Sacha Chernyavskiy

Yes, exacltly. 1) `flax.nnx` is not fully working, there are some minor fixes (e.g. conv dimension calculation and tests) left 2) the implementation does work, however, there happen some problems...

Surely, we will notify you!

@lanctot, for both APIs: `linen` and `nnx` there are now tests, and they're passing. The only minor things left from the development side are ~~model export and the changelog~~ benchmarks....

@lanctot , I ran for much longer time a TTT experiment  Doesn't look good, does it? Can the picture tell you what I can look at to find bugs?

@lanctot The main difference between graphs is in buffer/batch size: 2 ** 16 and 2 ** 10, that were default values for the model I use default value of averaging...

will share some progress tomorrow, you may approve checks later

I guess, we're making slight progress, do we not?  give it a look @lanctot

The latest plots (minor tweaks and fixes here and there). Maybe, using much more resources (I used a toy config), there is smth here:

I found an example with hyperparameters for tic-tac-toe, and results look somewhat more intuitive (although, I had to reduce batch size fourfold due to the resource constraints)

@lanctot, now ttt's example works much better (see graphs, winrate is pretty nice for an example config). However, I want to test with more steps for connect4 to make sure...