google-deepmind / alphastar Goto Github PK

File "/home/jbx2050/alphastar/alphastar/unplugged/scripts/evaluate.py", line 96, in
app.run(main)
File "/home/jbx2050/alphastar/lib/python3.10/site-packages/absl/app.py", line 308, in run
_run_main(main, args)
File "/home/jbx2050/alphastar/lib/python3.10/site-packages/absl/app.py", line 254, in _run_main
sys.exit(main(argv))
File "/home/jbx2050/alphastar/alphastar/unplugged/scripts/evaluate.py", line 93, in main
eval_actor.run_episode(eval_actor.setup_agent())
File "/home/jbx2050/alphastar/lib/python3.10/site-packages/alphastar/modules/evaluator.py", line 460, in run_episode
run_loop.play_episode(
File "/home/jbx2050/alphastar/lib/python3.10/site-packages/alphastar/modules/run_loop.py", line 54, in play_episode
agent_output, logs = agent.step(timestep)
File "/home/jbx2050/alphastar/lib/python3.10/site-packages/alphastar/modules/evaluator.py", line 255, in step
return self._step(timestep, self._get_params())
File "/home/jbx2050/alphastar/lib/python3.10/site-packages/alphastar/modules/evaluator.py", line 138, in _step
with jax_utils.no_jax_compilation_allowed(
File "/usr/lib/python3.10/contextlib.py", line 135, in enter
return next(self.gen)
File "/home/jbx2050/alphastar/lib/python3.10/site-packages/alphastar/commons/jax_utils.py", line 54, in no_jax_compilation_allowed
previous_backend_compile = jax._src.dispatch.backend_compile
AttributeError: module 'jax' has no attribute '_src'

Evaluation is great, how can we actually use it to play in the standard py-sc2 environment ?

I ran the following script and received this error

python alphastar/unplugged/scripts/train.py --config=alphastar/unplugged/configs/alphastar_supervised.py:alphastar.dummy --config.train.max_number_of_frames=16 --config.train.learner_kwargs.batch_size=4 --config.train.datasource.kwargs.shuffle_buffer_size=16 --config.train.optimizer_kwargs.lr_frames_before_decay=4 --config.train.learner_kwargs.unroll_len=3 --config.train.datasource.name=DummyDataSource

Traceback (most recent call last):
File "/xxx/home/xxx/alphastar/alphastar/unplugged/scripts/train.py", line 57, in
from alphastar.architectures import architectures
File "/xxx/home/xxx/alphastar/alphastar/architectures/architectures.py", line 22, in
from alphastar.architectures.dummy import dummy
File "/xxx/home/xxx/alphastar/alphastar/architectures/dummy/dummy.py", line 27, in
from alphastar.architectures.components import units
File "/xxx/home/xxx/alphastar/alphastar/architectures/components/units.py", line 25, in
from alphastar.architectures.components.static_data import unit_encoder_data
File "/xxx/home/xxx/alphastar/alphastar/architectures/components/static_data/unit_encoder_data.py", line 23, in
from pysc2.env.converter.cc.game_data.python import uint8_lookup
ImportError: cannot import name 'uint8_lookup' from 'pysc2.env.converter.cc.game_data.python' (/home/scratch/xxx/pysc2/pysc2/env/converter/cc/game_data/python/init.py)

Python: 3.9.0

Can you please provide the complete package version?

Hey!

I have been studying your paper and trying to get inspiration for my current DRL projects. I did not find the transformer you mention in the paper to add memory to the model. I found a paper from you called "STABILIZING TRANSFORMERS FOR REINFORCEMENT LEARNING" where you mention transformers for RL are too unstable to learn and propose some updates to the original arch. I am facing this problem on my projects and I was curious about your approach to solve this issue in this alphastar implementation! Did you use the approach of the paper?

On the other hand, I did not find the value network in the model. Does your value net need a complex and large MLP? Do you use any residual blocks?

Hope we can discuss this topics. You can reach me out if you prefer at [email protected]. Having the oportunity to see and learn from this alphastar implementation is really cool!

Regards,

Rodrigo

OS: Ubuntu-22.04.3
python: 3.10.12

Followed the quick start guide and installed AlphaStar in a python virtual environment using 'pip install .' from the main directory where AlphaStar and setup.py is located. I then went to the unplugged quickstart and tried to train with dummy data and got the following jax error.

Traceback (most recent call last):
File "/home/matt/dev/alphastar/alphastar/unplugged/scripts/train.py", line 56, in
from acme.jax import utils
File "/home/matt/dev/rocket/lib/python3.10/site-packages/acme/jax/utils.py", line 190, in
devices: Optional[Sequence[jax.xla.Device]] = None,
File "/home/matt/dev/rocket/lib/python3.10/site-packages/jax/_src/deprecations.py", line 53, in getattr
raise AttributeError(f"module {module!r} has no attribute {name!r}")
AttributeError: module 'jax' has no attribute 'xla'

Jax versions:
jax-0.4.14
jaxlib-0.4.14.

I've tried with a previous version that another user posted about setting them to both 0.3.2, but that did not work.
I've also tried decrementing the python version to 3.9 since that's what's recommended but that did not work.

Any help on what I can try would be much appreciated.