marload / deeprl-tensorflow2 Goto Github PK

In PPO_Discrete each reward is multiplied by 0.01 and in PPO_Continuous reward is also modified. I don't understand why do these modification, what does these modification do?

In the Actor net, It seems that from_logit should be set to False in tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True) as you added a softmax in the last layer:)

'Viewer' object has no attribute 'isopen'
File "E:\anaconda\envs\tf2\lib\site-packages\gym\envs\classic_control\rendering.py", line 81, in close
AttributeError: 'Viewer' object has no attribute 'isopen'
Traceback (most recent call last):
File "E:\anaconda\envs\tf2\lib\site-packages\gym\envs\classic_control\rendering.py", line 165, in del
if self.isopen and sys.meta_path:
if self.isopen and sys.meta_path:
if self.isopen and sys.meta_path:
self.close()
File "E:\anaconda\envs\tf2\lib\site-packages\gym\envs\classic_control\rendering.py", line 165, in del
if self.isopen and sys.meta_path:
AttributeError: 'Viewer' object has no attribute 'isopen'
AttributeError: 'Viewer' object has no attribute 'isopen'
self.close()
File "E:\anaconda\envs\tf2\lib\site-packages\gym\envs\classic_control\rendering.py", line 81, in close
AttributeError: 'Viewer' object has no attribute 'isopen'
AttributeError: 'Viewer' object has no attribute 'isopen'
if self.isopen and sys.meta_path:
File "E:\anaconda\envs\tf2\lib\site-packages\gym\envs\classic_control\rendering.py", line 81, in close
AttributeError: 'Viewer' object has no attribute 'isopen'
if self.isopen and sys.meta_path:
AttributeError: 'Viewer' object has no attribute 'isopen'

l don't konw how to fix it

..

Hi @marload,

Great repository you have here 😄! I am running your DQN script and I am trying to solve CartPole with it (consistently get >200 score).

I ran the script with the default parameters, but the agent is having trouble learning a successful policy. All I get is fluctuating scores between 10 and 100 for the first 800 episodes I trained it on. There was one episode with >200 but it was early in the training and having in mind that eps would have been very high at this point I think this must have been due to chance.

So my question is - if you have trained a successful agent with this algorithm can you provide me with "working" parameters? Or maybe DQN is just unstable in nature and I should run the script a couple of more times and hope for something better?

I have not reviewed the code thoroughly, because I wanted to see it working first, but at first glance, it looks clean and simple.

Anyway, thanks for posting it on Reddit, not sure why it was deleted. I hope I can learn a thing or two from it since I am working on something similar at the moment. 😄

Have a great day!

[No changes have been made to the code.
tensorflow version is 2.2, will this affect it?

The issue is not DQN specific, it's the only module (DQN_Discrete.py) which I tried to run on my mbp and on google colab. It runs okay, but both runs seem to take almost the same time. To activate the GPU, I added the following lines to main():

physical_devices = tf.config.experimental.list_physical_devices('GPU')
if len(physical_devices) > 0:
    tf.config.experimental.set_memory_growth(physical_devices[0], True)

Update:
wandb report shows 0% GPU utilization, you can check the graphs after a few minutes from starting the training here

Hello everyone.
I am trying to use A3C continuous. But I am getting some error saying "unrecognized arguments". Please see the attached picture.

How to solve this?