Refactor/distribution #22
Conversation
|
Desiment
left a comment
Desiment
left a comment
There was a problem hiding this comment.
Add docstrings please
Also, about constraints - I missed how our discussion about distribution type and constrains resolved. From what I see nothing changed
| if np.any(vals <= 0.0): | ||
| return float("-inf") | ||
| return float(np.sum(np.log(vals))) | ||
| def log_likelihood(self, batch: Sample) -> float: ... |
There was a problem hiding this comment.
We decided to remove this method?
| s = 0.0 | ||
| for k in support.iter_leq(x): | ||
| s += float(pmf_func(float(k))) | ||
| return float(np.clip(s, 0.0, 1.0)) |
There was a problem hiding this comment.
For distributions with infinite support, this method will not work. For example, if X, Y ~ Poiss(1), are independent, then X - Y has infinite support (that is there infinite amount of points in support.iter_leq(x)).
| pass | ||
|
|
||
| # 3) Cursor-like API: first(), next(x) | ||
| if hasattr(support, "first") and hasattr(support, "next"): |
There was a problem hiding this comment.
Yeah, "first" may not exists. As well as the "last".
X ~ Poiss(1) - last not exisits, support is N
Y ~ -Poiss(1) - first not exists, support is -N
X - Y support is Z
| ---------- | ||
| distribution : Distribution | ||
| **options : Any | ||
| Unused (kept for a uniform API with continuous fitters). |
There was a problem hiding this comment.
But API for continuous fitters has an option for "left/right" ppf
| Semantics | ||
| --------- | ||
| ``CDF(x) = sup { q ∈ [0,1] : PPF(q) ≤ x }`` |
There was a problem hiding this comment.
Again, this depends on the definition of PPF, and how it is computed (as left or right). Note that this does not matter in contiuous case (or, at least, should not)
Done:
Need to be done: