Intel® Distribution of OpenVINO™ Toolkit
Community support and discussions about the Intel® Distribution of OpenVINO™ toolkit, OpenCV, and all things computer vision-related on Intel® platforms.

LeakyReLU activations missing from xml file after conversion from pb

Slain__John
Beginner
93 Views

Hello there!

Recently i converted a pb-file to the openvino format using the mo_tf.py module. It successfully converted and generated the xml and bin files. However, there are no leaky relu functions inbetween the convolutions which means that the edges are pointing to layers that don't exist in the xml file. Is there a reason why the conversion would skip certain layers that seem to be supported? Looking at the xml file it is obvious that the conversion program noticed all the layers since all the connections at the end exist. However, it just decided to forgo writing in half of them. If more information is needed I'd be happy to add it.

0 Kudos
1 Reply
Shubha_R_Intel
Employee
93 Views

Dear Slain, John,

It looks like LeakyReLU only got supported in Tensorflow starting R1.13 according to the below doc:

https://www.tensorflow.org/api_docs/python/tf/nn/leaky_relu

Until Tensorflow R1.13, people had to find workarounds for LeakyReLU

That's why when you search for it in the OpenVino Supported Frameworks Doc you will not find LeakyRelu for Tensorflow.  Unfortunately, at this moment tensorflow 1.13 is not supported by OpenVino but it is scheduled to be supported.

But if I understand you correctly, you are saying that mo_tf.py was able to successfully generate IR . This is believable, like you said - mo_tf.py noticed it and decided to replace it with something else. That "something else" is explained as follows:

LeakyRelu is translated to Relu with negative_slope in IR.

Hope it helps. Thanks for using OpenVino !

Shubha

 

 

Reply