Introduction

Most "bad input shape" errors in scikit-learn SVM code come from one rule: X must be a two-dimensional array of shape (n_samples, n_features), while y must usually be a one-dimensional target vector. If either side has the wrong shape, SVC.fit raises a ValueError.

The confusing part is that NumPy, pandas, and slicing operations can quietly change dimensionality. A column can turn into a one-dimensional series, or a target can become a two-dimensional frame, and the SVM then rejects it.

What Shape the SVM Expects

For a normal classification task:

• 'X is a matrix'
• each row is one sample
• each column is one feature
• 'y is a vector with one label per sample'

This is correct:

1import numpy as np
2from sklearn.svm import SVC
3
4X = np.array([
5    [1.0, 2.0],
6    [1.5, 1.8],
7    [5.0, 8.0],
8    [6.0, 9.0],
9])
10y = np.array([0, 0, 1, 1])
11
12model = SVC(kernel="linear")
13model.fit(X, y)

Here X.shape is (4, 2) and y.shape is (4,), which is exactly what the estimator wants.

The Most Common Shape Mistake

The most frequent bug appears when you have a single feature and pass a flat array instead of a two-dimensional matrix.

This fails:

1import numpy as np
2from sklearn.svm import SVC
3
4X = np.array([1.0, 2.0, 3.0, 4.0])
5y = np.array([0, 0, 1, 1])
6
7model = SVC()
8model.fit(X, y)

Typical error text looks like this:

Expected 2D array, got 1D array instead

The fix is to reshape the input so each value becomes a row with one feature:

1X = X.reshape(-1, 1)
2
3model = SVC()
4model.fit(X, y)

The -1 tells NumPy to infer the number of rows, and 1 means one feature column.

pandas-Specific Issues

With pandas, the difference between df["feature"] and df[["feature"]] matters.

1import pandas as pd
2from sklearn.svm import SVC
3
4df = pd.DataFrame({
5    "height": [150, 160, 170, 180],
6    "weight": [50, 60, 70, 85],
7    "label": [0, 0, 1, 1],
8})
9
10X_bad = df["height"]
11y = df["label"]
12
13print(X_bad.shape)  # (4,)

df["height"] is a series, so it is one-dimensional. For an SVM feature matrix, use:

1X = df[["height"]]
2y = df["label"]
3
4model = SVC()
5model.fit(X, y)

df[["height"]] keeps the two-dimensional table shape.

The same idea applies when selecting multiple columns. Make sure your column list stays inside double brackets so you still get a frame instead of a series.

Mismatched Training and Prediction Shapes

You can also get shape errors during prediction. If the model was trained on two features, a future call to predict must also provide two features per sample.

1import numpy as np
2from sklearn.svm import SVC
3
4X = np.array([
5    [1.0, 2.0],
6    [2.0, 3.0],
7    [8.0, 8.0],
8    [9.0, 10.0],
9])
10y = np.array([0, 0, 1, 1])
11
12model = SVC()
13model.fit(X, y)
14
15sample = np.array([[4.0, 5.0]])
16print(model.predict(sample))

This is correct because sample.shape is (1, 2): one sample, two features.

This is wrong:

sample = np.array([4.0, 5.0])
model.predict(sample)

That is a one-dimensional array. Wrap it in another list or reshape it:

sample = sample.reshape(1, -1)

Practical Debugging Checklist

Before calling fit or predict, print shapes directly:

print("X:", X.shape)
print("y:", y.shape)

For classification with scikit-learn SVMs, a good default checklist is:

• 'X.ndim == 2'
• 'y.ndim == 1'
• 'len(X) == len(y)'
• training and prediction inputs use the same number of feature columns

If you use preprocessing pipelines, check the output of the transformer too. A custom preprocessing step can flatten arrays accidentally.

Common Pitfalls

The biggest pitfall is assuming a single feature can be passed as a simple list. In scikit-learn, one feature still needs a two-dimensional matrix with one column.

Another common mistake is passing y as a data frame with shape (n, 1) instead of a series or flat array. Some estimators handle that quietly, but many workflows behave better when the target is explicitly one-dimensional.

Slicing is another source of bugs. In pandas, df["col"] and df[["col"]] are not interchangeable. One gives a series, the other gives a frame.

Finally, do not guess. Print X.shape and y.shape before training. Most shape bugs become obvious immediately once you inspect the arrays.

Summary

• 'X for an SVM must usually have shape (n_samples, n_features).'
• 'y should usually be a one-dimensional label array.'
• A single feature still needs a two-dimensional input such as reshape(-1, 1).
• In pandas, use df[["col"]] for feature matrices and df["label"] for targets.
• Training and prediction must use the same number of feature columns.
• Printing shapes is the fastest way to diagnose these errors.