Runtime error relating to atari.Action class

[ugurkanates]

Tensor.contains only supports Tensor or scalar, but you passed in a <class 'torchcule_atari.AtariAction'>

Just trying bunch of examples provided by NVIDIA but seems sadly there is a bug ?

python 3.7.0 I installed apex from github and did python setup install ( pip version was having issues?) pytorch 1.4.0 cuda version 10.2

full stack

`Traceback (most recent call last):
  File "vtrace_main.py", line 30, in <module>
    vtrace_main()
  File "vtrace_main.py", line 27, in vtrace_main
    sys.exit(main(vtrace_parser_options, worker))
  File "/home/paypaytr/Desktop/ReinforcementLearning/nvidiaAtari/cule/examples/utils/launcher.py", line 160, in main
    maybe_restart(args, worker)
  File "/home/paypaytr/Desktop/ReinforcementLearning/nvidiaAtari/cule/examples/utils/launcher.py", line 125, in maybe_restart
    dispatch(args, worker)
  File "/home/paypaytr/Desktop/ReinforcementLearning/nvidiaAtari/cule/examples/utils/launcher.py", line 80, in dispatch
    worker(args.local_rank, ngpus_per_node, args)
  File "/home/paypaytr/Desktop/ReinforcementLearning/nvidiaAtari/cule/examples/vtrace/train.py", line 35, in worker
    train_env, test_env, observation = env_initialize(args, env_device)
  File "/home/paypaytr/Desktop/ReinforcementLearning/nvidiaAtari/cule/examples/utils/initializers.py", line 79, in env_initialize
    clip_rewards=args.clip_rewards, frameskip=4)
  File "/home/paypaytr/anaconda3/lib/python3.7/site-packages/torchcule-0.1.0-py3.7-linux-x86_64.egg/torchcule/atari/env.py", line 95, in __init__
    self.fire_reset = torchcule_atari.FIRE in self.action_set
  File "/home/paypaytr/anaconda3/lib/python3.7/site-packages/torch/tensor.py", line 510, in __contains__
    type(element)
RuntimeError: Tensor.__contains__ only supports Tensor or scalar, but you passed in a <class 'torchcule_atari.AtariAction'>.

`

[ifrosio]

If you want to move quick, can you try the docker image? That should work out of the box.

[ugurkanates]

@ifrosio yes thats what I did. I may also update my fork and do a PR that works with all current version of libraries and Ubuntu 20.04 LTS. This project is too good to left out die.

[ifrosio]

For docker instructions we posted:

FROM nvidia/cuda:10.0-cudnn7-devel-ubuntu18.04 as base ... RUN pip install torch==1.2.0

It looks like you are using different CUDA / pytorch versions. Any particular reason for this or can you try downgrading to CUDA 10 / pytorch 1.2 / python 3.6?

[ugurkanates]

@ifrosio Well there are two things , 1 - able to use latest versions of torch 2- For CUDA I believe NVIDIA handles that part , I don't think you can even downgrade CUDA no ? I mean I can handle torch with simple conda/ virtual env but yeah atleast CUDA part is I believe should be updated or choseable. My GPU isn't even up to date it's just one of the old models came with laptops 2GB VRAM model.