History

dnn_backend_native_layer_mathunary: add abs support

more math unary operations will be added here It can be tested with the model file generated with below python scripy: import tensorflow as tf import numpy as np import imageio in_img = imageio.imread('input.jpeg') in_img = in_img.astype(np.float32)/255.0 in_data = in_img[np.newaxis, :] x = tf.placeholder(tf.float32, shape=[1, None, None, 3], name='dnn_in') x1 = tf.subtract(x, 0.5) x2 = tf.abs(x1) y = tf.identity(x2, name='dnn_out') sess=tf.Session() sess.run(tf.global_variables_initializer()) graph_def = tf.graph_util.convert_variables_to_constants(sess, sess.graph_def, ['dnn_out']) tf.train.write_graph(graph_def, '.', 'image_process.pb', as_text=False) print("image_process.pb generated, please use \ path_to_ffmpeg/tools/python/convert.py to generate image_process.model\n") output = sess.run(y, feed_dict={x: in_data}) imageio.imsave("out.jpg", np.squeeze(output))

Ting Fu committed 6 years ago

Commit f73cc611 parent b6d65977 files · +145 −2Browse files

Unified · Split

7 files
  1. libavfilter/dnn/Makefile +1 −0
  2. libavfilter/dnn/dnn_backend_native.h +1 −0
  3. libavfilter/dnn/dnn_backend_native_layer_mathunary.c +80 −0
  4. libavfilter/dnn/dnn_backend_native_layer_mathunary.h +45 −0
  5. libavfilter/dnn/dnn_backend_native_layers.c +2 −0
  6. tools/python/convert_from_tensorflow.py +15 −1
  7. tools/python/convert_header.py +1 −1
libavfilter/dnn/Makefile+1−0
@@ -6,6 +6,7 @@ OBJS-$(CONFIG_DNN) += dnn/dnn_backend_native_layer_con · 5 unmodified lines
6OBJS-$(CONFIG_DNN) += dnn/dnn_backend_native_layer_depth2space.o6OBJS-$(CONFIG_DNN) += dnn/dnn_backend_native_layer_depth2space.o
7OBJS-$(CONFIG_DNN) += dnn/dnn_backend_native_layer_maximum.o7OBJS-$(CONFIG_DNN) += dnn/dnn_backend_native_layer_maximum.o
8OBJS-$(CONFIG_DNN) += dnn/dnn_backend_native_layer_mathbinary.o8OBJS-$(CONFIG_DNN) += dnn/dnn_backend_native_layer_mathbinary.o
9OBJS-$(CONFIG_DNN) += dnn/dnn_backend_native_layer_mathunary.o
9 10
10DNN-OBJS-$(CONFIG_LIBTENSORFLOW) += dnn/dnn_backend_tf.o11DNN-OBJS-$(CONFIG_LIBTENSORFLOW) += dnn/dnn_backend_tf.o
11 12
libavfilter/dnn/dnn_backend_native.h+1−0
@@ -42,6 +42,7 @@ typedef enum { · 41 unmodified lines
42 DLT_MIRROR_PAD = 3,42 DLT_MIRROR_PAD = 3,
43 DLT_MAXIMUM = 4,43 DLT_MAXIMUM = 4,
44 DLT_MATH_BINARY = 5,44 DLT_MATH_BINARY = 5,
45 DLT_MATH_UNARY = 6,
45 DLT_COUNT46 DLT_COUNT
46} DNNLayerType;47} DNNLayerType;
47 48
libavfilter/dnn/dnn_backend_native_layer_mathunary.c+80−0
@@ -0,0 +1,80 @@
1/*
2 * Copyright (c) 2020
3 *
4 * This file is part of FFmpeg.
5 *
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.
10 *
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.
15 *
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
19 */
20
21/**
22 * @file
23 * DNN native backend implementation.
24 */
25
26#include "dnn_backend_native.h"
27#include "libavutil/avassert.h"
28#include "dnn_backend_native_layer_mathunary.h"
29
30int dnn_load_layer_math_unary(Layer *layer, AVIOContext *model_file_context, int file_size)
31{
32 DnnLayerMathUnaryParams *params;
33 int dnn_size = 0;
34 params = av_malloc(sizeof(*params));
35 if(!params)
36 return 0;
37
38 params->un_op = (int32_t)avio_rl32(model_file_context);
39 dnn_size += 4;
40 layer->params = params;
41 layer->input_operand_indexes[0] = (int32_t)avio_rl32(model_file_context);
42 layer->output_operand_index = (int32_t)avio_rl32(model_file_context);
43 dnn_size += 8;
44
45 return dnn_size;
46
47}
48
49int dnn_execute_layer_math_unary(DnnOperand *operands, const int32_t *input_operand_indexes,
50 int32_t output_operand_index, const void *parameters)
51{
52 const DnnOperand *input = &operands[input_operand_indexes[0]];
53 DnnOperand *output = &operands[output_operand_index];
54 const DnnLayerMathUnaryParams *params = (const DnnLayerMathUnaryParams *)parameters;
55 int dims_count;
56 const float *src;
57 float *dst;
58
59 for (int i = 0; i < 4; ++i)
60 output->dims[i] = input->dims[i];
61
62 output->data_type = input->data_type;
63 output->length = calculate_operand_data_length(output);
64 output->data = av_realloc(output->data, output->length);
65 if (!output->data)
66 return DNN_ERROR;
67
68 dims_count = calculate_operand_dims_count(output);
69 src = input->data;
70 dst = output->data;
71
72 switch (params->un_op) {
73 case DMUO_ABS:
74 for (int i = 0; i < dims_count; ++i)
75 dst[i] = FFABS(src[i]);
76 return 0;
77 default:
78 return -1;
79 }
80}
libavfilter/dnn/dnn_backend_native_layer_mathunary.h+45−0
@@ -0,0 +1,45 @@
1/*
2 * Copyright (c) 2020
3 *
4 * This file is part of FFmpeg.
5 *
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.
10 *
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.
15 *
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
19 */
20
21/**
22 * @file
23 * DNN inference functions interface for native backend.
24 */
25
26#ifndef AVFILTER_DNN_DNN_BACKEND_NATIVE_LAYER_MATHUNARY_H
27#define AVFILTER_DNN_DNN_BACKEND_NATIVE_LAYER_MATHUNARY_H
28
29#include "libavformat/avio.h"
30#include "dnn_backend_native.h"
31
32typedef enum {
33 DMUO_ABS = 0,
34 DMUO_COUNT
35} DNNMathUnaryOperation;
36
37typedef struct DnnLayerMathUnaryParams{
38 DNNMathUnaryOperation un_op;
39} DnnLayerMathUnaryParams;
40
41int dnn_load_layer_math_unary(Layer *layer, AVIOContext *model_file_context, int file_size);
42int dnn_execute_layer_math_unary(DnnOperand *operands, const int32_t *input_operand_indexes,
43 int32_t output_operand_index, const void *parameters);
44
45#endif
libavfilter/dnn/dnn_backend_native_layers.c+2−0
@@ -25,6 +25,7 @@ · 24 unmodified lines
25#include "dnn_backend_native_layer_depth2space.h"25#include "dnn_backend_native_layer_depth2space.h"
26#include "dnn_backend_native_layer_maximum.h"26#include "dnn_backend_native_layer_maximum.h"
27#include "dnn_backend_native_layer_mathbinary.h"27#include "dnn_backend_native_layer_mathbinary.h"
28#include "dnn_backend_native_layer_mathunary.h"
28 29
29LayerFunc layer_funcs[DLT_COUNT] = {30LayerFunc layer_funcs[DLT_COUNT] = {
30 {NULL, NULL},31 {NULL, NULL},
@@ -33,4 +34,5 @@ LayerFunc layer_funcs[DLT_COUNT] = { · 2 unmodified lines
33 {dnn_execute_layer_pad, dnn_load_layer_pad},34 {dnn_execute_layer_pad, dnn_load_layer_pad},
34 {dnn_execute_layer_maximum, dnn_load_layer_maximum},35 {dnn_execute_layer_maximum, dnn_load_layer_maximum},
35 {dnn_execute_layer_math_binary, dnn_load_layer_math_binary},36 {dnn_execute_layer_math_binary, dnn_load_layer_math_binary},
37 {dnn_execute_layer_math_unary, dnn_load_layer_math_unary},
36};38};
tools/python/convert_from_tensorflow.py+15−1
@@ -70,8 +70,9 @@ class TFConverter: · 69 unmodified lines
70 self.converted_nodes = set()70 self.converted_nodes = set()
71 self.conv2d_scope_names = set()71 self.conv2d_scope_names = set()
72 self.conv2d_scopename_inputname_dict = {}72 self.conv2d_scopename_inputname_dict = {}
73 self.op2code = {'Conv2D':1, 'DepthToSpace':2, 'MirrorPad':3, 'Maximum':4, 'MathBinary':5}73 self.op2code = {'Conv2D':1, 'DepthToSpace':2, 'MirrorPad':3, 'Maximum':4, 'MathBinary':5, 'MathUnary':6}
74 self.mathbin2code = {'Sub':0, 'Add':1, 'Mul':2, 'RealDiv':3, 'Minimum':4}74 self.mathbin2code = {'Sub':0, 'Add':1, 'Mul':2, 'RealDiv':3, 'Minimum':4}
75 self.mathun2code = {'Abs':0}
75 self.mirrorpad_mode = {'CONSTANT':0, 'REFLECT':1, 'SYMMETRIC':2}76 self.mirrorpad_mode = {'CONSTANT':0, 'REFLECT':1, 'SYMMETRIC':2}
76 self.name_operand_dict = {}77 self.name_operand_dict = {}
77 78
@@ -286,6 +287,17 @@ class TFConverter: · 208 unmodified lines
286 np.array([output_operand_index], dtype=np.uint32).tofile(f)287 np.array([output_operand_index], dtype=np.uint32).tofile(f)
287 288
288 289
290 def dump_mathunary_to_file(self, node, f):
291 self.layer_number = self.layer_number + 1
292 self.converted_nodes.add(node.name)
293 i0_node = self.name_node_dict[node.input[0]]
294 np.array([self.op2code['MathUnary'], self.mathun2code[node.op]], dtype=np.uint32).tofile(f)
295 input_operand_index = self.add_operand(i0_node.name, Operand.IOTYPE_INPUT)
296 np.array([input_operand_index], dtype=np.uint32).tofile(f)
297 output_operand_index = self.add_operand(node.name, Operand.IOTYPE_OUTPUT)
298 np.array([output_operand_index],dtype=np.uint32).tofile(f)
299
300
289 def dump_layers_to_file(self, f):301 def dump_layers_to_file(self, f):
290 for node in self.nodes:302 for node in self.nodes:
291 if node.name in self.converted_nodes:303 if node.name in self.converted_nodes:
@@ -307,6 +319,8 @@ class TFConverter: · 15 unmodified lines
307 self.dump_maximum_to_file(node, f)319 self.dump_maximum_to_file(node, f)
308 elif node.op in self.mathbin2code:320 elif node.op in self.mathbin2code:
309 self.dump_mathbinary_to_file(node, f)321 self.dump_mathbinary_to_file(node, f)
322 elif node.op in self.mathun2code:
323 self.dump_mathunary_to_file(node, f)
310 324
311 325
312 def dump_operands_to_file(self, f):326 def dump_operands_to_file(self, f):
tools/python/convert_header.py+1−1
@@ -23,4 +23,4 @@ str = 'FFMPEGDNNNATIVE' · 22 unmodified lines
23major = 123major = 1
24 24
25# increase minor when we don't have to re-convert the model file25# increase minor when we don't have to re-convert the model file
26minor = 526minor = 6

Keyboard shortcuts

General

Show keyboard shortcuts
?
Focus the current search
/

Issues and pull requests

Open filters
f

Issues, pull requests and Inbox

Select next item
j↓
Select previous item
k↑
Open selected item
Enter
Dismiss filters or clear selection
Esc

Repository code

Go to file
t