Here's an amazing fact about Numpy. Don't fall into this trap!pic.twitter.com/7I64zLEfSN
Show this thread
It would be reasonable to error out at `x + y` due to unclear broadcasting expectations. It would also be reasonable to broadcast the (32,) vector to (32, 1). But how about a third, insane option?
-
-
Saw similar thing recently in another context, better to explicitly reshape instead of relying on broadcasting rules.
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Yeah, definitely learned this the hard way when assessing the accuracy of a ML model ensemble. The way it was composed, one of the models returned predictions in the shape (N, 1). I was wondering why the error was larger than I expected when things looked good visually.
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Looks like NumPy does its broadcast-matching from back to front, so y's 32 matches with x's 1, and... welp. Sometimes make me nervous that batch sizes and feature dims are often powers of 2, and such errors can be very invisible, especially with a big reduce op at the end.
-
When unit testing models I always set the sizes of everything to prime numbers for this reason. It's the closest thing to reasonable typechecking available.
End of conversation
New conversation -
-
-
I don't know what's meant to be the trap here. Python is such an awful mess of random behaviour but the broadcasting rules are one area where you can successfully reason about Python code algebraically
-
I was surprised to find that broadcasting could make an array smaller. But it makes perfect algebraic sense. (Don't look at attached screenshot if you want to set yourself this as a puzzle :)pic.twitter.com/9mTMVplg6z
- Show replies
New conversation -
-
-
Lol, but how 2048?
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
that implies the existence of reasonable broadcasting operations :-)
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Allowing trailing comma is way more trouble than the occasional convenience it buys.
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.