Is it possible to get uint8 observations for Atari?

[AdityaGudimella]

Hello,
One common trick to reduce replay buffer memory footprint when storing observations is to multiply each observation by 255 and storing it as a numpy.uint8 dtype in the replay buffer. Is it possible to get the uint8 states directly from envpool?

[Trinkle23897]

It is uint8 already.

In [1]: import envpool                        
                                                                      
In [2]: e=envpool.make_gym("Pong-v5")         
                                                                      
In [3]: e.observation_space                                    
Out[3]: Box(0, 255, (4, 84, 84), uint8)
                                                                      
In [4]: e.reset()                   
Out[4]:                                          
array([[[[ 52,  52,  52, ...,  87,  87,  87],
         [ 87,  87,  87, ...,  87,  87,  87],
         [ 87,  87,  87, ...,  87,  87,  87],
         ...,
         [236, 236, 236, ..., 236, 236, 236],
         [236, 236, 236, ..., 236, 236, 236],
         [236, 236, 236, ..., 236, 236, 236]],
                                                                      
        [[ 52,  52,  52, ...,  87,  87,  87],
         [ 87,  87,  87, ...,  87,  87,  87],
         [ 87,  87,  87, ...,  87,  87,  87],
         ...,
         [236, 236, 236, ..., 236, 236, 236],
         [236, 236, 236, ..., 236, 236, 236],
         [236, 236, 236, ..., 236, 236, 236]],
                                                                      
        [[ 52,  52,  52, ...,  87,  87,  87],
         [ 87,  87,  87, ...,  87,  87,  87],
         [ 87,  87,  87, ...,  87,  87,  87],
         ...,
         [236, 236, 236, ..., 236, 236, 236],
         [236, 236, 236, ..., 236, 236, 236],
         [236, 236, 236, ..., 236, 236, 236]],
                                                                      
        [[ 52,  52,  52, ...,  87,  87,  87],
         [ 87,  87,  87, ...,  87,  87,  87],
         [ 87,  87,  87, ...,  87,  87,  87],
         ...,
         [236, 236, 236, ..., 236, 236, 236],
         [236, 236, 236, ..., 236, 236, 236],
         [236, 236, 236, ..., 236, 236, 236]]]], dtype=uint8)

In [5]: import numpy as np

In [6]: e.step(np.array([0]))
Out[6]: 
(array([[[[ 52,  52,  52, ...,  87,  87,  87],
          [ 87,  87,  87, ...,  87,  87,  87],
          [ 87,  87,  87, ...,  87,  87,  87],
          ...,
          [236, 236, 236, ..., 236, 236, 236],
          [236, 236, 236, ..., 236, 236, 236],
          [236, 236, 236, ..., 236, 236, 236]],
  
         [[ 52,  52,  52, ...,  87,  87,  87],
          [ 87,  87,  87, ...,  87,  87,  87],
          [ 87,  87,  87, ...,  87,  87,  87],
          ...,
          [236, 236, 236, ..., 236, 236, 236],
          [236, 236, 236, ..., 236, 236, 236],
          [236, 236, 236, ..., 236, 236, 236]],
  
         [[ 52,  52,  52, ...,  87,  87,  87],
          [ 87,  87,  87, ...,  87,  87,  87],
          [ 87,  87,  87, ...,  87,  87,  87],
          ...,
          [236, 236, 236, ..., 236, 236, 236],
          [236, 236, 236, ..., 236, 236, 236],
          [236, 236, 236, ..., 236, 236, 236]],
  
         [[ 52,  52,  52, ...,  87,  87,  87],
          [ 87,  87,  87, ...,  87,  87,  87],
          [ 87,  87,  87, ...,  87,  87,  87],
          ...,
          [236, 236, 236, ..., 236, 236, 236],
          [236, 236, 236, ..., 236, 236, 236],
          [236, 236, 236, ..., 236, 236, 236]]]], dtype=uint8),
 array([0.], dtype=float32),
 array([False]),
 {'env_id': array([0], dtype=int32),
  'players': {'env_id': array([0], dtype=int32)},
  'lives': array([0], dtype=int32),
  'reward': array([0.], dtype=float32),
  'TimeLimit.truncated': array([False])})

[AdityaGudimella]

You're right. I'm so sorry. Thank you for the amazing library. I'm using it in my project (will be public in a few months) too.

[Trinkle23897]

Glad to hear that! Best wishes for your new project!