"expected zero arguments for construction of ClassDict (for numpy.dtype)" when calling UDF that returns FloatType()

Question

I believe it is related to this one: Spark Error:expected zero arguments for construction of ClassDict (for numpy.core.multiarray._reconstruct)

I have a dataframe

id col_1 col_2
1 [1,2] [1,3]
2 [2,1] [3,4]

I want to create another column that is a cosine distance between col_1 and col_2.

from scipy.spatial.distance import cosine

def cosine_distance(a,b):
    try:
        return cosine(a, b)
    except Exception as e:
        return 0.0 # in case division by zero

And I defined a udf:

cosine_distance_udf = udf (cosine_distance, FloatType())

And finally:

new_df = df.withColumn('cosine_distance', cosine_distance_udf('col_1', 'col_2'))

And I have the error: PickleException: expected zero arguments for construction of ClassDict (for numpy.dtype)

What did I do wrong?

Answer 1

The reason for the error is clear when you check the return type of cosine:

type(cosine([1, 2], [1, 3]))
# numpy.float64

However, np.float64 is a subclass of float:

issubclass(np.float64, float)
# True

So, with a small change, to your function,

def cosine_distance(a, b):
    try:
        return float(cosine(a, b)) # cosine(a, b).item()
    except Exception as e:
        return 0.0 # in case division by zero

This will work

df.withColumn('cosine_distance', cosine_distance_udf('col_1', 'col_2')).show()

+------+------+---------------+
| col_1| col_2|cosine_distance|
+------+------+---------------+
|[1, 2]|[3, 4]|     0.01613009|
|[2, 1]|[3, 4]|     0.10557281|
+------+------+---------------+