-
Notifications
You must be signed in to change notification settings - Fork 725
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
ACKTR model crashes using CnnLnLstmPolicy #387
Comments
Hello, It seems that it may be related from your custom environment. from stable_baselines.common.cmd_util import make_atari_env
from stable_baselines import ACKTR
env = make_atari_env("BreakoutNoFrameskip-v4", num_env=2, seed=1)
# Reduce number of steps to avoid memory issue
model = ACKTR("CnnLnLstmPolicy", env, n_steps=4, verbose=1)
model.learn(1000)
|
That code snippet does not work for me
The same happens for other atari environments using ACKTR + CnnLstmPolicies, like
But it works fine on MlpLstmPolicy
This seems to be an ACKTR-specific issue for me, PPO2 works for all listed examples |
What is your gym version ? (+ associated, like atary-py) |
These are all I could think of |
The error seems related to tensorflow version (I could reproduce the bug in google colab) |
@araffin |
So 1.8.0 cpu version |
Thank you. |
Describe the bug
Describe the bug
ACKTR example code crashes when modified to use MlpLnLstmPolicy. Apparent bug in KFAC code
Code example
results in:
System Info
Describe the characteristic of your environment:
pip install from cloned repo
GeForce GTX 1080
Python 3.6.8
1.13.1
Additional context
seems to expect bprop to be a batch x channel tensor, but is instead a batch tensor, stems from optim.compute_and_apply_stats(self.joint_fisher, var_list=params), joint_fisher is a (838980, 32) tensor
The text was updated successfully, but these errors were encountered: