gym-unity: Int64 isInstance(int) not successful on Windows #1448
Description
I am currently trying to run the gym-unity wrapper on Windows.
As mentioned before, the Readme is not up to date, and DQN Baseline example is not working, but this can be easily fixed by using the current method defintion for deepq.learn().
While running the GridWorld example, i encountered a different error:
Traceback (most recent call last):
File "C:/Users/ex/Documents/ml-agents/train_agent.py", line 30, in <module>
main()
File "C:/Users/ex/Documents/ml-agents/train_agent.py", line 23, in main
gamma=0.99
File "C:\Users\ex\AppData\Local\Programs\Python\Python36\lib\site-packages\baselines\deepq\deepq.py", line 282, in learn
new_obs, rew, done, _ = env.step(env_action)
File "C:\Users\ex\AppData\Local\Programs\Python\Python36\lib\site-packages\gym_unity\envs\unity_env.py", line 131, in step
info = self._env.step(action)[self.brain_name]
File "C:\Users\ex\AppData\Local\Programs\Python\Python36\lib\site-packages\mlagents\envs\environment.py", line 327, in step
for brain_name in list(vector_action.keys()) + list(memory.keys()) + list(text_action.keys()):
AttributeError: 'numpy.int64' object has no attribute 'keys'
While debugging, i found the reason to be a failing type check in gym_unity\envs\environment.py, line 281:
if isinstance(vector_action, (int, np.int_, float, np.float_, list, np.ndarray)):
vector_action is an int64 datatype and on Windows np.int_ evaluates to a standard int32 datytype.
Similar typechecks are also at different spots in the code.
A hotfix for me was to cast the value in \baselines\deepq\deepq.py, line 280:
if isinstance(action, np.int64):
env_action = np.int32(action)
else:
env_action = action
but this obviously is not a real solution. Does the action value have to be an int64 integer? Has anyone else encountered a similar problem?
I might add that normal baseline examples (e.g. the CartPole training) work fine.
System Specs:
- MS Windows 7 Enterprise 64-Bit
- Python 3.6.7
- Numpy 1.14
- Tensorflow 1.7.1 CPU-Version