Interpolated class missing `bound_args_indices`

[FabiJa]

Hi,
a have a problem with the Interpolated class in the development branch. As a starting point I tried the example from the source:

import matplotlib.pyplot as plt
import numpy as np
import pymc as pm
import arviz as az
from scipy.stats import gamma
plt.style.use('arviz-darkgrid')
rv = gamma(1.99)
x = np.linspace(rv.ppf(0.01),rv.ppf(0.99), 1000)
points = np.linspace(x[0], x[-1], 50)
pdf = rv.pdf(points)
interpolated = pm.Interpolated.dist(points, pdf)
fig, ax = plt.subplots(1, 1)
ax.plot(x, rv.pdf(x), 'C0', linestyle = '--',  label='Original Gamma pdf',alpha=0.8,lw=2)
ax.plot(points, pdf, color='black', marker='o', label='Lattice Points',alpha=0.5,linestyle='')
ax.plot(x, np.exp(interpolated.logp(x).eval()),'C1',label='Interpolated pdf',alpha=0.8,lw=3)
r = interpolated.random(size=1000)
ax.hist(r, density=True, alpha=0.4,align ='mid',color='grey')
ax.legend(loc='best', frameon=False)
plt.show()

I get the following error:

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
/tmp/ipykernel_5967/2908023609.py in <module>
     13 ax.plot(x, rv.pdf(x), 'C0', linestyle = '--',  label='Original Gamma pdf',alpha=0.8,lw=2)
     14 ax.plot(points, pdf, color='black', marker='o', label='Lattice Points',alpha=0.5,linestyle='')
---> 15 ax.plot(x, np.exp(interpolated.logp(x).eval()),'C1',label='Interpolated pdf',alpha=0.8,lw=3)
     16 r = interpolated.random(size=1000)
     17 ax.hist(r, density=True, alpha=0.4,align ='mid',color='grey')

~/anaconda3/envs/pymc3_jax/lib/python3.9/site-packages/pymc/distributions/distribution.py in fn(*args, **kwargs)
    128 def _make_nice_attr_error(oldcode: str, newcode: str):
    129     def fn(*args, **kwargs):
--> 130         raise AttributeError(f"The `{oldcode}` method was removed. Instead use `{newcode}`.`")
    131 
    132     return fn

AttributeError: The `rv.logp(x)` method was removed. Instead use `pm.logp(rv, x)`.

Any advice?
Thx.

• PyMC3 Version: main

[ricardoV94]

In the dev branch you should write: np.exp(pm.logp(interpolated, x).eval())

Does that solve it?

[FabiJa]

Hi,
a quick test gives me the following error:

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
/tmp/ipykernel_7375/4243939004.py in <module>
     13 ax.plot(x, rv.pdf(x), 'C0', linestyle = '--',  label='Original Gamma pdf',alpha=0.8,lw=2)
     14 ax.plot(points, pdf, color='black', marker='o', label='Lattice Points',alpha=0.5,linestyle='')
---> 15 ax.plot(x, np.exp(pm.logp(interpolated, x).eval()),'C1',label='Interpolated pdf',alpha=0.8,lw=3)
     16 r = interpolated.random(size=1000)
     17 ax.hist(r, density=True, alpha=0.4,align ='mid',color='grey')

~/anaconda3/envs/pymc3_jax/lib/python3.9/site-packages/pymc/distributions/logprob.py in logp(var, rv_values, **kwargs)
    354         var.tag.value_var = at.as_tensor_variable(value_var, dtype=var.dtype)
    355 
--> 356     return logpt(var, rv_values, **kwargs)
    357 
    358 

~/anaconda3/envs/pymc3_jax/lib/python3.9/site-packages/pymc/distributions/logprob.py in logpt(var, rv_values, jacobian, scaling, transformed, cdf, sum, **kwargs)
    163         if rv_var is not None:
    164             # Make sure that the value is compatible with the random variable
--> 165             rv_value = rv_var.type.filter_variable(rv_value.astype(rv_var.dtype))
    166 
    167         if rv_value_var is None:

~/anaconda3/envs/pymc3_jax/lib/python3.9/site-packages/aesara/tensor/type.py in filter_variable(self, other, allow_convert)
    243                 return other2
    244 
--> 245         raise TypeError(
    246             f"Cannot convert Type {other.type} "
    247             f"(of Variable {other}) into Type {self}. "

TypeError: Cannot convert Type TensorType(float64, vector) (of Variable TensorConstant{[0.1463393...61971835]}) into Type TensorType(float64, scalar). You can try to manually convert TensorConstant{[0.1463393...61971835]} into a TensorType(float64, scalar).

[ricardoV94]

Ugh, the problem is that the Interpolated dist is a scalar and you are asking it to evaluate a vector of x.

We should simplify the process for users, but in the meantime you can do this when you define the interpolated dist:

interpolated = pm.Interpolated.dist(points, pdf, size=len(x)) 

[FabiJa]

That's it. Thanks.

[FabiJa]

Trying the example updating priors I get the following error:

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
/tmp/ipykernel_7478/954443530.py in <module>
      1 with pm.Model() as model_update:
      2     # Priors are posteriors from previous iteration
----> 3     R0 = from_posterior("R0", trace["R0"])
      4     lam = from_posterior("lambda", trace["lambda"])
      5     sigma = pm.HalfCauchy("sigma", 1, shape=2)

/tmp/ipykernel_7478/2112757308.py in from_posterior(param, samples)
     12     x = np.concatenate([[x[0] - 3 * width], x, [x[-1] + 3 * width]])
     13     y = np.concatenate([[0], y, [0]])
---> 14     return Interpolated(param, x, y)

~/anaconda3/envs/pymc3_jax/lib/python3.9/site-packages/pymc/distributions/continuous.py in __new__(cls, *args, **kwargs)
    163         transform = kwargs.get("transform", UNSET)
    164         if transform is UNSET:
--> 165             kwargs["transform"] = cls.default_transform()
    166         return super().__new__(cls, *args, **kwargs)
    167 

~/anaconda3/envs/pymc3_jax/lib/python3.9/site-packages/pymc/distributions/continuous.py in default_transform(cls)
    169     def default_transform(cls):
    170         if cls.bound_args_indices is None:
--> 171             raise ValueError(
    172                 f"Must specify bound_args_indices for {cls.__name__} bounded distribution"
    173             )

ValueError: Must specify bound_args_indices for Interpolated bounded distribution

Looking at the class Interpolated, I guess the definition of bound_args_indices = (lb ub) # Lower, Upper is missing, or?

[ricardoV94]

Sounds correct.

CC @kc611