Fixes necessary to upgrade to TF2.2 / Python 3.8 / Ubuntu 20.04

[lanctot]

These are some fixes requires to get OpenSpiel compiling and running on Ubuntu 20.04 under TF 2.2 to address #166. Note that TF2.2 requires Python 3.8.

In particular, this PR changes the following:

• add tf.disable_v2_behavior in several places
• upgrade some pip package versions (matplotlib, cvxopt, attrs).
• remove dependency on dm-sonnet, which does not play well with disable_v2_behavior
• add simple_nets.py , a very simple pure TF replacement for sonnet (MLPs and dense layers)

Tested (so far):

• I have run kuhn_nfsp.py to ensure the graph looks like what we expect
• Similarly for kuhn_policy_gradient.py

Not done yet:

• Apply Deep CFR fix we received earlier today ( DeepCFR bug fixes #248 )
• Apply A2C fix we received earlier today ( Fix bug in get_weights when using a2c loss.  #247 )
• Get opinions on changes from people who are more familiar with TF than me :)

[googlebot]

Thanks for your pull request. It looks like this may be your first contribution to a Google open source project (if not, look below for help). Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

📝 Please visit https://cla.developers.google.com/ to sign.

Once you've signed (or fixed any issues), please reply here with @googlebot I signed it! and we'll verify it.

---

What to do if you already signed the CLA

Individual signers

• It's possible we don't have your GitHub username or you're using a different email address on your commit. Check your existing CLA data and verify that your email is set on your git commits.

Corporate signers

• Your company has a Point of Contact who decides which employees are authorized to participate. Ask your POC to be added to the group of authorized contributors. If you don't know who your Point of Contact is, direct the Google project maintainer to go/cla#troubleshoot (Public version).
• The email used to register you as an authorized contributor must be the email used for the Git commit. Check your existing CLA data and verify that your email is set on your git commits.
• The email used to register you as an authorized contributor must also be attached to your GitHub account.

ℹ️ Googlers: Go here for more info.

[lanctot]

@googlebot I signed it!

[googlebot]

CLAs look good, thanks!

ℹ️ Googlers: Go here for more info.

[lanctot]

@elkhrt @fzvinicius @rfaulkner @finbarrtimbers @jblespiau do you mind taking a look at this when you have a chance?

No rush. Even if the Travis-CI tests pass, we should discuss the best time to import it (Travis-CI does not officially support Ubuntu 20.04 yet).

[jblespiau]

For this line and the others, why are you changing the versions to be >=? For Tensorflow, I understand. Why not bump it to 2.0?

For the others, I usually prefer having a specific version, except for libraries that have very strong guarantees about backward compatibility, because by having a very large space for possible number versions, it's more likely that at least one set of version fails (because the cross-product of available versions is very large).

I expect the packages to introduce breaking changes between major versions. thus, at the very least we should use attrs >= 19.1.1, < 20.0.0 to prevent potential breakages.

Thus, I would suggest to
(a) either use specific versions. It's simple, everyone understands it, and it's deterministic.
(b) If you want to give more freedom, please make sure to have a single major version. For example:
cvopt ~=1.2 will accept versions after 1.2 and before 2.0.

For simplicity, I would suggest (a) for all most of the libs. For tensorflow we can keep >= as you tested with both (why not jump to 2.0?).

[jblespiau]

(let's just change what you need for this CL? Likely just TF, we can look at whether to upgrade other dependencies/do the upgrade of other dependencies when you are back. I have an internal CL).

[jblespiau]

Remove the future imports.

[jblespiau]

All the "tf.disable_v2_behavior()" are indeed correct. Internally, we have this done by a dependency in the BUILD file, which is not exported, which is why it was breaking.

As far as I have understood,
(a) sonnet can be used with TF2.
(b) you want to support both TF1 and TF2?

Which is why you created the alternative MLP and other classes? I did not look into Sonnet, but is the same version of sonnet working with both TF1 an TF2? It is still possible to use sonnet instead of having extra classes?

Otherwise, everything seems good to me.

[lanctot]

Thanks @jblespiau , most of these have been resolved by email.

For everyone else: the choices to loosen the required versions is to help with this transition period, i.e. since TF2.2 needs python >= 3.8 (which at least requires installing separately on Ubuntu 18.04, but also seemed to cause problems for us, on Debian 10), I'd prefer to keep Python 3.6/3.7 + TF1.15 support until it is clearly safe to remove it.

I'm running the Travis tests now to see if these changes will still work on Ubuntu 18.04 + MacOS.

[lanctot]

I'm running the Travis tests now to see if these changes will still work on Ubuntu 18.04 + MacOS.

They passed! I'll import it soon, likely some time this week.