implicit cast from Tensor to TensorIndex

[yueyinqiu]

closes #1266

---

It includes some aggressive changes. Careful checks are required.

[yueyinqiu]

• I have removed the implicit operator from TensorIndex to Tensor.
  • I don't understand why it should exist to just throw an exception... but the behavior is even tested in the unit tests.
• I have removed the two indexers this[params Tensor[] indices] and this[params long[] indices] since tensors could now be implicitly cast to TensorIndex.
  • However if there is something else, say a user defined class A, can be implicitly cast to a tensor, which used to be able to be used as the indices, is now unable to be used because it won't be automatically cast twice. This is happening for int? in TorchVision/AutoAugment.cs.

There might be other potential risks since it is an implicit operator to be added.

[NiklasGustafsson]

Hmmm. Not sure about this one. Please elaborate.

[yueyinqiu]

1. add implicit operator TensorIndex(Tensor tensor) into TensorIndex

This is what we want.

2. remove implicit operator Tensor(TensorIndex value) of TensorIndex

This operator is just throwing an exception. I don't know why it should exist, but this is even tested in a unit test.

3. remove Tensor this[params Tensor[] indices] of Tensor

Well... Now I suppose it shouldn't be removed... We need an override for a array.

4. remove Tensor this[params Tensor[] indices] of Tensor

Similar to 3.

[NiklasGustafsson]

This operator is just throwing an exception.

I added it to make sure that some non-C# language (who knows?) doesn't apply different rules for implicit conversions.

[yueyinqiu]

This operator is just throwing an exception.

I added it to make sure that some non-C# language (who knows?) doesn't apply different rules for implicit conversions.

Sorry, I haven't catch it. So seems that we could remove this now? Or we shall still keep it?

[yueyinqiu]

Amazing... A unit test failed because of the removal of that exception-throwing operator...

It's this line here:

If we don't have the implicit operator TensorIndex -> Tensor, it does not know which indexer should be used (params Tensor[] or params TensorIndex[] are both possible choices)...

And when we remove it and add the new Tensor -> TensorIndex, it will choose the params Tensor[] indexer... But the tensor converted from (0, 2) is a float tensor, which cannot be used as indices...

That's a bit...

Well... Do we really need the implicit operator from a float tuple to a tensor? Actually I suggest to make them explicit. Or we could add (long, long) -> Tensor as well, and then the unit test would pass. (But this may break other things...)

So amazing...

---

A possible solution would be still to remove Tensor this[params Tensor[] indices]. And add Tensor this[Tensor[] indices] for arrays.

[NiklasGustafsson]

Implicit operators are complicated to get right, and since they're invisible in the code that uses them, hard to debug. I actually had to ask Mads Torgersen (lead PM on C#) to help me get this right. The operator you are trying to remove was important to declare but should never be called, as I remember it.

I'd rather we left these operators alone.