More consistency with array-methods

It’s sometimes hard to tell in numpy whether a given numpy function will also be a method of an array.

e.g. You can go arr.max() or np.max(arr), but you can’s go arr.abs() (you need to go np.abs(arr).

Same goes for most functions: median, sign, exp, etc. It would be nice, because I find it much more readable to look at:

mean_rmse = ((out-targ)**2).mean(axis=1).sqrt().mean(axis=0)

Than:

mean_rmse = np.mean(np.sqrt(np.mean((out-targ)**2, axis =1)), axis=0)

[njsmith]

I’m a bit confused at why this is a surprise, given that everyone else commenting in this issue has been talking about why it’s a bad idea and whether there’s any way we can remove methods from ndarray.

[petered]

Thanks Ralf, the array API standard looks like a great initiative. Though I still after all these years would strongly prefer a method-based syntax, both for readability (avoiding the bracket-matching nightmare discussed at top of post) and ability to re-use code like

def sigmoid(x): 
    return 1/(1+(-x).exp())

And have it be automatically valid in numpy, tensorflow, pytorch, etc. Though it would be nice if there were some “extension function” syntax that would enable us to call this like x.sigmoid().