Tensorflow Inception network generates wrong output

[idata]

Hi,

I am trying to run inceptionV1and V3 using the examples on NC AP Zoo for Tensorflow and the pre-trained checkpoints. Compilation and generation of the graph is successful but unfortunately, no matter what I am feeding to the graph the output is always the same!

('Number of categories:', 1001) Start download to NCS... ******************************************************************************* inception-v3 on NCS ******************************************************************************* (112, 'nematode, nematode worm, roundworm', 0.021683) (550, 'envelope', 0.01712) (624, 'letter opener, paper knife, paperknife', 0.01226) (474, 'can opener, tin opener', 0.01178) (513, 'corkscrew, bottle screw', 0.010948) ******************************************************************************* Finished

Apparently even the probability output is too small and also different from the expected outputs from running the exact same checkpoint using Tensorflow for the same test image. Any idea?

I am curios to know why OUTPUT_NODE_FLAG = -on is InceptionV3/Predictions/Reshape_1 while the output is coming from "out" after "InceptionV3/Predictions/Softmax".

In all of the Tensorflow example MobileNet is the only graph which generating the correct output. I also have tried the Caffe version of GoogleNet and AlexNet which all work correctly with my sticker.

Thanks.

[idata]

@Minerva Thanks for bringing this to our attention. Please do me a favor and try running this command in your inception_v1 and v3 directory: python3 ./run.py and let me know what you get. Thanks.

[idata]

I have got a similar issue, and by applying your suggestion things got solved.

However, if I want to use ROS, I am bounded to use python2.7 instead of python3. Is there any reason that the inception network cannot work with python2.7, even though release note on V1.09 claims that the API supports python2.7 as well.

[idata]

@sradmard Please see my comment @ https://ncsforum.movidius.com/discussion/596/two-problem-about-applying-movidius-on-the-raspberry-pi3#latest

[idata]

I also had the same issue, running as python3 resolved it, thanks @Tome_at_Intel