[BugFix] Vectorized priority update in replay buffers #1598
Conversation
Signed-off-by: Matteo Bettini <[email protected]>
facebook-github-bot
added
the
CLA Signed
vmoens
added
bug
| priority = torch.tensor( | ||
| [self._get_priority_item(td) for td in data], | ||
| dtype=torch.float, | ||
| device=data.device, | ||
| ) |
There was a problem hiding this comment.
I guess we assume that the stack dim is 0 but it could not be (?)
I think we can consider the priority can be stacked, no? At the end of the day it's supposed to be one priority per item. Maybe I'm missing something
There was a problem hiding this comment.
this was the previous treatment so I am not super sure what was going on or why things were this way.
my guess is that it is assuming 0 as stack dim because in expand it stacks on 0 and also because that is the dim of the indeces and priority
are you suggesting to completely remove this and always go vectorized? I am down to try.
There was a problem hiding this comment.
maybe it was this way because the priority can have different shapes along the stack dim?
That is the only explanation I can guess
Co-authored-by: Vincent Moens <[email protected]>
Co-authored-by: Vincent Moens <[email protected]>
Signed-off-by: Matteo Bettini <[email protected]>
…rioritised_buffer
Signed-off-by: Matteo Bettini <[email protected]>
Signed-off-by: Matteo Bettini <[email protected]> Co-authored-by: Vincent Moens <[email protected]>
This is a patch for #1574
It fixes the core problem highlighted in that issue but there are still points that will need attention in a future refactoring of this class as I am not sure it is compatible with all the cases it aims to support. I am happy to elucidate more about this if needed.