Ensure the validator is referenced in this decorator (capability_validator=rand_validator)

Same for the decorators below

Based on the numpy and Torch docstrings (ATen, Numpy), I think you would need to unpack the integers with something like np.random.randn(*args[0]). Additionally, .to would not have effect in Numpy

This should be in the validator, since converters should not throw errors

To remove

Numpy would use permutation, as here

As above, I think you would need to unpack the integers with something like np.random.rand(*args[0]). Additionally, .to would not have effect in Numpy

Thanks for the review. I think np.random.rand(*args) should also work, since it passes locally. Let me cross check.

To remove

Can be removed since it is unused

Same as above

Same as above

This fails on my machine, since np.random.randn does not accept tuples, and args is a tuple containing a tuple:

>>> args = ((1, 2, 3,),)
>>> np.random.randn(*args)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "mtrand.pyx", line 1270, in numpy.random.mtrand.RandomState.randn
  File "mtrand.pyx", line 1431, in numpy.random.mtrand.RandomState.standard_normal
  File "_common.pyx", line 636, in numpy.random._common.cont
TypeError: 'tuple' object cannot be interpreted as an integer

@gs-olive Thanks for pointing this out.
I think this got missed in the tests, since in the tests the input is None, and the compilation is not invoked.
I changedthe above to have np.random.rand(*args[0]). Also in the test I modified it to have an input (hacky way, I can as well use the harness.py with the below model)

def test_rand(self, name, op, shape_or_input):
        class TestModule(nn.Module):
            def __init__(self, rand_op, size):
                super().__init__()
                self.rand_op = rand_op
                self.size = size
            def forward(self, x):
                b = x + 1
                self.size[0] = b
                return self.rand_op(self.size)

But right now I am running into segmentation fault. I am looking into this further.

Same error as above on my machine

Does this only test the shapes of the results? If so, can the name be updated to run_test_compare_shapes_only or something similar?

In this case it is checking the shape of the Tensors. But it is supposed to be a generic function where the callable can check other attribute of the Tensor. For example, in the case of rand and randn we compare the data type as well. So I think the name should be run_test_comparator but you can let me know if you think otherwise.

Ok that makes sense, I think the name is reasonable then. If possible, maybe run_test_compare_tensor_attributes_only just to delineate that the actual values in the tensors are ignored, but it is also okay as is

Ok makes sense, I will rename it to the above then.