[Frontend][ONNX] No support for `RandomUniform`

[SamKG]

This is mainly meant as a post for discussion purposes - trying to figure out how to best approach supporting this in the API.

Currently, there's no clear way to add support for the random operations.
ONNX relies on the existence of a global key (the op API doesn't specify any key); However, TVM doesn't seem to expose any such value.

Any feedback or advice on how to approach this would be appreciated.

[mbrookhart]

tvm/python/tvm/relay/frontend/onnx.py

Lines 3397 to 3418 in 6d88bdd

	class RandomUniform(OnnxOpConverter):
	"""Operator converter for random_uniform"""
	@classmethod
	def _impl_v1(cls, inputs, attr, params):
	dtype = get_type(attr.get("dtype", 1))
	high = attr.get("high", 1.0)
	low = attr.get("low", 0.0)
	seed = attr.get("seed", None)
	shape = attr["shape"]
	assert dtype in [
	"float32",
	"float64",
	], "Only float random value generation is currently supported."
	if seed is None:
	seed = np.random.randint(1e6)
	key = _random.threefry_key(seed)
	output = _op.random.uniform(key, shape, dtype=dtype, low=low, high=high)
	_, vals = _expr.TupleWrapper(output, 2)
	return vals

We have RandomUniform, but we haven't looked into the rest yet. There is a random number generator based on ThreeFry Counter-based RNG, but it's probably not going to match the seed ONNX uses, so I would recommend testing some statistics on the output vs ONNX (mean, std, etc), rather than directly trying to get the same output.

The threefry generator generates random bits if I remember correctly, you can see how uniform transforms that into a uniform distribution here:

tvm/python/tvm/topi/random/kernel.py

Line 483 in 6d88bdd

def uniform(gen, low, high, out_shape, out_dtype):

[AndrewZhaoLuo]

#8426 Has implemented this.

[AndrewZhaoLuo]

@SamKG is this issue able to be closed? I believe this is implemented now.