Example error caused by create_tensor(... keep_dimensions) defaulting to False

[NoneJou072]

Description

Hello, when I run the example files provided by the [official document] (https://github.com/Toni-SM/skrl/blob/main/docs/source/examples/deepmind/dm_manipulation_stack_sac.py), an error comes as follows.

Traceback (most recent call last):
  File "/home/zhr/zhr/skrl/docs/source/examples/deepmind/dm_manipulation_stack_sac.py", line 179, in <module>
    trainer = SequentialTrainer(cfg=cfg_trainer, env=env, agents=agent_sac)
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/zhr/zhr/skrl/skrl/trainers/torch/sequential.py", line 55, in __init__
    self.agents.init(trainer_cfg=self.cfg)
  File "/home/zhr/zhr/skrl/skrl/agents/torch/sac/sac.py", line 201, in init
    self.memory.create_tensor(name="states", size=self.observation_space, dtype=torch.float32)
  File "/home/zhr/zhr/skrl/skrl/memories/torch/base.py", line 197, in create_tensor
    size = self._get_space_size(size, keep_dimensions)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/zhr/zhr/skrl/skrl/memories/torch/base.py", line 111, in _get_space_size
    raise ValueError("keep_dimensions=True cannot be used with Dict spaces")
ValueError: keep_dimensions=True cannot be used with Dict spaces

The reason was that I used the 'dict' type as the obs space.

I noticed that a similar error was mentioned in https://github.com/Toni-SM/skrl/pull/201. Do you have any temporary solutions for using 'dict' obs space, thanks.

What skrl version are you using?

1.3.0

What ML framework/library version are you using?

Pytorch version

Additional system information

Python 3.10, Ubuntu20.04

[larsrpe]

I am facing the same issues. It does not seem like the Memory class handels dict observations. One work around is to concatenate the observations instead.

[Toni-SM]

Hi @NoneJou072 and @larsrpe

Keep dimensions change (set to True) introduced in version 1.3.0 was reverted in develop branch.

https://github.com/Toni-SM/skrl/blob/eff72951a4522e1a237e8a6df200b89fc6656a75/skrl/memories/torch/base.py#L132

In addition, spaces utilities were added to develop branch to handle fundamental (Box, Discrete and MultiDiscrete) and composite (Tuple and Dict) spaces better.

---

Please, use the develop branch while the new release is being completed

[larsrpe]

Thank you for the clarification. And sorry for not checking the devleop branch right away! Keep up the good work!