[FFmpeg-devel] [PATCH 1/6] dnn/native: add native support for 'add'

[Guo, Yejun]

It can be tested with the model file generated with below python script:

import tensorflow as tf
import numpy as np
import imageio

in_img = imageio.imread('input.jpg')
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')
z1 = 0.039 + x
z2 = x + 0.042
z3 = z1 + z2
z4 = z3 - 0.381
z5 = z4 - x
y = tf.math.maximum(z5, 0.0, 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))

Signed-off-by: Guo, Yejun <yejun.guo at intel.com>
---
 libavfilter/dnn/dnn_backend_native_layer_mathbinary.c | 13 +++++++++++++
 libavfilter/dnn/dnn_backend_native_layer_mathbinary.h |  1 +
 tools/python/convert_from_tensorflow.py               | 15 +++++++--------
 tools/python/convert_header.py                        |  2 +-
 4 files changed, 22 insertions(+), 9 deletions(-)

diff --git a/libavfilter/dnn/dnn_backend_native_layer_mathbinary.c b/libavfilter/dnn/dnn_backend_native_layer_mathbinary.c
index 3b8bab8..3fe337f 100644
--- a/libavfilter/dnn/dnn_backend_native_layer_mathbinary.c
+++ b/libavfilter/dnn/dnn_backend_native_layer_mathbinary.c
@@ -107,6 +107,19 @@ int dnn_execute_layer_math_binary(DnnOperand *operands, const int32_t *input_ope
             }
         }
         return 0;
+    case DMBO_ADD:
+        if (params->input0_broadcast || params->input1_broadcast) {
+            for (int i = 0; i < dims_count; ++i) {
+                dst[i] = params->v + src[i];
+            }
+        } else {
+            const DnnOperand *input1 = &operands[input_operand_indexes[1]];
+            const float *src1 = input1->data;
+            for (int i = 0; i < dims_count; ++i) {
+                dst[i] = src[i] + src1[i];
+            }
+        }
+        return 0;
     default:
         return -1;
     }
diff --git a/libavfilter/dnn/dnn_backend_native_layer_mathbinary.h b/libavfilter/dnn/dnn_backend_native_layer_mathbinary.h
index 6b684d1..3c5bc6b 100644
--- a/libavfilter/dnn/dnn_backend_native_layer_mathbinary.h
+++ b/libavfilter/dnn/dnn_backend_native_layer_mathbinary.h
@@ -32,6 +32,7 @@
 
 typedef enum {
     DMBO_SUB = 0,
+    DMBO_ADD = 1,
     DMBO_COUNT
 } DNNMathBinaryOperation;
 
diff --git a/tools/python/convert_from_tensorflow.py b/tools/python/convert_from_tensorflow.py
index 2485f16..9a495c0 100644
--- a/tools/python/convert_from_tensorflow.py
+++ b/tools/python/convert_from_tensorflow.py
@@ -71,7 +71,7 @@ class TFConverter:
         self.conv2d_scope_names = set()
         self.conv2d_scopename_inputname_dict = {}
         self.op2code = {'Conv2D':1, 'DepthToSpace':2, 'MirrorPad':3, 'Maximum':4, 'MathBinary':5}
-        self.mathbin2code = {'Sub':0}
+        self.mathbin2code = {'Sub':0, 'Add':1}
         self.mirrorpad_mode = {'CONSTANT':0, 'REFLECT':1, 'SYMMETRIC':2}
         self.name_operand_dict = {}
 
@@ -255,8 +255,7 @@ class TFConverter:
         np.array([input_operand_index, output_operand_index], dtype=np.uint32).tofile(f)
 
 
-    def dump_sub_to_file(self, node, f):
-        assert(node.op == 'Sub')
+    def dump_mathbinary_to_file(self, node, f):
         self.layer_number = self.layer_number + 1
         self.converted_nodes.add(node.name)
         i0_node = self.name_node_dict[node.input[0]]
@@ -264,15 +263,13 @@ class TFConverter:
         np.array([self.op2code['MathBinary'], self.mathbin2code[node.op]], dtype=np.uint32).tofile(f)
         if i0_node.op == 'Const':
             scalar = i0_node.attr['value'].tensor.float_val[0]
-            assert(i0_node.name.find('sub/x'))
-            np.array([1], dtype=np.uint32).tofile(f)
+            np.array([1], dtype=np.uint32).tofile(f)            # broadcast: 1
             np.array([scalar], dtype=np.float32).tofile(f)
-            np.array([0], dtype=np.uint32).tofile(f)
+            np.array([0], dtype=np.uint32).tofile(f)            # broadcast: 0
             input_operand_index = self.add_operand(i1_node.name, Operand.IOTYPE_INPUT)
             np.array([input_operand_index], dtype=np.uint32).tofile(f)
         elif i1_node.op == 'Const':
             scalar = i1_node.attr['value'].tensor.float_val[0]
-            assert(i1_node.name.find('sub/y'))
             np.array([0], dtype=np.uint32).tofile(f)
             input_operand_index = self.add_operand(i0_node.name, Operand.IOTYPE_INPUT)
             np.array([input_operand_index], dtype=np.uint32).tofile(f)
@@ -309,7 +306,9 @@ class TFConverter:
             elif node.op == 'Maximum':
                 self.dump_maximum_to_file(node, f)
             elif node.op == 'Sub':
-                self.dump_sub_to_file(node, f)
+                self.dump_mathbinary_to_file(node, f)
+            elif node.op == 'Add':
+                self.dump_mathbinary_to_file(node, f)
 
 
     def dump_operands_to_file(self, f):
diff --git a/tools/python/convert_header.py b/tools/python/convert_header.py
index 6576fca..7027022 100644
--- a/tools/python/convert_header.py
+++ b/tools/python/convert_header.py
@@ -23,4 +23,4 @@ str = 'FFMPEGDNNNATIVE'
 major = 1
 
 # increase minor when we don't have to re-convert the model file
-minor = 1
+minor = 2
-- 
2.7.4