google / brax

Massively parallel rigidbody physics simulation on accelerator hardware.
Apache License 2.0
2.34k stars 255 forks source link

1 ppo_test fails #200

Open erwincoumans opened 2 years ago

erwincoumans commented 2 years ago

Probably harmless but just in case:

Traceback (most recent call last):
  File "ppo_test.py", line 45, in testTrain
    self.assertGreater(metrics['eval/episode_reward'], 100 * 0.995)
AssertionError: DeviceArray(97.05763, dtype=float32) not greater than 99.5

This was on Windows, pip list output

brax                0.0.10                   f:\dev\brax
jax                 0.3.14                   f:\dev\jax
jaxlib              0.3.7
erikfrey commented 2 years ago

We just pushed some big changes to training - might work for you if you pip install from github now, or I'll let you know when these changes have made it to a version.

But yes either way, there is some instability here. If this keeps happening I'll lower the threshold.