@@ -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;
}
@@ -32,6 +32,7 @@
typedef enum {
DMBO_SUB = 0,
+ DMBO_ADD = 1,
DMBO_COUNT
} DNNMathBinaryOperation;
@@ -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):
@@ -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
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@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(-)