Remove explicit optimization from math.logsumexp

[ricardoV94]

Aesara already performs one pass of this optimization (which will be exactly equivalent to the pymc3 one after pymc-devs/pytensor#465). As things stand, the final graph contains two passes of the same optimization, because the optimized output still contains the same form as the original unoptimized graph log(sum(exp(x))) -> x_max + << log(sum(exp(x - x_max))) >>

https://github.com/pymc-devs/pymc3/blob/c62100c99df9a74194db89863776f155a95c076a/pymc3/math.py#L189-L194

import aesara
import aesara.tensor as at
from pymc3.math import logsumexp

x = at.vector('x')
y = logsumexp(x)
f = aesara.function([x], y)
aesara.dprint(f)
Elemwise{Composite{(i0 + log(i1) + i2)}}[(0, 0)] [id A] ''   9
 |InplaceDimShuffle{x} [id B] ''   5
 | |Reduce{maximum}{0} [id C] 'max'   4
 |   |Elemwise{sub,no_inplace} [id D] ''   3
 |     |x [id E]
 |     |Elemwise{Composite{Switch(IsInf(i0), i1, i0)}}[(0, 0)] [id F] ''   2
 |       |InplaceDimShuffle{x} [id G] ''   1
 |       | |Reduce{maximum}{0} [id H] 'max'   0
 |       |   |x [id E]
 |       |TensorConstant{(1,) of 0} [id I]
 |InplaceDimShuffle{x} [id J] ''   8
 | |Sum{acc_dtype=float64} [id K] ''   7
 |   |Elemwise{Composite{exp((i0 - i1))}}[(0, 0)] [id L] ''   6
 |     |Elemwise{sub,no_inplace} [id D] ''   3
 |     |InplaceDimShuffle{x} [id B] ''   5
 |Elemwise{Composite{Switch(IsInf(i0), i1, i0)}}[(0, 0)] [id F] ''   2

y = at.log(at.sum(at.exp(x), axis=None, keepdims=True))
f = aesara.function([x], y)
aesara.dprint(f)
Elemwise{Composite{(i0 + log(i1))}}[(0, 0)] [id A] ''   5
 |InplaceDimShuffle{x} [id B] ''   1
 | |Reduce{maximum}{0} [id C] 'max'   0
 |   |x [id D]
 |InplaceDimShuffle{x} [id E] ''   4
   |Sum{acc_dtype=float64} [id F] ''   3
     |Elemwise{Composite{exp((i0 - i1))}} [id G] ''   2
       |x [id D]
       |InplaceDimShuffle{x} [id B] ''   1

[brandonwillard]

Looks like the second graph was truncated.

[ricardoV94]

Looks like the second graph was truncated.

Thanks, fixed it