This needs documentation in _torch_docs.py and tests in test_torch.py.

These are the behaviours on corner cases otherwise working fine :

=>torch.randint(5,(3))
Traceback (most recent call last):
File "", line 1, in
TypeError: randint(): argument 'size' (position 2) must be tuple of ints, not int

=>torch.randint(5,3)
Traceback (most recent call last):
File "", line 1, in
TypeError: randint(): argument 'size' (position 2) must be tuple of ints, not int

=>torch.randint(5,10,3)
Traceback (most recent call last):
File "", line 1, in
TypeError: randint() received an invalid combination of arguments - got (int, int, int), but expected one of:
(int high, tuple of ints size, torch.Generator generator, torch.dtype dtype, int device, bool requires_grad)
(int low, int high, tuple of ints size, torch.Generator generator, torch.dtype dtype, int
device, bool requires_grad)

=>torch.randint(3,5,())
4
[torch.FloatTensor of size ()]

The first outcome is bothering me. Any fix? (Others are fine in my opinion)

@pytorchbot test this please

The behavior in the first case is just a Python-ism. (3) is the same as 3; the one element tuple is (3, ).

@ezyang yeah that case is working (I should have included it in my examples). Also torch.randn((3)) works right, so therefore torch.randint((3,5,(4)) must also work for sake of continuity right?

EDIT :
Some additional inputs

torch.randint(6,22,(3,))
7
21
6
[torch.FloatTensor of size (3,)]

torch.randint(6,22,(2,2))
19 11
18 9
[torch.FloatTensor of size (2,2)]

@pytorchbot test this please

So, perhaps what you are suggesting is that randint should have an option randint(low, high, *args), where args are sizes?

@ezyang No we need to pass the size as a tuple only as we provide
randint(5,(2,2)) to give random integers in range 0-5
If we dont enclose size in tuple then we cant differentiate between randint(5,2,(2)) and randint(5,(2,2))

EDIT : I am suggesting to have option for Intlist to accept tuple of size 1 as well, which it somehow does for torch.randn

No, what I am saying is that (3) is not a tuple of size one.

>>> (3)
3
>>> (3,)
(3,)

It accidentally works in torch.randn((3)) case because randn has an overload torch.randn(*size_args).

Best not to get too crazy with the overloads here, IMO.

@ezyang
Ohh i see!! So i believe currently it's working fine, as we cant provide size_args overloading coz it'll complicate things. Will write the documentation and tests then, and also mention in docs specifically that for tensor of size 3 use sizes as (3,)

By the way if I overload it as torch.randint(int,int,int) then torch.randint(1,5,(3)) would work right. I can do that if you want. (Hacky I know :P )

I hate it! Leave it the way it is ;)

Looks reasonable, but this still needs doc and tests.

Will commit by tommorow @ssnl :)

I am having doubt how should I document the function.
As I have used function overloading I can create a function signature like this :

torch.randint(low=0,high,sizes) (low can be optional)

Now I think it might confuse some people as (in general) it's not possible for first argument to be optional before second and third arguments (They dont know I used function overloading for the feature).

Do you think it's alright to document it like this??

We already have docs where first arg is written as optional, e.g. http://pytorch.org/docs/master/torch.html#torch.arange. So it should be fine.

@ssnl I guess it's ready to merge now?

@ssnl is there anything more to add or correct now?

@ssnl

Can we merge this now?

randint doesn't show up in site docs: http://pytorch.org/docs/master/search.html?q=randint&check_keywords=yes&area=default

It'll probably come in next release!

I think some autofunction entry is needed, so that Sphinx would pull up the docs from the doc string

Looks like @vadimkantorov is right.
@Naman-ntc can you add an entry for randint in https://github.com/pytorch/pytorch/blob/master/docs/source/torch.rst ? Probably something like

.. autofunction:: randint

in the random functions section, look here

Ohh I see!
Makes sense, will make a new PR for it.