Stable Baselines3 RuntimeError: mat1 and mat2 must have the same dtype

Question

I am trying to implement SAC with a custom environment in Stable Baselines3 and I keep getting the error in the title. The error occurs with any off policy algorithm not just SAC.

Traceback:

File "<MY PROJECT PATH>\src\main.py", line 70, in <module>
  main()
File "<MY PROJECT PATH>\src\main.py", line 66, in main
  model.learn(total_timesteps=timesteps, reset_num_timesteps=False, tb_log_name=f"sac_{num_cars}_cars")
File "<MY PROJECT PATH>\venv\lib\site-packages\stable_baselines3\sac\sac.py", line 309, in learn
  return super().learn(
File "<MY PROJECT PATH>\venv\lib\site-packages\stable_baselines3\common\off_policy_algorithm.py", line 375, in learn
  self.train(batch_size=self.batch_size, gradient_steps=gradient_steps)
File "<MY PROJECT PATH>\venv\lib\site-packages\stable_baselines3\sac\sac.py", line 256, in train
  current_q_values = self.critic(replay_data.observations, replay_data.actions)
File "<MY PROJECT PATH>\venv\lib\site-packages\torch\nn\modules\module.py", line 1190, in _call_impl
  return forward_call(*input, **kwargs)
File "<MY PROJECT PATH>\venv\lib\site-packages\stable_baselines3\common\policies.py", line 885, in forward
  return tuple(q_net(qvalue_input) for q_net in self.q_networks)
File "<MY PROJECT PATH>\venv\lib\site-packages\stable_baselines3\common\policies.py", line 885, in <genexpr>
  return tuple(q_net(qvalue_input) for q_net in self.q_networks)
File "<MY PROJECT PATH>\venv\lib\site-packages\torch\nn\modules\module.py", line 1190, in _call_impl
  return forward_call(*input, **kwargs)
File "<MY PROJECT PATH>\venv\lib\site-packages\torch\nn\modules\container.py", line 204, in forward
  input = module(input)
File "<MY PROJECT PATH>\venv\lib\site-packages\torch\nn\modules\module.py", line 1190, in _call_impl
  return forward_call(*input, **kwargs)
File "<MY PROJECT PATH>\venv\lib\site-packages\torch\nn\modules\linear.py", line 114, in forward
  return F.linear(input, self.weight, self.bias)
RuntimeError: mat1 and mat2 must have the same dtype

Action and observation spaces:

self.action_space = Box(low=-1., high=1., shape=(2,), dtype=np.float)
self.observation_space = Box(
    np.array(
        [-np.inf] * (9 * 40) + [-np.inf] * 3 + [-np.inf] * 3 + [-np.inf] * 3
        + [0.] + [0.] + [0.] + [-1.] + [0.] * 4 + [0.] * 4 + [0.] * 4,
        dtype=np.float
    ),
    np.array(
        [np.inf] * (9 * 40) + [np.inf] * 3 + [np.inf] * 3 + [np.inf] * 3
        + [np.inf] + [1.] + [1.] + [1.] + [1.] * 4 + [np.inf] * 4 + [np.inf] * 4,
        dtype=np.float
    ),
    dtype=np.float
)

Observations are returned in the step and reset methods as a numpy array of floats.

Is there something I'm missing which is causing this error? If I use one of the environments that come with gym such as pendulum it works fine which is why I think I have a problem with my custom environment.

Thanks in advance for any help and please let me know if more information is required.

Trafalgar

np.float should return a float64 by default. For some reason, F.linear seems to struggle with precision. If your program allows the use of a lower precision, a quick fix might be to replace dtype=np.float with dtype='float32'. – Salic

Answer 1

Change the inputs to float32 , default the loader set the type as float64.

inputs = inputs.to(torch.float32)

Egyptian