hongshaorou/Tianshou
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

SamplingConfig: Improve docstrings of replay_buffer_save_only_last_obs, replay_buffer_stack_num

Browse Source
This commit is contained in:
Dominik Jain 2024-04-29 17:12:28 +02:00
parent 74737416ff
commit 250a129cc4
1 changed files with 9 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
tianshou/highlevel/config.py
View File

@ -118,9 +118,12 @@ class SamplingConfig(ToStringMixin):
replay_buffer_ignore_obs_next: bool = False replay_buffer_ignore_obs_next: bool = False
replay_buffer_save_only_last_obs: bool = False replay_buffer_save_only_last_obs: bool = False
"""if True, only the most recent frame is saved when appending to experiences rather than the """if True, for the case where the environment outputs stacked frames (e.g. because it
full stacked frames. This avoids duplicating observations in buffer memory. Set to False to is using a `FrameStack` wrapper), save only the most recent frame so as not to duplicate
save stacked frames in full. observations in buffer memory. Specifically, if the environment outputs observations `obs` with
shape (N, ...), only obs[-1] of shape (...) will be stored.
Frame stacking with a fixed number of frames can then be recreated at the buffer level by setting
:attr:`replay_buffer_stack_num`.
""" """
replay_buffer_stack_num: int = 1 replay_buffer_stack_num: int = 1
@ -128,6 +131,9 @@ class SamplingConfig(ToStringMixin):
the number of consecutive environment observations to stack and use as the observation input the number of consecutive environment observations to stack and use as the observation input
to the agent for each time step. Setting this to a value greater than 1 can help agents learn to the agent for each time step. Setting this to a value greater than 1 can help agents learn
temporal aspects (e.g. velocities of moving objects for which only positions are observed). temporal aspects (e.g. velocities of moving objects for which only positions are observed).
If the environment already stacks frames (e.g. using a `FrameStack` wrapper), this should either not
be used or should be used in conjunction with :attr:`replay_buffer_save_only_last_obs`.
""" """
@property @property