I'm a bit confused as why the following two slicing and index scenarios yield different shapes. Especially when x[:, [1], :, [1]] creates a shape of (1, 2, 2), as the slicing : at dim 0 should reserve the dimensions (2)? How was the shape determined in this case?
x = np.arange(2 * 2 * 2 * 2).reshape((2, 2, 2, 2))
print(f"x:\n{x}")
a = x[:, [1], [1], :]
print("================")
print(f"shape of x[:, [1], [1], :] == {a.shape}")
print(a)
print("================")
b = x[:, [1], :, [1]]
print(f"shape of x[:, [1], :, [1]] == {b.shape}")
print(b)
The results are
x:
[[[[ 0 1]
[ 2 3]]
[[ 4 5]
[ 6 7]]]
[[[ 8 9]
[10 11]]
[[12 13]
[14 15]]]]
================
shape of x[:, [1], [1], :] == (2, 1, 2)
[[[ 6 7]]
[[14 15]]]
================
shape of x[:, [1], :, [1]] == (1, 2, 2)
[[[ 5 7]
[13 15]]]
