Difference between np.dot and np.multiply with np.sum in binary cross-entropy loss calculation
np.dot
is the dot product of two matrices.
|A B| . |E F| = |A*E+B*G A*F+B*H|
|C D| |G H| |C*E+D*G C*F+D*H|
Whereas np.multiply
does an element-wise multiplication of two matrices.
|A B| ⊙ |E F| = |A*E B*F|
|C D| |G H| |C*G D*H|
When used with np.sum
, the result being equal is merely a coincidence.
>>> np.dot([[1,2], [3,4]], [[1,2], [2,3]])
array([[ 5, 8],
[11, 18]])
>>> np.multiply([[1,2], [3,4]], [[1,2], [2,3]])
array([[ 1, 4],
[ 6, 12]])
>>> np.sum(np.dot([[1,2], [3,4]], [[1,2], [2,3]]))
42
>>> np.sum(np.multiply([[1,2], [3,4]], [[1,2], [2,3]]))
23
What you're doing is calculating the binary cross-entropy loss which measures how bad the predictions (here: A2
) of the model are when compared to the true outputs (here: Y
).
Here is a reproducible example for your case, which should explain why you get a scalar in the second case using np.sum
In [88]: Y = np.array([[1, 0, 1, 1, 0, 1, 0, 0]])
In [89]: A2 = np.array([[0.8, 0.2, 0.95, 0.92, 0.01, 0.93, 0.1, 0.02]])
In [90]: logprobs = np.dot(Y, (np.log(A2)).T) + np.dot((1.0-Y),(np.log(1 - A2)).T)
# `np.dot` returns 2D array since its arguments are 2D arrays
In [91]: logprobs
Out[91]: array([[-0.78914626]])
In [92]: cost = (-1/m) * logprobs
In [93]: cost
Out[93]: array([[ 0.09864328]])
In [94]: logprobs = np.sum(np.multiply(np.log(A2), Y) + np.multiply((1 - Y), np.log(1 - A2)))
# np.sum returns scalar since it sums everything in the 2D array
In [95]: logprobs
Out[95]: -0.78914625761870361
Note that the np.dot
sums along only the inner dimensions which match here (1x8) and (8x1)
. So, the 8
s will be gone during the dot product or matrix multiplication yielding the result as (1x1)
which is just a scalar but returned as 2D array of shape (1,1)
.
Also, most importantly note that here np.dot
is exactly same as doing np.matmul
since the inputs are 2D arrays (i.e. matrices)
In [107]: logprobs = np.matmul(Y, (np.log(A2)).T) + np.matmul((1.0-Y),(np.log(1 - A2)).T)
In [108]: logprobs
Out[108]: array([[-0.78914626]])
In [109]: logprobs.shape
Out[109]: (1, 1)
Return result as a scalar value
np.dot
or np.matmul
returns whatever the resulting array shape would be, based on input arrays. Even with out=
argument it's not possible to return a scalar, if the inputs are 2D arrays. However, we can use np.asscalar()
on the result to convert it to a scalar if the result array is of shape (1,1)
(or more generally a scalar value wrapped in an nD array)
In [123]: np.asscalar(logprobs)
Out[123]: -0.7891462576187036
In [124]: type(np.asscalar(logprobs))
Out[124]: float
ndarray of size 1 to scalar value
In [127]: np.asscalar(np.array([[[23.2]]]))
Out[127]: 23.2
In [128]: np.asscalar(np.array([[[[23.2]]]]))
Out[128]: 23.2
If Y
and A2
are (1,N) arrays, then np.dot(Y,A.T)
will produce a (1,1) result. It is doing a matrix multiplication of a (1,N) with a (N,1). The N's
are summed, leaving the (1,1).
With multiply
the result is (1,N). Sum all values, and the result is a scalar.
If Y
and A2
were (N,) shaped (same number of elements, but 1d), the np.dot(Y,A2)
(no .T
) would also produce a scalar. From np.dot
documentation:
For 2-D arrays it is equivalent to matrix multiplication, and for 1-D arrays to inner product of vectors
Returns the dot product of a and b. If a and b are both scalars or both 1-D arrays then a scalar is returned; otherwise an array is returned.
squeeze
reduces all size 1 dimensions, but still returns an array. In numpy
an array can have any number of dimensions (from 0 to 32). So a 0d array is possible. Compare the shape of np.array(3)
, np.array([3])
and np.array([[3]])
.