[Python] Why does np.unique return fewer/more unique rows than MATLAB’s unique with identical...

I’m trying to replicate MATLAB behavior using NumPy in Python.

I have a 2D matrix of shape (248, 2) in both MATLAB and Python (via NumPy). When I use the unique function to filter out duplicate rows:

• MATLAB’s unique(A, 'rows') returns shape (240, 2)
• NumPy’s np.unique(A, axis=0) returns shape (232, 2)

Why is there a difference in the number of unique rows, even though the input matrices are the same?

I suspect it has to do with floating-point precision. Does MATLAB use a tolerance internally for comparing rows in unique? And if so, how can I replicate MATLAB’s behavior accurately in NumPy?

Code I’m using:

# Example: MATLAB
a = unique(a, ‘rows’)


# Trying to mimic MATLAB behavior
unique_rows = np.unique(a, axis=0)
print("Shape after unique:", a.shape)`

Continue reading...