Stable-baselines: A2C evaluation error (parallel environments n_envs)

[pesapower]

Hi, I'm trying to train and evaluate a A2C model using 4 parallel environments for the training and just 1 environment for the evaluation. The code I'm using is the following:

# Initialize parallel environments (training)
train_env = make_vec_env(MyTrainEnv, env_kwargs={...}, n_envs=4, vec_env_cls=DummyVecEnv)
test_env = make_vec_env(MyTestEnv, env_kwargs={...}, n_envs=1, vec_env_cls=DummyVecEnv)

# Create callback for evaluating model during training process
eval_callback = EvalCallback(test_env,
                             log_path='./logs/test/',
                             eval_freq=1000,
                             deterministic=True,
                             render=False,
                             n_eval_episodes=10)

model = A2C(MlpLstmPolicy, train_env, tensorboard_log='./logs/train/')

# Start training
model.learn(total_timesteps=1000000, callback=eval_callback)

The training procede good until the eval_callback is triggered, at that moment the following error occurs:

  File "C:\Users\pesap\miniconda3\envs\DeepRL-HedgeFund\lib\site-packages\stable_baselines\a2c\a2c.py", line 263, in learn
    rollout = self.runner.run(callback)
  File "C:\Users\pesap\miniconda3\envs\DeepRL-HedgeFund\lib\site-packages\stable_baselines\common\runners.py", line 48, in run
    return self._run()
  File "C:\Users\pesap\miniconda3\envs\DeepRL-HedgeFund\lib\site-packages\stable_baselines\a2c\a2c.py", line 368, in _run
    if self.callback.on_step() is False:
  File "C:\Users\pesap\miniconda3\envs\DeepRL-HedgeFund\lib\site-packages\stable_baselines\common\callbacks.py", line 99, in on_step
    return self._on_step()
  File "C:\Users\pesap\miniconda3\envs\DeepRL-HedgeFund\lib\site-packages\stable_baselines\common\callbacks.py", line 305, in _on_step
    return_episode_rewards=True)
  File "C:\Users\pesap\miniconda3\envs\DeepRL-HedgeFund\lib\site-packages\stable_baselines\common\evaluation.py", line 54, in evaluate_policy
    action, state = model.predict(obs, state=state, deterministic=deterministic)
  File "C:\Users\pesap\miniconda3\envs\DeepRL-HedgeFund\lib\site-packages\stable_baselines\common\base_class.py", line 819, in predict
    actions, _, states, _ = self.step(observation, state, mask, deterministic=deterministic)
  File "C:\Users\pesap\miniconda3\envs\DeepRL-HedgeFund\lib\site-packages\stable_baselines\common\policies.py", line 505, in step
    {self.obs_ph: obs, self.states_ph: state, self.dones_ph: mask})
  File "C:\Users\pesap\miniconda3\envs\DeepRL-HedgeFund\lib\site-packages\tensorflow_core\python\client\session.py", line 956, in run
    run_metadata_ptr)
  File "C:\Users\pesap\miniconda3\envs\DeepRL-HedgeFund\lib\site-packages\tensorflow_core\python\client\session.py", line 1156, in _run
    (np_val.shape, subfeed_t.name, str(subfeed_t.get_shape())))
ValueError: Cannot feed value of shape (1, 15) for Tensor 'input/Ob:0', which has shape '(4, 15)'

In particular the last error message is:

Cannot feed value of shape (1, 15) for Tensor 'input/Ob:0', which has shape '(4, 15)'

It seems that the model still require to receive inputs from 4 parallel environments during the evaluation process (it's not possible to use EvalCallback with n_envs > 1).

Any suggestion? Thanks in advance!

Bests,
Nicola

[Miffyli]

See documentation:

One current limitation of recurrent policies is that you must test them with the same number of environments they have been trained on.

See this comment for a possible solution: https://github.com/hill-a/stable-baselines/issues/166#issuecomment-502350843

Edit: See post below. The easiest solution likely is to update stable-baselines with `pip install --upgrade git+https://github.com/hill-a/stable-baselines``.

[araffin]

Related PR and issue: https://github.com/hill-a/stable-baselines/pull/1017 and https://github.com/hill-a/stable-baselines/issues/1015

[pesapower]

Thank you very much @Miffyli and @araffin !
I upgraded stable-baselines to the latest version and now it seems to work like a charm, great work!