Reconsider equivalence between type.broadcastable and type.shape #1170
Replies: 1 comment 1 reply
-
|
Don't forget that we're trying to phase out use of The only reason |
Beta Was this translation helpful? Give feedback.
-
|
I know that, but many old rewrites relied on the older interpretation of broadcastable which is not the same as the new one. |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
-
Right now
TensorType("float64", shape=(None, 5, 1)).broadcastable == (False, False, True)However, due to enabling dynamic broadcasting, this is not correct as that type can broadcast along the first dimension when that has a size of 1 at runtime. In many rewrites we check for conditions like
var1.dtype.broadcastable == var2.dtype.broadcastable, but we shouldn't assume they are the same when they come fromNonetype shape dimensions.One option would be to change to the following:
TensorType("float64", shape=(None, 5, 1)).broadcastable == (np.nan, False, True), or any other type that respects(self == self) is Falseasnp.nandoes.Related to #1089 and #1122
Beta Was this translation helpful? Give feedback.
All reactions