Discussion: methods vs. function in R and Python

[skeydan]

Currently, with tensor modification, our strategy seems to be:

• R functions (torch_) only when a fresh object is returned
• R methods (x$what()) only when the existing object is modified

e.g. tx$view() but torch_reshape

I assume this is why I don't see an x$clone()
(also I notice we have x$t_() but not x$t())

But, a person coming from Python might expect $ access to work independently of whether there are side effects or not (there is an x.clone(), etc.)

Not saying we have to follow this ... but partly we do: For example, there is x$mm as well as x$mm_ in R ...

What do you think?

[dfalbel]

IMHO we should closely follow python here. Ie if a method exists in Python it should also exist in R.

x$t() exists in R the side:

> x <- torch_randn(10, 5)
> x$t()
torch_tensor 
 1.5990 -1.2676  1.4600 -0.6393 -0.8593  1.0309  0.4477  1.0159  0.0463  0.0362
 2.1157  0.7238  0.0673  0.7179  0.1117 -1.0217 -2.2464  0.0867  1.2638 -0.7934
-1.2125 -1.0230 -0.0952 -1.1130  0.2201 -0.3773  0.5066 -0.8243 -0.3650 -0.2140
-0.2426  0.0644  0.2774 -0.9705  1.0606 -1.2880  0.8776 -0.4154  0.1012 -1.9847
-0.6819 -0.0620 -1.3663 -1.2802  0.7495 -0.3198 -1.4872 -0.9426  1.0952  0.2087
[ CPUFloatType{5,10} ]

clone is a special case, because it was not possible to add a clone method to an R6 class.
so we renamed it to copy. We should be able to call it clone again, since we are no longer using R6 classes for tensors.

[skeydan]

Oh, I completely agree! Perfect. Yes, then it'd be nice to have a clone as well!

Why I thought there was no x$t():

t$mm_(t$t())
Error: attempt to apply non-function

(but tbh I didn't investigate)

[dfalbel]

I think in this case, mm_ is the one that doesn't exist.

[skeydan]

oh, right - it really does not exist (in Python) :-)

Closing this and opening a separate one for clone :-)