setting an array element with a sequence

[LeiYangGH]

PS C:\G\ML-From-Scratch> python mlfromscratch/examples/polynomial_regression.py
Finding regularization constant using cross validation:
Traceback (most recent call last):
File "C:\G\ML-From-Scratch\mlfromscratch\examples\polynomial_regression.py", line 81, in
main()
File "C:\G\ML-From-Scratch\mlfromscratch\examples\polynomial_regression.py", line 32, in main
cross_validation_sets = k_fold_cross_validation_sets(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\user.conda\envs\tor\Lib\site-packages\mlfromscratch-0.0.4-py3.12.egg\mlfromscratch\utils\data_manipulation.py", line 145, in k_fold_cross_validation_sets
ValueError: setting an array element with a sequence. The requested array has an inhomogeneous shape after 2 dimensions. The detected shape was (10, 4) + inhomogeneous part.

[levietnguyen]

Simply Change return np.array(sets) in data_manipulation.py to return sets

Explanation:

Why Returning sets Works Instead of np.array(sets)

The fix works because of how NumPy handles arrays with complex nested structures. Here's why:

The Original Problem

When the code attempted to return np.array(sets), it was trying to convert a list of lists (where each inner list contains arrays of potentially different shapes) into a homogeneous NumPy array.

Why This Fails

NumPy arrays are designed to hold elements of the same type and, more importantly, elements that can be arranged in a regular grid. When you have a list like sets where each element contains:

• X_train: A 2D array with shape (varying_size, n_features)
• X_test: A 2D array with shape (smaller_size, n_features)
• y_train: A 1D array with shape (varying_size,)
• y_test: A 1D array with shape (smaller_size,)

NumPy can't create a regular grid structure to hold these heterogeneous shapes.

What Happens When np.array(sets) is Used

When NumPy tries to convert this irregular structure to an array, it attempts to create what's called a "ragged array"—an array where the size of dimensions can vary.

• NumPy does not directly support ragged arrays, so it tries to use an object array (i.e., an array of Python objects).
• This approach fails when the objects themselves have complex, nested structures with inconsistent dimensions.

Why Returning sets Works

By returning sets as a plain Python list instead of converting it to a NumPy array:

✅ You preserve the original structure without forcing it into a rigid NumPy array format.
✅ The consumer code can iterate through the list and access each fold's data directly.
✅ Each element retains its original NumPy array properties for X_train, X_test, y_train, and y_test.

Final Fix:

Replace:

return np.array(sets)

With:

return sets

This ensures that the k-fold cross-validation sets remain correctly formatted and usable.