Known issues with colab

[lanctot]

It seems that the some packages on colab servers have been upgraded (e.g. sonnet) causing some backward-compatibility issues with the Python learning algorithms. (See some discussion here: #190)

We know about it, but we're not sure when we will be able to fix this. If anybody works out what to fix, a PR would be greatly appreciated!

[lanctot]

Quick update that we are working on getting OpenSpiel to work on Ubuntu 20.04 (see #166). I think those fixes will also mostly fix colab, so stay tuned!

[yarncraft]

Hello again @lanctot, I have been able to construct a Dockerized OpenSpiel JupyterLab so that users can experiment with the code and write new algorithms in the popular Jupyter Notebook environment.

Thanks to Docker you can also deploy it on a larger VM without a hassle (we are doing this with a Compute Optimized Azure instance at the moment) and have an amazing development workflow.

When I have some time left I will push my changes!

[lanctot]

This looks awesome! Thanks @yarncraft for your continued contributions :)

[sullins2]

This might not be a complete fix, but in some cases simply reverting to Tensorflow 1.x will make everything work. I have been using policy_gradients.py, for example, and currently Colab will spit out a bunch of errors. Simply run a cell with this first and it will work:

%tensorflow_version 1.x

[lanctot]

Nice, thanks @sullins2!

And, actually, small update on this: thanks to @dhennes, our Monday update will have a fix that should solve a lot of the current problems with colab, too!

[yarncraft]

I submitted a PR that offers a good alternative:
#219