[FIXED] Numpy indexing with ndarray and PyTorch tensor

Issue

I find numpy array indexing works differently with ndarrray and PyTorch tensor of shape (1,) and want to know why. Please see the case below:

import numpy as np
import torch as th

x = np.arange(10)

y = x[np.array([1])]
z = x[th.tensor([1])]
print(y, z)

y would be array[2] while z being just 2. What is the difference exactly?

Solution

Note that integer tensors of a single element can be converted to an index:

>>> torch.tensor([1]).__index__()
1
>>> torch.tensor([1, 2]).__index__()
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: only integer tensors of a single element can be converted to an index

When the index passed in is tensor, ndarray cannot recognize it, so it attempts to call its __index__ method. If the conversion is successful, it is treated as an integer:

if (PyLong_CheckExact(obj) || !PyArray_Check(obj)) {
    // it calls PyNumber_Index() internally
    npy_intp ind = PyArray_PyIntAsIntp(obj);

    if (error_converting(ind)) {
        PyErr_Clear();
    }
    else {
        index_type |= HAS_INTEGER;
        indices[curr_idx].object = NULL;
        indices[curr_idx].value = ind;
        indices[curr_idx].type = HAS_INTEGER;
        used_ndim += 1;
        new_ndim += 0;
        curr_idx += 1;
        continue;
    }
}

_{Source code from NumPy.}

Therefore, the effect of code of OP is equivalent to the following code:

>>> np.arange(10)[1]
1

Answered By - Mechanic Pig