train CA demo - Dead Kernel

[williamgilpin]

train CA demo - Dead Kernel

[williamgilpin]

Hi @21dmohan thanks for checking out my project! It shouldn't be that intensive, if I recall correctly I also just used normal tensorflow, not tensorflow GPU. My specs also are pretty comparable:

Do you happen to know which line in cell 3 is crashing it? Is it potentially a problem with plt.imshow?
Or is it definitely the sess.run? If the latter, then I'm concerned that this might be a tensorflow version issue---when I wrote this I was using tensorflow 1.

[21dmohan]

Interesting, it appears to be an issue with plt.imshow (the kernel dies at the fourth cell):

Do you have any suggestions as to how to fix this?

[williamgilpin]

Okay that's a relief. What is out.shape? Does matplotlib otherwise work? For example,

plt.imshow(np.identity(3))

[21dmohan]

It appears that my matplotlib is not working:

I'll try and sort that out! Thank you so much! I'll let you know if I have any questions :)

[21dmohan]

Quick update (in case someone else has this issue).

Following the suggestion from this post: https://superuser.com/questions/1406008/jupyter-kernel-dies-when-attempting-a-simple-plot-python

I just had to install nomkl: conda install nomkl

[williamgilpin]

That's good to know, thank you for posting your fix here. I hope you enjoy the package---let me know if you have any other questions.

[williamgilpin]

I'm re-opening this issue because I lost an hour to this bug today. Here is what I needed to do in order to get this code (and all my other tensorflow projects) functional today.

1. Make a fresh conda environment, but don't use the latest version of python because tensorflow doesn't support it.

conda create -n nn python=3.6.9

2. Install the following packages all together. For some reason everything breaks if nomkl is installed after the other packages

conda install nomkl numpy scipy scikit-learn numexpr tensorflow matplotlib

[williamgilpin]

It looks like the latest version of TensorFlow plays nicely with everything, and so I will re-close this issue for now.