Resetting environment state in PPO `training_step` function

[nic-barbara]

Hi there,

I'm currently working with a recurrent PPO version of the PPO training algorithm in brax. For recurrent PPO, it's best to evaluate each training step (within a training epoch) over a complete trajectory of the environment - i.e: starting from just after it was last terminated/reset up until it terminates again.

To test things out, I'm looking to reset the environment in the PPO training_step function. However, I'm a little confused as to how to do this efficiently and correctly. I've tried replacing these lines:

training_state, state, key = carry
key_sgd, key_generate_unroll, new_key = jax.random.split(key, 3)

with this:

training_state, _, key = carry
key_sgd, key_generate_unroll, new_key = jax.random.split(key, 3)
state = reset_fn(key_envs)

where reset_fn = jax.jit(jax.vmap(env.reset)). Doing this means I end up with an error regarding array sizes (the 256 comes from me using num_envs = 256 for training).

  File "/home/nbar5346/code/Robust-RL/robustrl/experiment.py", line 219, in train
    _, params, metrics = train_fn(environment=self.env,
  File "/home/nbar5346/code/Robust-RL/robustrl/recurrentppo/train.py", line 575, in train
    training_epoch_with_timing(training_state, env_state, epoch_keys)
  File "/home/nbar5346/code/Robust-RL/robustrl/recurrentppo/train.py", line 487, in training_epoch_with_timing
    result = training_epoch(training_state, env_state, key)
  File "/home/nbar5346/code/Robust-RL/robustrl/recurrentppo/train.py", line 466, in training_epoch
    (training_state, state, _), loss_metrics = jax.lax.scan(
  File "/home/nbar5346/code/Robust-RL/robustrl/recurrentppo/train.py", line 428, in training_step
    (state, _), data = jax.lax.scan(f, init_carry, (),
  File "/home/nbar5346/code/Robust-RL/robustrl/recurrentppo/train.py", line 416, in f
    next_state, data = acting.generate_unroll(
  File "/home/nbar5346/code/Robust-RL/robustrl/recurrentppo/acting.py", line 86, in generate_unroll
    (final_state, _), data = jax.lax.scan(
  File "/home/nbar5346/code/Robust-RL/robustrl/recurrentppo/acting.py", line 82, in f
    nstate, transition = actor_step(
  File "/home/nbar5346/code/Robust-RL/robustrl/recurrentppo/acting.py", line 50, in actor_step
    nstate = env.step(env_state, actions)
  File "/home/nbar5346/code/Robust-RL/venv/lib/python3.10/site-packages/brax/envs/wrappers/training.py", line 122, in step
    state = self.env.step(state, action)
  File "/home/nbar5346/code/Robust-RL/venv/lib/python3.10/site-packages/brax/envs/wrappers/training.py", line 228, in step
    res = jax.vmap(step, in_axes=[self._in_axes, 0, 0])(
ValueError: vmap got inconsistent sizes for array axes to be mapped:
  * most axes (154 of them) had size 1, e.g. axis 0 of argument s.pipeline_state.solver_niter of type int32[1,256];
  * some axes (7 of them) had size 256, e.g. axis 0 of argument sys.body_ipos of type float32[256,3,3]

I'm using my own versions of the train.py and acting.py files which are slightly modified versions from brax v0.10.5. I have made only minimal changes to those files so I doubt they are the cause of this issue. I can train a policy as normal without introducing the additional reset.

Any help would be greatly appreciated, and happy to provide more info as required!

[btaba]

Hi @nic-barbara

It looks like you need to reshape key_envs to correspond to the correct batch size (see here).