How can I sample continuous transitions

[heyuanYao-pku]

I'm writing a custom training process, which is autoregressive . And it needs to sample $n$ continuous transitions from the replay buffer so that it can be used as the ground truth of roll out.
$n$ is perhaps 16 or 24 and so on.....

[heyuanYao-pku]

Or perhaps, is there a convenient way to sample from buffer with a custom function?

[ymd-h]

@heyuanYao-pku
Hi, thank you for your feed backs.

I don't find any convenient ways, but there are workarounds.

---

Option 1:
You can transform to $n$-length batch shape before adding.
Additional local buffer can be used, however, if you want to utilize stack_compress option, stack dimension must be the last dimension.
https://ymd_h.gitlab.io/cpprb/features/memory_compression/

---

Option 2:
You can use private method _encode_sample, which takes index.
You might generate random number outside, or slide PER indices after sampling.

Note:
_encode_sample is private method and not guaranteed within API compatibility, however, I know some users use the method, so that we don't change without enough reason.

---

FYI:
(I am not sure about the detail of your needs, so it might be nonsense for you.)

For online periodical roll out, some users utilize ReplayBuffer.
When buffer becomes full, dump with get_all_transitions() method, then clear() it.