[Numpy-discussion] Consider improving numpy.outer's behavior with zero-dimensional vectors
josef.pktd at gmail.com
josef.pktd at gmail.com
Fri Apr 17 12:09:27 EDT 2015
On Fri, Apr 17, 2015 at 11:22 AM, Neil Girdhar <mistersheik at gmail.com> wrote:
> On Fri, Apr 17, 2015 at 10:47 AM, <josef.pktd at gmail.com> wrote:
>> On Fri, Apr 17, 2015 at 10:07 AM, Sebastian Berg
>> <sebastian at sipsolutions.net> wrote:
>> > On Do, 2015-04-16 at 15:28 -0700, Matthew Brett wrote:
>> >> Hi,
>> > <snip>
>> >> So, how about a slight modification of your proposal?
>> >> 1) Raise deprecation warning for np.outer for non 1D arrays for a few
>> >> versions, with depraction in favor of np.multiply.outer, then
>> >> 2) Raise error for np.outer on non 1D arrays
>> > I think that was Neil's proposal a bit earlier, too. +1 for it in any
>> > case, since at least for the moment I doubt outer is used a lot for non
>> > 1-d arrays. Possible step 3) make it work on higher dims after a long
>> > period.
>> sounds ok to me
>> Some random comments of what I remember or guess in terms of usage
>> I think there are at most very few np.outer usages with 2d or higher
>> (statsmodels has two models that switch between 2d and 1d
>> parameterization where we don't use outer but it has similar
>> characteristics. However, we need to control the ravel order, which
>> IIRC is Fortran)
>> The current behavior of 0-D scalars in the initial post might be
>> useful if a numpy function returns a scalar instead of a 1-D array in
>> size=1. np.diag which is a common case, doesn't return a scalar (in my
>> version of numpy).
>> I don't know any use case where I would ever want to have the 2d
>> behavior of np.multiply.outer.
I only understand part of your example, but it looks similar to what
we are doing in statsmodels.
> My use case is pretty simple. Given an input vector x, and a weight matrix
> W, and a model y=Wx, I calculate the gradient of the loss L with respect W.
> It is the outer product of x with the vector of gradients dL/dy. So the
> code is simply:
> W -= outer(x, dL_by_dy)
if you sum/subtract over all the values, isn't this the same as
> Sometimes, I have some x_indices and y_indices. Now I want to do:
> W[x_indices, y_indices] -= outer(x[x_indices], dL_by_dy[y_indices])
> Unfortunately, if x_indices or y_indices are "int" or slice in some way that
> removes a dimension, the left side will have fewer dimensions than the
> right. np.multipy.outer does the right thing without the ugly cases:
> if isinstance(x_indices, int): … # ugly hacks follow.
My usual hacks are either to use np.atleast_1d or np.atleast_1d or
np.squeeze if there is shape mismatch in some cases.
>> I guess we will or would have applications for outer along an axis,
>> for example if x.shape = (100, 10), then we have
>> x[:,None, :] * x[:, :, None] (I guess)
>> Something like this shows up reasonably often in econometrics as
>> "Outer Product". However in most cases we can avoid constructing this
>> matrix and get the final results in a more memory efficient or faster
>> (example an array of covariance matrices)
> Not sure I see this. outer(a, b) should return something that has shape:
> (a.shape + b.shape). If you're doing it "along an axis", you mean you're
> reshuffling the resulting shape vector?
No I'm not reshaping the full tensor product.
It's a vectorized version of looping over independent outer products
np.array([outer(xi, yi) for xi,yi in zip(x, y)])
(which I would never use with outer)
but I have code that works similar for a reduce (or reduce_at) loop over this.
>> > - Sebastian
>> >> Best,
>> >> Matthew
>> >> _______________________________________________
>> >> NumPy-Discussion mailing list
>> >> NumPy-Discussion at scipy.org
>> >> http://mail.scipy.org/mailman/listinfo/numpy-discussion
>> > _______________________________________________
>> > NumPy-Discussion mailing list
>> > NumPy-Discussion at scipy.org
>> > http://mail.scipy.org/mailman/listinfo/numpy-discussion
>> NumPy-Discussion mailing list
>> NumPy-Discussion at scipy.org
> NumPy-Discussion mailing list
> NumPy-Discussion at scipy.org
More information about the NumPy-Discussion