dnn_backend_native_layer_mathunary: add abs support
authorTing Fu <ting.fu@intel.com>
Mon, 25 May 2020 14:46:26 +0000 (22:46 +0800)
committerGuo, Yejun <yejun.guo@intel.com>
Thu, 28 May 2020 03:04:21 +0000 (11:04 +0800)
commitf73cc61bf5aa383048979f4de2023877c522f6be
tree61508bdb3d3751f1fc01db2cbc6a93c0e76ac5d0
parentb6d6597bef66531ec07c07a7125b88aee38fb220
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))

Signed-off-by: Ting Fu <ting.fu@intel.com>
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
libavfilter/dnn/Makefile
libavfilter/dnn/dnn_backend_native.h
libavfilter/dnn/dnn_backend_native_layer_mathunary.c [new file with mode: 0644]
libavfilter/dnn/dnn_backend_native_layer_mathunary.h [new file with mode: 0644]
libavfilter/dnn/dnn_backend_native_layers.c
tools/python/convert_from_tensorflow.py
tools/python/convert_header.py