[Question] Why don't tensors implement the Copy trait?

[EddieMataEwy]

Hi, I was just wondering why tensors implement the Clone trait but not the Copy trait.
What is the reason for this?

I just find myself writing .clone() over and over again, and I don't see any reason why you couldn't derive Copy as well.

Thank you for your time.

[nathanielsimard]

I don't think Copy works by just calling clone implicitly, it's a marker trait that tells the compiler the type can be copied directly in memory. However, a tensor clone isn't a copy, it just increments a reference, so the reason is probably the same why Copy isn't implemented for Arc.

[EddieMataEwy]

I see. In retrospect that was a dumb question. I just saw the Tensors are implemented as Arc Arrays.
In this case, would it be possible to have many of the tensor ops take a reference to a tensor so that the compiler implements dereference coercion automatically? Or would we run into similar issues?

[nathanielsimard]

Most of the optimizations of Burn are because we use owned tensors as arguments for all our operations. Using the ownership system to know when a tensor can be reused. This allows us to capture the graph and is fundamental in our optimization strategy with the upcoming burn-fusion. So, using tensor as reference would actually make this impossible and slow down the framework. I agree that calling clone when reusing a tensor isn't really pretty, but you can actually use clippy to minimize the number of clones and, in the same way, optimize your model!

[EddieMataEwy]

Gotcha! Thanks for the explanation and for your time. Your work here is amazing. Keep going!