jcwleo / curiosity-driven-exploration-pytorch Goto Github PK

Hello, could you please explain the version of each package you are using?

In the following locations of the code:

https://github.com/jcwleo/curiosity-driven-exploration-pytorch/blob/master/envs.py#L188-L189
https://github.com/jcwleo/curiosity-driven-exploration-pytorch/blob/master/envs.py#L286-L287

history is updated assuming the history size is 4. Shouldn't it instead be

self.history[:self.history_size-1, :, :] = self.history[1:, :, :]
self.history[self.history_size-1, :, :] = self.pre_proc(obs)

There was an error when I ran the file eval.py.can you tell me why?

Int reward values are anomaly small for sparse reward environments.
This is normal? if so, why?

Venture example:

I got many errors like:

File ".../curiosity-driven-exploration-pytorch/envs.py", line 266, in run
obs, reward, done, info = self.env.step(action)
File ".../envs/p3-torch10/lib/python3.6/site-packages/nes_py/wrappers/binary_to_discrete_space_env.py", line 67, in step
return self.env.step(self._action_map[action])
File ".../envs/p3-torch10/lib/python3.6/site-packages/gym/wrappers/time_limit.py", line 31, in step
observation, reward, done, info = self.env.step(action)
File ".../envs/p3-torch10/lib/python3.6/site-packages/gym/envs/atari/atari_env.py", line 88, in step
action = self._action_set[a]
IndexError: index 130 is out of bounds for axis 0 with size 4

Hi! For the inverse net in the ICM model, you did not use a SoftMax activation function.
But according to the paper:
In case a_t (action) is discrete, the output of g is a soft-max distribution across all possible actions.

Hi,

I can only see that you optimize the intrinsic loss in your code. Can you point me to the line where you add the intrinsic rewards to the actual environment/extrinsic rewards?

In some areas of your code I can see comments like
# total reward = int reward
which would, according to the original paper, be wrong, no?

Thank you.

Hi, I can't wrap my head around why you're using residual connections/blocks. I've seen it in other ICM implementations as well but where is that written in the paper? On what basis do we use it?

Thanks a lot!

Hello, I would like to ask whether you have tested your implementation in ViZDoom environment, like what was done in the original paper? Thanks!

Hello, I trained the model according to the parameters at the beginning but it never converged, I would like to ask what is the problem, can you provide the parameters when you trained the model? This advice is very important for me. The results of my training are shown in the figure below and the results are very bad.