BUG: infs or NaNs when performing logistic regression on the MNIST dataset.

[ylefay]

Describe the issue:

Thanks a lot for your work on this library.

I am trying to perform a logistic regression using ADVI implementation of PyMC on the MNIST digit dataset.
However, it fails because of NaN of inf values.
The CSV file containing the dataset is available at: https://drive.google.com/file/d/1eEKzfmEu6WKdRlohBQiqi3PhW_uIVJVP/view
(source: https://git-disl.github.io/GTDLBench/datasets/mnist_datasets/).

Reproduceable code example:

import numpy as np
import pandas as pd
import pymc as pm

# Performing preprocessing on the MNIST dataset: selecting the digits 0 and 8, normalizing the features and deleting the rows full of zeros.
def mnist_dataset():
    def preprocess(data):
        idx = (data[:, 0] == 0) + (data[:, 0] == 8)
        data = data[idx]
        idx = data[:, 0] == 8
        data[idx, 0] = 1
        data[~idx, 0] = 0
        labels = data[:, 0]
        data = data[:, 1:]
        data = data / 255
        data = np.delete(data, np.where(np.product(data == 0, axis=0) == 1), axis=-1)
        return data, labels.astype(dtype=float)

    mnist_train = np.array(pd.read_csv("./mnist/mnist_train.csv", header=None))
    mnist_train = preprocess(mnist_train)
    return mnist_train

flipped_predictors, response = mnist_dataset()
dim = flipped_predictors.shape[1]
with pm.Model() as logistic_model:
    cov = np.identity(dim) * 25
    beta = pm.MvNormal('beta', mu=np.zeros(dim), cov=cov)
    logit_theta = pm.Deterministic('logit_theta', flipped_predictors @ beta)
    y = pm.Bernoulli("y", logit_p=logit_theta, observed=response)
with logistic_model:
    callback = pm.variational.callbacks.CheckParametersConvergence(diff='absolute')
    start_means = {'beta': np.zeros(dim)}
    start_sigma = {'beta': np.ones(dim)}
    approx = pm.fit(n=100, callbacks=[callback], start=start_means,
                    start_sigma=start_sigma)

Error message:

<details>
/home/user/LSVI/venv/bin/python /home/user/LSVI/LSVI/experiments/logisticRegression/mnist/script.py 
Fitting: ━━━━━━━━━                             25% 0:00:03 Average Loss = 43,696
Traceback (most recent call last):
  File "/home/user/LSVI/venv/lib/python3.12/site-packages/pytensor/compile/function/types.py", line 959, in __call__
    self.vm()
  File "/home/user/LSVI/venv/lib/python3.12/site-packages/pytensor/graph/op.py", line 524, in rval
    r = p(n, [x[0] for x in i], o)
        ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/user/LSVI/venv/lib/python3.12/site-packages/pytensor/tensor/slinalg.py", line 298, in perform
    outputs[0][0] = scipy.linalg.solve_triangular(
                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/user/LSVI/venv/lib/python3.12/site-packages/scipy/linalg/_basic.py", line 334, in solve_triangular
    b1 = _asarray_validated(b, check_finite=check_finite)
         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/user/LSVI/venv/lib/python3.12/site-packages/scipy/_lib/_util.py", line 321, in _asarray_validated
    a = toarray(a)
        ^^^^^^^^^^
  File "/home/user/LSVI/venv/lib/python3.12/site-packages/numpy/lib/function_base.py", line 630, in asarray_chkfinite
    raise ValueError(
ValueError: array must not contain infs or NaNs

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "/home/user/LSVI/LSVI/experiments/logisticRegression/mnist/script.py", line 37, in <module>
    approx = pm.fit(n=100, callbacks=[callback], start=start_means,
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/user/LSVI/venv/lib/python3.12/site-packages/pymc/variational/inference.py", line 775, in fit
    return inference.fit(n, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/user/LSVI/venv/lib/python3.12/site-packages/pymc/variational/inference.py", line 155, in fit
    state, means, covs = self._iterate_with_loss(
                         ^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/user/LSVI/venv/lib/python3.12/site-packages/pymc/variational/inference.py", line 233, in _iterate_with_loss
    e = step_func()
        ^^^^^^^^^^^
  File "/home/user/LSVI/venv/lib/python3.12/site-packages/pytensor/compile/function/types.py", line 972, in __call__
    raise_with_op(
  File "/home/user/LSVI/venv/lib/python3.12/site-packages/pytensor/link/utils.py", line 524, in raise_with_op
    raise exc_value.with_traceback(exc_trace)
  File "/home/user/LSVI/venv/lib/python3.12/site-packages/pytensor/compile/function/types.py", line 959, in __call__
    self.vm()
  File "/home/user/LSVI/venv/lib/python3.12/site-packages/pytensor/graph/op.py", line 524, in rval
    r = p(n, [x[0] for x in i], o)
        ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/user/LSVI/venv/lib/python3.12/site-packages/pytensor/tensor/slinalg.py", line 298, in perform
    outputs[0][0] = scipy.linalg.solve_triangular(
                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/user/LSVI/venv/lib/python3.12/site-packages/scipy/linalg/_basic.py", line 334, in solve_triangular
    b1 = _asarray_validated(b, check_finite=check_finite)
         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/user/LSVI/venv/lib/python3.12/site-packages/scipy/_lib/_util.py", line 321, in _asarray_validated
    a = toarray(a)
        ^^^^^^^^^^
  File "/home/user/LSVI/venv/lib/python3.12/site-packages/numpy/lib/function_base.py", line 630, in asarray_chkfinite
    raise ValueError(
ValueError: array must not contain infs or NaNs
Apply node that caused the error: SolveTriangular{trans=0, unit_diagonal=False, lower=True, check_finite=True, b_ndim=2}([[5. 0. 0. ... 0. 0. 5.]], ExpandDims{axis=1}.0)
Toposort index: 31
Inputs types: [TensorType(float64, shape=(784, 784)), TensorType(float64, shape=(784, 1))]
Inputs shapes: [(784, 784), (784, 1)]
Inputs strides: [(8, 6272), (8, 8)]
Inputs values: ['not shown', 'not shown']
Outputs clients: [[Transpose{axes=[1, 0]}(SolveTriangular{trans=0, unit_diagonal=False, lower=True, check_finite=True, b_ndim=2}.0), SolveTriangular{trans=0, unit_diagonal=False, lower=False, check_finite=True, b_ndim=2}([[5. 0. 0. ... 0. 0. 5.]], SolveTriangular{trans=0, unit_diagonal=False, lower=True, check_finite=True, b_ndim=2}.0)]]

HINT: Re-running with most PyTensor optimizations disabled could provide a back-trace showing when this node was created. This can be done by setting the PyTensor flag 'optimizer=fast_compile'. If that does not work, PyTensor optimizations can be disabled with 'optimizer=None'.
HINT: Use the PyTensor flag `exception_verbosity=high` for a debug print-out and storage map footprint of this Apply node.
</details>

PyMC version information:

pymc==5.16.2, installed via pip, using python 3.12 on arch-linux

Context for the issue:

This is part of academic work on variational inference.

[welcome]

]
🎉 Welcome to PyMC! 🎉 We're really excited to have your input into the project! 💖

If you haven't done so already, please make sure you check out our Contributing Guidelines and Code of Conduct.

[N-0-MAD]

So you just want to get rid of NaN values from the Dataset right?

[ylefay]

So you just want to get rid of NaN values from the Dataset right?

There is no NaN values in the dataset.