How can I resolve the 'RuntimeError: Numpy is not available' error in scaler.inverse_transform?

Below is one explanation and a couple of workarounds to fix the error.

What’s Happening

When you call

PYTHONCopied
original_scores = scaler.inverse_transform(features[i].numpy().reshape(1, -1)).flatten()

you are taking a single sample (a row) from your already standardized data and trying to invert the scaling. Although reshaping with (1, -1) generally produces a 2D array, sometimes issues occur if the shape of the sample isn’t consistent with what StandardScaler was originally fitted on.

For example, if your original data had N features per sample (with shape (n_samples, N)), the scaler is expecting a 2D array with exactly N columns. In some cases—even if it looks right—the manual reshape of a tensor sample can cause a mismatch. (Occasionally errors like "Found array with X features, but expected Y" appear if there is any mismatch in dimensionality.)

How to Fix It

A common workaround is to apply the inverse transformation in one call on the entire (scaled) dataset rather than doing it sample by sample. This leaves less room for shape errors. For example:

PYTHONCopied
# Instead of doing scaler.inverse_transform for each sample:
original_features = scaler.inverse_transform(features.numpy())

for i in range(12):
    original_scores = original_features[i]  # This is now the original feature vector for sample i.
    if y_pred[i] == 1:
        print('Học sinh đạt danh hiệu với điểm:', original_scores)
    else:
        print('Học sinh không đạt danh hiệu với điểm:', original_scores)

If for some reason you must work sample by sample, ensure that the array you pass has the exact same number of columns as the scaler was fitted with. For example, double-check that:

1. The scaler is fitted on a 2D array with shape (n_samples, n_features).
2. Each sample being inverse transformed is reshaped into (1, n_features).

Given that you already use features[i].numpy().reshape(1, -1), this should work provided that each sample has all the features. If you still see an error then printing the shape before inverse transforming might help:

PYTHONCopied
sample_data = features[i].numpy().reshape(1, -1)
print("Sample shape:", sample_data.shape)  # Should be (1, n_features)
original_scores = scaler.inverse_transform(sample_data).flatten()

Summary

• The error likely arises from a mismatch in the expected shape of data for scaling.
• A robust solution is to inverse transform the whole dataset at once.
• Always verify that the shape of the data passed to scaler.inverse_transform exactly matches the shape used during fit (i.e. (1, n_features) for one sample).

This approach should resolve the error you’re encountering with Python 3.12.7.