channel;
+        }
+    }
+
+    return 0;
+}

libavfilter/dnn/dnn_backend_native_layer_avgpool.h

@@ -0,0 +1,40 @@
+/*
+ * Copyright (c) 2020
+ *
+ * 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 native backend.
+ */
+
+#ifndef AVFILTER_DNN_DNN_BACKEND_NATIVE_LAYER_AVGPOOL_H
+#define AVFILTER_DNN_DNN_BACKEND_NATIVE_LAYER_AVGPOOL_H
+
+#include "dnn_backend_native.h"
+
+typedef struct AvgPoolParams{
+    int32_t strides, kernel_size;
+    DNNPaddingParam padding_method;
+} AvgPoolParams;
+
+int dnn_load_layer_avg_pool(Layer *layer, AVIOContext *model_file_context, int file_size, int operands_num);
+int dnn_execute_layer_avg_pool(DnnOperand *operands, const int32_t *input_operand_indexes,
+                             int32_t output_operand_index, const void *parameters);
+
+#endif

libavfilter/dnn/dnn_backend_native_layer_conv2d.h

@@ -24,12 +24,11 @@
 #include "dnn_backend_native.h"
 
 typedef enum {RELU, TANH, SIGMOID, NONE, LEAKY_RELU} DNNActivationFunc;
-typedef enum {VALID, SAME, SAME_CLAMP_TO_EDGE} DNNConvPaddingParam;
 
 typedef struct ConvolutionalParams{
     int32_t input_num, output_num, kernel_size;
     DNNActivationFunc activation;
-    DNNConvPaddingParam padding_method;
+    DNNPaddingParam padding_method;
     int32_t dilation;
     int32_t has_bias;
     float *kernel;

libavfilter/dnn/dnn_backend_native_layers.c

@@ -26,6 +26,7 @@
 #include "dnn_backend_native_layer_maximum.h"
 #include "dnn_backend_native_layer_mathbinary.h"
 #include "dnn_backend_native_layer_mathunary.h"
+#include "dnn_backend_native_layer_avgpool.h"
 
 LayerFunc layer_funcs[DLT_COUNT] = {
     {NULL, NULL},
@@ -35,4 +36,5 @@ LayerFunc layer_funcs[DLT_COUNT] = {
     {dnn_execute_layer_maximum,     dnn_load_layer_maximum},
     {dnn_execute_layer_math_binary, dnn_load_layer_math_binary},
     {dnn_execute_layer_math_unary,  dnn_load_layer_math_unary},
+    {dnn_execute_layer_avg_pool,  dnn_load_layer_avg_pool},
 };

tools/python/convert_from_tensorflow.py

@@ -67,10 +67,12 @@ class TFConverter:
         self.edges = {}
         self.conv_activations = {'Relu':0, 'Tanh':1, 'Sigmoid':2, 'None':3, 'LeakyRelu':4}
         self.conv_paddings = {'VALID':0, 'SAME':1}
+        self.pool_paddings = {'VALID':0, 'SAME':1}
         self.converted_nodes = set()
         self.conv2d_scope_names = set()
         self.conv2d_scopename_inputname_dict = {}
-        self.op2code = {'Conv2D':1, 'DepthToSpace':2, 'MirrorPad':3, 'Maximum':4, 'MathBinary':5, 'MathUnary':6}
+        self.op2code = {'Conv2D':1, 'DepthToSpace':2, 'MirrorPad':3, 'Maximum':4,
+                        'MathBinary':5, 'MathUnary':6, 'AvgPool':7}
         self.mathbin2code = {'Sub':0, 'Add':1, 'Mul':2, 'RealDiv':3, 'Minimum':4}
         self.mathun2code  = {'Abs':0, 'Sin':1, 'Cos':2, 'Tan':3, 'Asin':4,
                 'Acos':5, 'Atan':6, 'Sinh':7, 'Cosh':8, 'Tanh':9, 'Asinh':10,
@@ -300,6 +302,37 @@ class TFConverter:
         np.array([output_operand_index],dtype=np.uint32).tofile(f)
 
 
+    def dump_avg_pool_to_file(self, node, f):
+        assert(node.op == 'AvgPool')
+        self.layer_number = self.layer_number + 1
+        self.converted_nodes.add(node.name)
+        node0 = self.name_node_dict[node.input[0]]
+        strides = node.attr['strides']
+
+        # Tensorflow do not support pooling strides in batch dimension and
+        # current native NN do not support pooling strides in channel dimension, added assert() here.
+        assert(strides.list.i[1]==strides.list.i[2])
+        assert(strides.list.i[0]==1)
+        assert(strides.list.i[3]==1)
+        strides = strides.list.i[1]
+        filter_node = node.attr['ksize']
+        input_name = node.input[0]
+
+        # Tensorflow do not support pooling ksize in batch dimension and channel dimension.
+        assert(filter_node.list.i[0]==1)
+        assert(filter_node.list.i[3]==1)
+        filter_height = filter_node.list.i[1]
+        filter_width = filter_node.list.i[2]
+
+        padding = node.attr['padding'].s.decode("utf-8")
+        np.array([self.op2code[node.op], strides, self.pool_paddings[padding], filter_height],
+                 dtype=np.uint32).tofile(f)
+
+        input_operand_index = self.add_operand(input_name, Operand.IOTYPE_INPUT)
+        output_operand_index = self.add_operand(node.name, Operand.IOTYPE_OUTPUT)
+        np.array([input_operand_index, output_operand_index],dtype=np.uint32).tofile(f)
+
+
     def dump_layers_to_file(self, f):
         for node in self.nodes:
             if node.name in self.converted_nodes:
@@ -313,6 +346,8 @@ class TFConverter:
 
             if node.op == 'Conv2D':
                 self.dump_simple_conv2d_to_file(node, f)
+            if node.op == 'AvgPool':
+                self.dump_avg_pool_to_file(node, f)
             elif node.op == 'DepthToSpace':
                 self.dump_depth2space_to_file(node, f)
             elif node.op == 'MirrorPad':