Help with custom DataFit for Tweedie Regression

[fsaforo1]

I am trying to create a custom datafit for Tweedie GLM. I followed the tutorial for Poisson but it errors out with a not so useful error statement. See below the implementation and error.

from numba import jit
import numpy as np
from numpy.linalg import norm
from numba import njit
from numba import float64, int64, bool_

from skglm.datafits.base import BaseDatafit
from skglm.utils.sparse_ops import spectral_norm, _sparse_xj_dot

class Tweedie(BaseDatafit):
    r"""Tweedie datafit.

    The datafit reads:

    .. math::
        1 / n_"samples" \sum_(i=1)^(n_"samples") \left(\frac{y_i^{2-p}}{(1-p)(2-p)} - \frac{y_i e^{(1-p) Xw_i}}{1-p} + \frac{e^{(2-p) Xw_i}}{2-p}\right)
    """

    def __init__(self, p=1.5):
        self.p = p

    def get_spec(self):
        pass

    def params_to_dict(self):
        return dict(p=self.p)

    def initialize(self, X, y):
        if np.any(y < 0):
            raise ValueError(
                "Target vector `y` should only take non-negative values "
                "when fitting a Tweedie model.")

    def initialize_sparse(self, X_data, X_indptr, X_indices, y):
        if np.any(y < 0):
            raise ValueError(
                "Target vector `y` should only take non-negative values "
                "when fitting a Tweedie model.")

    @jit
    def raw_grad(self, y, Xw):
        """Compute gradient of datafit w.r.t ``Xw``."""
        p = self.p
        return ((np.exp((1 - p) * Xw) - y * np.exp((2 - p) * Xw)) / len(y))

    @jit
    def raw_hessian(self, y, Xw):
        """Compute Hessian of datafit w.r.t ``Xw``."""
        p = self.p
        return ((1 - p) * np.exp((1 - p) * Xw) + (2 - p) * y * np.exp((2 - p) * Xw)) / len(y)

    @jit
    def value(self, y, w, Xw):
        p = self.p
        term1 = y**(2-p) / ((1-p)*(2-p))
        term2 = y * np.exp((1-p) * Xw) / (1-p)
        term3 = np.exp((2-p) * Xw) / (2-p)
        return (np.sum(term1 - term2 + term3)) / len(y)

    @jit
    def gradient(self, X, y, Xw):
        return X.T @ self.raw_grad(y, Xw)

    @jit
    def gradient_scalar(self, X, y, w, Xw, j):
        return (X[:, j] @ self.raw_grad(y, Xw)) / len(y)

    @jit
    def full_grad_sparse(self, X_data, X_indptr, X_indices, y, Xw):
        n_features = X_indptr.shape[0] - 1
        grad = np.zeros(n_features, dtype=X_data.dtype)
        for j in range(n_features):
            grad[j] = 0.
            for i in range(X_indptr[j], X_indptr[j + 1]):
                grad[j] += X_data[i] * self.raw_grad(y, Xw)[X_indices[i]]
        return grad

    @jit
    def gradient_scalar_sparse(self, X_data, X_indptr, X_indices, y, Xw, j):
        grad = 0.
        for i in range(X_indptr[j], X_indptr[j + 1]):
            idx_i = X_indices[i]
            grad += X_data[i] * self.raw_grad(y, Xw)[idx_i]
        return grad / len(y)

    @jit
    def intercept_update_step(self, y, Xw):
        return np.sum(self.raw_grad(y, Xw))

from skglm.solvers import GramCD
from skglm.penalties import L1_plus_L2
from skglm.estimators import GeneralizedLinearEstimator as SKGLM

skmod = SKGLM(
    datafit=Tweedie(p=1.5),
    penalty=L1_plus_L2(alpha=1.0, l1_ratio=1.0),
    solver=GramCD(max_iter=1000, fit_intercept=True)
)

skmod.fit(X, y)

I get the error message below:

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
Cell In[63], line 7
      1 skmod = SKGLM(
      2     datafit=Tweedie(p=1.5),
      3     penalty=L1_plus_L2(alpha=1.0, l1_ratio=1.0),
      4     solver=GramCD(max_iter=1000, fit_intercept=True)
      5 )
----> 7 skmod.fit(X_trans, y)

File /blah/blah/miniconda/envs/ipykernel_python3.9/lib/python3.9/site-packages/skglm/estimators.py:252, in GeneralizedLinearEstimator.fit(self, X, y)
    249 self.datafit = self.datafit if self.datafit else Quadratic()
    250 self.solver = self.solver if self.solver else AndersonCD()
--> 252 return _glm_fit(X, y, self, self.datafit, self.penalty, self.solver)

File /blah/blah/miniconda/envs/ipykernel_python3.9/lib/python3.9/site-packages/skglm/estimators.py:104, in _glm_fit(X, y, model, datafit, penalty, solver)
    101 n_samples, n_features = X_.shape
    103 penalty_jit = compiled_clone(penalty)
--> 104 datafit_jit = compiled_clone(datafit, to_float32=X.dtype == np.float32)
    105 if issparse(X):
    106     datafit_jit.initialize_sparse(X_.data, X_.indptr, X_.indices, y)

File /blah/blah/miniconda/envs/ipykernel_python3.9/lib/python3.9/site-packages/skglm/utils/jit_compilation.py:79, in compiled_clone(instance, to_float32)
     63 def compiled_clone(instance, to_float32=False):
     64     """Compile instance to a jitclass.
     65
     66     Parameters
   (...)
     77         Return a jitclass.
     78     """
---> 79     return jit_cached_compile(
     80         instance.__class__,
     81         instance.get_spec(),
     82         to_float32,
     83     )(**instance.params_to_dict())

File /blah/blah/miniconda/envs/ipykernel_python3.9/lib/python3.9/site-packages/skglm/utils/jit_compilation.py:60, in jit_cached_compile(klass, spec, to_float32)
     57 if to_float32:
     58     spec = spec_to_float32(spec)
---> 60 return jitclass(spec)(klass)

File /blah/blah/miniconda/envs/ipykernel_python3.9/lib/python3.9/site-packages/numba/experimental/jitclass/decorators.py:77, in jitclass.<locals>.wrap(cls)
     74 else:
     75     from numba.experimental.jitclass.base import (register_class_type,
     76                                                   ClassBuilder)
---> 77     cls_jitted = register_class_type(cls, spec, types.ClassType,
     78                                      ClassBuilder)
     80     # Preserve the module name of the original class
     81     cls_jitted.__module__ = cls.__module__

File /blah/blah/miniconda/envs/ipykernel_python3.9/lib/python3.9/site-packages/numba/experimental/jitclass/base.py:213, in register_class_type(cls, spec, class_ctor, builder)
    211     msg = "class members are not yet supported: {0}"
    212     members = ', '.join(others.keys())
--> 213     raise TypeError(msg.format(members))
    215 for k, v in props.items():
    216     if v.fdel is not None:

TypeError: class members are not yet supported: value, raw_grad, raw_hessian, gradient, gradient_scalar, full_grad_sparse, gradient_scalar_sparse, intercept_update_step

[Badr-MOUFAD]

Hi @fsaforo1 👋

You should remove the jit decorator that you used in most methods of the Tweedie class : the jit compilation will be handled by GeneralizedLinearEstimator

[Badr-MOUFAD]

I highly recommend you open a Pull Request. That way we can follow along with you a give you a hand if needed and so that the community can benefit for the Tweedie implementation.

[mathurinm]

@fsaforo1 would you be able to send a Pull request to start implementing tweedie datafit ? If this is your first PR, a small contribution guide is at https://github.com/mathurinm/github-assignment/?tab=readme-ov-file#summary-how-to-contribute-to-an-existing-repository

[fsaforo1]

I managed to get it to fit but the estimates are way off. It's either the math is wrong or something is amiss. I will post a reproducible example when possible. I will make the pull request once everything checks out

[mathurinm]

It's better if you start the PR right now, even if not working, this way we can all look at the same code and help you

[QB3]

Let us know if you need help @fsaforo1 !

[mathurinm]

@fsaforo1 can you show your code, either in your repo either in a pull request, so that we work on it together and integrate it in skglm? Thanks a lot!