98 lines
3.2 KiB
Python
98 lines
3.2 KiB
Python
from typing import Any, Dict
|
|
import ast
|
|
import torch
|
|
from torch.distributions import Bernoulli
|
|
import scipy.stats as scst
|
|
|
|
|
|
class BernoulliEx(Bernoulli):
|
|
"""
|
|
A class to extend Bernoulli to add missing functionality.
|
|
|
|
pragma: skip_doc
|
|
"""
|
|
def cdf(self, x: torch.Tensor):
|
|
"""
|
|
Computes the cumulative distribution function (CDF) of the Bernoulli distribution.
|
|
|
|
Args:
|
|
x (torch.Tensor): Input tensor containing values where the CDF is evaluated.
|
|
|
|
Returns:
|
|
torch.Tensor: Tensor containing the cumulative probabilities.
|
|
"""
|
|
a = x.detach().cpu().numpy()
|
|
probs = self.probs.detach().cpu().numpy()
|
|
outputs = scst.bernoulli.cdf(a, probs)
|
|
return torch.tensor(outputs, dtype=torch.float32).to(x.device)
|
|
|
|
def icdf(self, q: torch.Tensor):
|
|
"""
|
|
Computes the inverse of cumulative distribution function (CDF) of the Bernoulli distribution.
|
|
|
|
Args:
|
|
q (torch.Tensor): Input tensor containing q.
|
|
|
|
Returns:
|
|
torch.Tensor: Tensor containing the inverse of cumulative probabilities.
|
|
"""
|
|
a = q.detach().cpu().numpy()
|
|
probs = self.probs.detach().cpu().numpy()
|
|
outputs = scst.bernoulli.ppf(a, probs)
|
|
return torch.tensor(outputs, dtype=torch.float32).to(q.device)
|
|
|
|
|
|
class PtdBernoulli:
|
|
"""
|
|
Instantiates a Bernoulli distribution object.
|
|
|
|
Args:
|
|
probs (str): Probabilities of the distsribution. You can specify a scalar or a tuple of float.
|
|
logits (str): Logits of the distribution. You can specify a scalar or a tuple of float.
|
|
|
|
category: PyTorch wrapper - Distribution
|
|
"""
|
|
|
|
@classmethod
|
|
def INPUT_TYPES(cls) -> Dict[str, Any]:
|
|
"""
|
|
Defines the input types for the function.
|
|
|
|
Returns:
|
|
Dict[str, Any]: A dictionary specifying the required input types.
|
|
"""
|
|
return {
|
|
"required": {
|
|
"probs": ("STRING", {"default": ""}),
|
|
"logits": ("STRING", {"default": ""})
|
|
}
|
|
}
|
|
|
|
RETURN_TYPES: tuple = ("PTDISTRIBUTION",)
|
|
FUNCTION: str = "f"
|
|
CATEGORY: str = "Distribution"
|
|
|
|
def f(self, probs: str, logits: str) -> tuple:
|
|
"""
|
|
Instantiates a Bernoulli distribution object.
|
|
|
|
Args:
|
|
probs (str): Probabilities of the distsribution. You can specify a scalar or a tuple of float.
|
|
logits (str): Logits of the distribution. You can specify a scalar or a tuple of float.
|
|
Returns:
|
|
tuple: (torch.distributions.bernoulli.Bernoulli)
|
|
"""
|
|
p = ast.literal_eval(probs) if probs else None
|
|
l = ast.literal_eval(logits) if logits else None
|
|
|
|
if p is not None and l is not None: # You cannot use if p because p or l can be "0".
|
|
raise ValueError("You can specify either probabilities or logits, but not both.")
|
|
if p is None and l is None:
|
|
raise ValueError("You have to specify either probabilities or logits")
|
|
|
|
if p is not None:
|
|
dist = BernoulliEx(probs=torch.tensor(p, dtype=torch.float32))
|
|
elif l is not None:
|
|
dist = BernoulliEx(logits=torch.tensor(l, dtype=torch.float32))
|
|
|
|
return (dist,) |