ValueError: operands could not be broadcast together with shapes

The error#

ValueError: operands could not be broadcast together with shapes (1000,3) (1000,)

PyTorch reports the same problem with different words:

RuntimeError: The size of tensor a (3) must match the size of tensor b (1000) at non-singleton dimension 1

What it means#

You combined two arrays element by element — with +, -, * or a comparison. Their shapes could not be matched up.

Broadcasting is NumPy's rule for stretching a smaller array across a bigger one. It compares shapes from the right. Two dimensions are compatible when they are equal, or when one of them is 1. A missing dimension counts as 1.

So (1000, 3) and (3,) work: the 3s line up, and the missing axis stretches. But (1000, 3) and (1000,) fail: the rightmost pair is 3 versus 1000.

Why it happens#

The most common route is a reduction that dropped an axis. X.mean(axis=1) on a (1000, 3) array returns shape (1000,). Subtracting that from X compares 3 with 1000 and fails.

The second route is mixing row-shaped and column-shaped data. A flat array (N,) and a column (N, 1) behave differently. Combining (N,) with (N, 1) does not error — it broadcasts to (N, N). That is usually a silent bug that shows up later as this error, or as a memory explosion.

How to fix it#

1. Print both shapes before the failing line. The message names the shapes, but you need to know which variable is which.

print(a.shape, b.shape)

2. Keep the reduced axis with keepdims=True. This is the fix for the mean/sum/max family.

X_centered = X - X.mean(axis=1, keepdims=True)   # (1000,3) - (1000,1) works

3. Add the axis yourself with reshape or np.newaxis.

b = b.reshape(-1, 1)        # (1000,) -> (1000,1)
b = b[:, np.newaxis]        # same thing

4. Flatten a stray column back to 1D when you meant a flat array.

y = y.ravel()               # (1000,1) -> (1000,)

scikit-learn wants targets flat, and this silences its column-vector warning too.

5. In PyTorch, use unsqueeze for the same job.

b = b.unsqueeze(1)          # (1000,) -> (1000,1)

How to prevent it#

Write the expected shape as a comment next to any line that changes it. Add cheap asserts at function boundaries, like assert X.ndim == 2. Be suspicious the moment (N,) and (N, 1) both exist in the same function — pick one convention and convert at the edges.

Related errors#

• mat1 and mat2 shapes cannot be multiplied — the matrix-multiplication version of a shape mismatch
• ValueError: cannot reshape array of size
• ValueError: setting an array element with a sequence
• stack expects each tensor to be equal size