ttbrunner / alphasheep Goto Github PK

the train_rl_agent.py was extended to a graphical display of the winrate
train_rl_agent_graph.zip

Update Readme by requirements and starting procedures.

There is a bug with the way we use Keras / the Adam Optimizer.
For any TF version >=2.0.0rc0, training crashes because Keras tries to get gradients in eager mode.
For this reason, the TF version is currently fixed to >2.0.0b;<2.0.0rc0 in requirements.txt.

See tensorflow/tensorflow#32646

Maybe we can use this import workaround. (Is it a workaround? Or did we call it in the wrong way?)

Currently, training is split into separate train and eval scripts. They synchronize on checkpoints on disk by renaming them.

The reason is that training is relatively fast, while evaluation is very slow.

There are several options to make this better:
a) Create a single, unified, train script that uses multiprocessing to spawn eval workers
b) Just put eval in a sequential training loop.
c) Find a way to perform eval much faster (enabling a) and b))