Deconvolution operation produces different results in tensorflow and CV- SDK

[Feldman__Daniel]

My goal was to convert unet architecture from tensorflow (with python API) to IR and to run the network on the inference engine.I run the IR and got different results from the tensorflow.

I was able to isolate the root cause of the different results to the transposed convolution (deconv) layer (convolution and max-pool layers works fine).

My execution environment is composed of two parts:

1. Ubuntu 17.10 x64 with python3.6, tensorflow 1.8, CV-SDK 2018.0.234, used for:

   • Creating and freezing the tensorflow model.

   • Inference of the frozen graph in tensorflow.

   • Converting the frozen graph to IR.

2. Windows 10 x64 with Visual Studio 2015, CV-SDK 2018.0.234 (also with 2018.1.249), used for:

   • Inference (with the engine)
      

I attached both the frozen graph and the IR which was used for the deconv isolated comparison (and the tensorflow code which was used for creating and inferring the frozen graph).
In addition, I added the code of the inference with the cv-sdk (it is mostly based on the hello_classification sample).

To illustrate the different results (with a matrix of ones as input), I calculated the sum of the output matrix:
Tensorflow output sum is -12.744938.
CV-SDK output sum is 2.86543.

Any help appreciated.

[Monique_J_Intel]

Hi Daniel,

I reproduced your issue and got the OpenVINO/CV-SDK sum at -12.7449. I have attached a screenshot from my ubuntu 16.04 skull canyon NUC and I am using the latest external version of OpenVINO which is version 2018.2.319 that was put out July 18. I highly encourage you to update your OpenVINO package to the latest version if you are still seeing this issue and let me know if the issue still persist.

Kind Regards,

Monique Jones