adam optim ERROR:If capturable=False, state_steps should not be CUDA tensors.

[jaried]

• I have marked all applicable categories:
• I have visited the source website
• I have searched through the issue tracker for duplicates
• I have mentioned version numbers, operating system and environment, where applicable:

  import tianshou, gym, torch, numpy, sys
  print(tianshou.__version__, gym.__version__, torch.__version__, numpy.__version__, sys.version, sys.platform)
  0.4.8 0.21.0 1.12.0+cu113 1.20.1 3.8.8 (default, Apr 13 2021, 15:08:03) [MSC v.1916 64 bit (AMD64)] win32

  The SAC algorithm has been normal before. Recently, after reading the saved policy weight and optimizer, the first learning error is as follows. How can I solve it?

  
  Epoch #1:   0%|                                                                                                                                                                                 | 23/57600 [00:02<1:37:11,  9.87it/s]
  ('If capturable=False, state_steps should not be CUDA tensors.',)
  ===========
  Traceback (most recent call last):
  File "D:\Tony\Documents\yunpan\invest\2022\Quant\code\factor\myfactor.py", line 684, in wrapper
  func(*args, **kw)
  File "tianshou_if.py", line 378, in sac_with_il_if
  result = offpolicy_trainer(
  File "D:\Anaconda3\lib\site-packages\tianshou\trainer\offpolicy.py", line 129, in offpolicy_trainer
  return OffpolicyTrainer(*args, **kwargs).run()
  File "D:\Anaconda3\lib\site-packages\tianshou\trainer\base.py", line 425, in run
  deque(self, maxlen=0)  # feed the entire iterator into a zero-length deque
  File "D:\Anaconda3\lib\site-packages\tianshou\trainer\base.py", line 282, in __next__
  self.policy_update_fn(data, result)
  File "D:\Anaconda3\lib\site-packages\tianshou\trainer\offpolicy.py", line 118, in policy_update_fn
  losses = self.policy.update(self.batch_size, self.train_collector.buffer)
  File "D:\Anaconda3\lib\site-packages\tianshou\policy\base.py", line 277, in update
  result = self.learn(batch, **kwargs)
  File "D:\Anaconda3\lib\site-packages\tianshou\policy\modelfree\sac.py", line 149, in learn
  td1, critic1_loss = self._mse_optimizer(
  File "D:\Anaconda3\lib\site-packages\tianshou\policy\modelfree\ddpg.py", line 158, in _mse_optimizer
  optimizer.step()
  File "D:\Anaconda3\lib\site-packages\torch\optim\optimizer.py", line 109, in wrapper
  return func(*args, **kwargs)
  File "D:\Anaconda3\lib\site-packages\torch\autograd\grad_mode.py", line 27, in decorate_context
  return func(*args, **kwargs)
  File "D:\Anaconda3\lib\site-packages\torch\optim\adam.py", line 157, in step
  adam(params_with_grad,
  File "D:\Anaconda3\lib\site-packages\torch\optim\adam.py", line 213, in adam
  func(params,
  File "D:\Anaconda3\lib\site-packages\torch\optim\adam.py", line 255, in _single_tensor_adam
  assert not step_t.is_cuda, "If capturable=False, state_steps should not be CUDA tensors."
  AssertionError: If capturable=False, state_steps should not be CUDA tensors.

[jaried]

Before executing offpolicy_trainer, I added the following code:

        actor_optim.param_groups[0]['capturable'] = True
        alpha_optim.param_groups[0]['capturable'] = True
        critic1_optim.param_groups[0]['capturable'] = True
        critic2_optim.param_groups[0]['capturable'] = True

can run, what is the reason?

[jaried]

This article says that it is caused by pressing ctrl+c to end the training? https://github.com/babysor/MockingBird/issues/631

But I use the previous version of the checkpoint file in question, and the same error is reported, why is this?

[jaried]

https://github.com/pytorch/pytorch/issues/80809

Someone said this:

Hi, I am also facing the same issue when I try to load the checkpoint and resume model training on the latest pytorch (1.12).

It seems to be related with a newly introduced parameter (capturable) for the Adam and AdamW optimizers. Currently two workarounds:

1. forcing capturable = True after loading the checkpoint (as suggested above) optim.param_groups[0]['capturable'] = True . This seems to slow down the model training by approx. 10% (YMMV depending on the setup).
2. Reverting pytorch back to previous versions (I have been using 1.11.0).

I'm wondering whether enforcing capturable = True may incur unwanted side effects.

I'm also wondering about whether forcing captureable=True would have unwanted side effects. I will also return to torch1.11.