Any plan to make Conv and DeConv use the Const architecture ?

[alexander-camuto]

Heya :)

If I recall the tract API recently(ish) made the transition from an architecture where the constantts involved in an operation are carried at the Op level, for instance:

Matmul {
a: Option<Tensor<i128>>
b: Option<Tensor<i128>>
}

to one where constants are represented as independent nodes that are then ingested by the op.

Const{
name: "a"
value: Tensor<i128>
}
Const{
name: "b"
value: Tensor<i128>
}
Matmul 

A notable exception to this are the Conv and Deconv nodes which still carry the kernel and bias with them. I was wondering if there way any plan to change this and move these two ops more towards the current API architecture.

I ask because we have some processing steps that operate on all constants but then have to apply special rules to Conv and DeConv nodes -- just wondering if this will change.

[kali]

Maybe... I am moving away from attributes and towards more inputs, a bit like ONNX has been doing. I can see another benefits to this approach in the Dynamic Convolution models (not sure if these are still around of if the hype has gone) where you compute the weights to the convolution as part of your network. But I don't have definite plan for that.

But I'll keep it in mind, who knows, maybe when I'm masochist enough to go near the convolution code again, I may do it :)

[alexander-camuto]

sweet thank you !

[alexander-camuto]

Hello @kali recent updates seem to have done this which is amazing. is that correct ?

[kali]

Absolutely. And it's been a huge pain, you definitely owe me one if we meet one day :)

[alexander-camuto]

very kind -- let me know if you want me to grab you a beer end of feb !