Guillaume Lagrange

> @laggui I'm down to provide an api for D2 + conv; but I don't want to lose the raw computation. Not sure I follow, what do you mean by...

> `raw computation`; I want to be able to explicitly re-use the _exact same_ shape calculations in library code that the `conv` family uses; so that we can tie contracts...

@nathanielsimard remember when I first made the tests strict and had to convert every data type for correct comparison then we decided to make it less strict? Where do we...

I think the first proposed solution is too restrictive, because not all backends support the same dtypes necessarily. And we might want to run tests in f16 at some point....

I think this is correctly handled. For strict `assert_eq`, we make sure to compare with the backend element type. Otherwise, we can have data equality (not strict on dtype), and...

Sorry for the late reply. Does it converge after more than one epoch? For a single epoch, the difference could be in the random initialization.

Thanks for checking with more epochs! There does seem to be an issue, at 10 epochs it should at least be trending towards a converging solution. That is apparently not...

Thanks for all the feedback 🙏 The example is relatively straightforward so I don't have anything else in mind, will probably have to check it out to see if it...

Most of the work would be in CubeCL, and would only be exposed to Burn through an API for cubecl backends. I think a step in that direction would be...

Hmmm maybe there are still some refs being held for some tensors. The handles for our cubecl backends are ref counted, which might be the issue within the prediction loop....