Add as_tensor_type method to TensorTable for framework column conversion

[oliverholworthy]

Adds a convenience method to to the TensorTable for converting columns between framework types.

Example usage

table = TensorTable({"feature": np.array([1, 2, 3])})

table.as_tensor_type(torch.Tensor).to_dict()
# => {'a': torch.tensor([1, 2, 3])}

table.as_tensor_type(TorchColumn).to_dict()
# => {'a': torch.tensor([1, 2, 3])}

table = TensorTable({"feature": np.array([1, 2, 3])})
table.as_tensor_type("unknown")
# => Raises `ValueError`
# ValueError: tensor_type argument must be a type. Received: <class 'str'> 
# Supported values are: 'numpy.ndarray', 'torch.Tensor', 'tensorflow.python.framework.ops.Tensor'. 
# Or TensorColumn Types: 'merlin.table.NumpyColumn', 'merlin.table.TorchColumn', 'merlin.table.TensorflowColumn'

[oliverholworthy]

One of the reasons for the method being called .to was to be consistent with the Merlin DTypes API. However, after some further consideration maybe it's not that consistent because to returns a different type from the original.

from merlin.dtypes import dtype

dtype("int64")
# => DType(name='int64', element_type=<ElementType.Int: 'int'>, element_size=64, element_unit=None, signed=True, shape=Shape(dims=None))

dtype("int64").to("tensorflow")
# => tf.int64

Perhaps we could borrow the name astype from pandas which does preserve the original type. Or if we'd like to reserve that for changing the dtypes of the columns rather than the tensor types, maybe .as_tensor_type

[karlhigley]

Those all sound like good suggestions to me, and I can see any of them working okay

[github-actions]

Documentation preview

https://nvidia-merlin.github.io/core/review/pr-285

[oliverholworthy]

I've re-requested your review @karlhigley since I've made some updates that change things quite a bit since you last reviewed.

• renamed .to to .as_tensor_type
• changed the argument type from a string to Union[Type, TensorColumn]