Hi,
I have just sent a PR (https://github.com/numpy/numpy/pull/5015), adding the possibility of having frozen dimensions in gufunc signatures. As a proof of concept, I have added a `cross1d` gufunc to `numpy.core.umath_tests`:
In [1]: import numpy as np In [2]: from numpy.core.umath_tests import cross1d
In [3]: cross1d.signature Out[3]: '(3),(3)->(3)'
In [4]: a = np.random.rand(1000, 3) In [5]: b = np.random.rand(1000, 3)
In [6]: np.allclose(np.cross(a, b), cross1d(a, b)) Out[6]: True
In [7]: %timeit np.cross(a, b) 10000 loops, best of 3: 76.1 us per loop
In [8]: %timeit cross1d(a, b) 100000 loops, best of 3: 13.1 us per loop
In [9]: c = np.random.rand(1000, 2) In [10]: d = np.random.rand(1000, 2)
In [11]: cross1d(c, d) --------------------------------------------------------------------------- ValueError Traceback (most recent call last) <ipython-input-11-72c66212e40c> in <module>() ----> 1 cross1d(c, d)
ValueError: cross1d: Operand 0 has a mismatch in its core dimension 0, with gufunc signature (3),(3)->(3) (size 2 is different from 3)
The speed up over `np.cross` is nice, and while `np.cross` is not the best of examples, as it needs to handle more sizes, in many cases this will allow producing gufuncs that work without a Python wrapper redoing checks that are best left to the iterator, such as dimension sizes.
It still needs tests, but before embarking on fully developing those, I wanted to make sure that there is an interest on this.
I would also like to further enhance gufuncs providing computed dimensions, e.g. making it possible to e.g. define `pairwise_cross` with signature '(n, 3)->($m, 3)', where the $ indicates that m is a computed dimension, that would have to be calculated by a function passed to the gufunc constructor and stored in the gufunc object, based on the other core dimensions. In this case it would make $m be n*(n-1), so that all pairwise cross products between 3D vectors could be computed.
The syntax with '$' is kind of crappy, so any suggestions on how to better express this in the signature are more than welcome, as well as any feedback on the merits (or lack of them) of implementing this.
Jaime
On Fri, Aug 29, 2014 at 1:14 AM, Jaime Fernández del Río jaime.frio@gmail.com wrote:
Hi,
I have just sent a PR (https://github.com/numpy/numpy/pull/5015), adding the possibility of having frozen dimensions in gufunc signatures. As a proof of concept, I have added a `cross1d` gufunc to `numpy.core.umath_tests`:
In [1]: import numpy as np In [2]: from numpy.core.umath_tests import cross1d
In [3]: cross1d.signature Out[3]: '(3),(3)->(3)'
In [4]: a = np.random.rand(1000, 3) In [5]: b = np.random.rand(1000, 3)
In [6]: np.allclose(np.cross(a, b), cross1d(a, b)) Out[6]: True
In [7]: %timeit np.cross(a, b) 10000 loops, best of 3: 76.1 us per loop
In [8]: %timeit cross1d(a, b) 100000 loops, best of 3: 13.1 us per loop
In [9]: c = np.random.rand(1000, 2) In [10]: d = np.random.rand(1000, 2)
In [11]: cross1d(c, d)
ValueError Traceback (most recent call last) <ipython-input-11-72c66212e40c> in <module>() ----> 1 cross1d(c, d)
ValueError: cross1d: Operand 0 has a mismatch in its core dimension 0, with gufunc signature (3),(3)->(3) (size 2 is different from 3)
The speed up over `np.cross` is nice, and while `np.cross` is not the best of examples, as it needs to handle more sizes, in many cases this will allow producing gufuncs that work without a Python wrapper redoing checks that are best left to the iterator, such as dimension sizes.
It still needs tests, but before embarking on fully developing those, I wanted to make sure that there is an interest on this.
I would also like to further enhance gufuncs providing computed dimensions, e.g. making it possible to e.g. define `pairwise_cross` with signature '(n, 3)->($m, 3)', where the $ indicates that m is a computed dimension, that would have to be calculated by a function passed to the gufunc constructor and stored in the gufunc object, based on the other core dimensions. In this case it would make $m be n*(n-1), so that all pairwise cross products between 3D vectors could be computed.
The syntax with '$' is kind of crappy, so any suggestions on how to better express this in the signature are more than welcome, as well as any feedback on the merits (or lack of them) of implementing this.
Some thoughts:
When I first saw the PR my first reaction was that maybe we should be allowing more general hooks for a gufunc to choose its core dimensions. Reading the code convinced me that this is a relatively minimal enhancement over what we're currently doing, so your current PR looks fine to me.
But, for your computed dimension idea I'm wondering if what we should do instead is just let a gufunc provide a C callback that looks at the input array dimensions and explicitly says somehow which dimensions it wants to treat as the core dimensions and what its output shapes will be. There's no rule that we have to extend the signature mini-language to be Turing complete, we can just use C :-).
It would be good to have a better motivation for computed gufunc dimensions, though. Your "all pairwise cross products" example would be *much* better handled by implementing the .outer method for binary gufuncs: pairwise_cross(a) == cross.outer(a, a). This would make gufuncs more consistent with ufuncs, plus let you do all-pairwise-cross-products between two different sets of cross products, plus give us all-pairwise-matrix-products for free, etc.
While you're messing around with the gufunc dimension matching logic, any chance we can tempt you to implement the "optional dimensions" needed to handle '@', solve, etc. elegantly? The rule would be that you can write something like (n?,k),(k,m?)->(n?,m?) and the ? dimensions are allowed to take on an additional value "nothing at all". If there's no dimension available in the input, then we act like it was reshaped to add a dimension with shape 1, and then in the output we squeeze this dimension out again. I guess the rules would be that (1) in the input, you can have ? dimensions at the beginning or the end of your shape, but not both at the same time, (2) any dimension that has a ? in one place must have it in all places, (3) when checking argument conformity, "nothing at all" only matches against "nothing at all", not against 1; this is because if we allowed (n?,m),(n?,m)->(n?,m) to be applied to two arrays with shapes (5,) and (1, 5), then it would be ambiguous whether the output should have shape (5,) or (1, 5).
-n
On Thu, Aug 28, 2014 at 5:40 PM, Nathaniel Smith njs@pobox.com wrote:
Some thoughts:
But, for your computed dimension idea I'm wondering if what we should do instead is just let a gufunc provide a C callback that looks at the input array dimensions and explicitly says somehow which dimensions it wants to treat as the core dimensions and what its output shapes will be. There's no rule that we have to extend the signature mini-language to be Turing complete, we can just use C :-).
It would be good to have a better motivation for computed gufunc dimensions, though. Your "all pairwise cross products" example would be *much* better handled by implementing the .outer method for binary gufuncs: pairwise_cross(a) == cross.outer(a, a). This would make gufuncs more consistent with ufuncs, plus let you do all-pairwise-cross-products between two different sets of cross products, plus give us all-pairwise-matrix-products for free, etc.
The outer for binary gufuncs sounds like a good idea. A reduce for binary gufuncs that allow it (like square matrix multiplication) would also be nice. But going back to the original question, the pairwise whatevers were just an example: one could come up with several others, e.g.:
(m),(n)->($p),($q) with $p = m - n and $q = n - 1, could be (I think) the signature of a polynomial division gufunc (m),(n)->($p), with $p = m - n + 1, could be the signature of a convolution or correlation gufunc (m)->($n), with $n = m / 2, could be some form of downsampling gufunc
While you're messing around with the gufunc dimension matching logic, any chance we can tempt you to implement the "optional dimensions" needed to handle '@', solve, etc. elegantly? The rule would be that you can write something like (n?,k),(k,m?)->(n?,m?) and the ? dimensions are allowed to take on an additional value "nothing at all". If there's no dimension available in the input, then we act like it was reshaped to add a dimension with shape 1, and then in the output we squeeze this dimension out again. I guess the rules would be that (1) in the input, you can have ? dimensions at the beginning or the end of your shape, but not both at the same time, (2) any dimension that has a ? in one place must have it in all places, (3) when checking argument conformity, "nothing at all" only matches against "nothing at all", not against 1; this is because if we allowed (n?,m),(n?,m)->(n?,m) to be applied to two arrays with shapes (5,) and (1, 5), then it would be ambiguous whether the output should have shape (5,) or (1, 5).
I definitely do not mind taking a look into it. I need to think a little more about the rules to convince myself that there is a consistent set of them that we can use. I also thought there may be a performance concern, that you may want to have different implementations when dimensions are missing, not automatically add a 1 and then remove it. It doesn't seem to be the case with neither `np.dot` nor `np.solve`, so maybe I am being overly cautious.
Thanks for your comments and ideas. I have a feeling there are some nice features hidden in here, but I can't seem to figure out what should they be on my own.
Jaime
On Fri, Aug 29, 2014 at 2:55 AM Jaime Fernández del Río < jaime.frio@gmail.com> wrote:
On Thu, Aug 28, 2014 at 5:40 PM, Nathaniel Smith njs@pobox.com wrote:
Some thoughts:
But, for your computed dimension idea I'm wondering if what we should do instead is just let a gufunc provide a C callback that looks at the input array dimensions and explicitly says somehow which dimensions it wants to treat as the core dimensions and what its output shapes will be. There's no rule that we have to extend the signature mini-language to be Turing complete, we can just use C :-).
It would be good to have a better motivation for computed gufunc dimensions, though. Your "all pairwise cross products" example would be *much* better handled by implementing the .outer method for binary gufuncs: pairwise_cross(a) == cross.outer(a, a). This would make gufuncs more consistent with ufuncs, plus let you do all-pairwise-cross-products between two different sets of cross products, plus give us all-pairwise-matrix-products for free, etc.
The outer for binary gufuncs sounds like a good idea. A reduce for binary gufuncs that allow it (like square matrix multiplication) would also be nice. But going back to the original question, the pairwise whatevers were just an example: one could come up with several others, e.g.:
(m),(n)->($p),($q) with $p = m - n and $q = n - 1, could be (I think)
the signature of a polynomial division gufunc (m),(n)->($p), with $p = m - n + 1, could be the signature of a convolution or correlation gufunc (m)->($n), with $n = m / 2, could be some form of downsampling gufunc
While you're messing around with the gufunc dimension matching logic, any chance we can tempt you to implement the "optional dimensions" needed to handle '@', solve, etc. elegantly? The rule would be that you can write something like (n?,k),(k,m?)->(n?,m?) and the ? dimensions are allowed to take on an additional value "nothing at all". If there's no dimension available in the input, then we act like it was reshaped to add a dimension with shape 1, and then in the output we squeeze this dimension out again. I guess the rules would be that (1) in the input, you can have ? dimensions at the beginning or the end of your shape, but not both at the same time, (2) any dimension that has a ? in one place must have it in all places, (3) when checking argument conformity, "nothing at all" only matches against "nothing at all", not against 1; this is because if we allowed (n?,m),(n?,m)->(n?,m) to be applied to two arrays with shapes (5,) and (1, 5), then it would be ambiguous whether the output should have shape (5,) or (1, 5).
I definitely do not mind taking a look into it. I need to think a little more about the rules to convince myself that there is a consistent set of them that we can use. I also thought there may be a performance concern, that you may want to have different implementations when dimensions are missing, not automatically add a 1 and then remove it. It doesn't seem to be the case with neither `np.dot` nor `np.solve`, so maybe I am being overly cautious.
Thanks for your comments and ideas. I have a feeling there are some nice features hidden in here, but I can't seem to figure out what should they be on my own.
Jaime
-- (__/) ( O.o) ( > <) Este es Conejo. Copia a Conejo en tu firma y ayúdale en sus planes de dominación mundial. _______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@scipy.org http://mail.scipy.org/mailman/listinfo/numpy-discussion
I'm not sure where this is at, given the current amount of work that is coming from the 1.10 release, but this sounds like a really great idea. Computed/fixed dimensions would allow gufuncs for things like: - polynomial multiplication, division, differentiation, and integration - convolutions - views of different types (see the corresponding discussion at http://permalink.gmane.org/gmane.comp.python.numeric.general/59847). Some of these examples would work better with gufuncs that can construct views and have an axes keyword, but this is exactly the kind of functionality that would be really great to have. Thanks for the great work! -Ian Henriksen
On Fri, Aug 29, 2014 at 10:55 AM, Jaime Fernández del Río < jaime.frio@gmail.com> wrote:
On Thu, Aug 28, 2014 at 5:40 PM, Nathaniel Smith njs@pobox.com wrote:
Some thoughts:
But, for your computed dimension idea I'm wondering if what we should do instead is just let a gufunc provide a C callback that looks at the input array dimensions and explicitly says somehow which dimensions it wants to treat as the core dimensions and what its output shapes will be. There's no rule that we have to extend the signature mini-language to be Turing complete, we can just use C :-).
It would be good to have a better motivation for computed gufunc dimensions, though. Your "all pairwise cross products" example would be *much* better handled by implementing the .outer method for binary gufuncs: pairwise_cross(a) == cross.outer(a, a). This would make gufuncs more consistent with ufuncs, plus let you do all-pairwise-cross-products between two different sets of cross products, plus give us all-pairwise-matrix-products for free, etc.
The outer for binary gufuncs sounds like a good idea. A reduce for binary gufuncs that allow it (like square matrix multiplication) would also be nice. But going back to the original question, the pairwise whatevers were just an example: one could come up with several others, e.g.:
(m),(n)->($p),($q) with $p = m - n and $q = n - 1, could be (I think)
the signature of a polynomial division gufunc (m),(n)->($p), with $p = m - n + 1, could be the signature of a convolution or correlation gufunc (m)->($n), with $n = m / 2, could be some form of downsampling gufunc
An example where a computed output dimension would be useful is with linalg.svd, as some resulting dimensions for a matrix (m, n) are based on min(m, n). This, coupled with the required keyword support makes it necessary to have 6 gufuncs to support the functionality.
I do think that the C callback solution would be enough, and just allow the signature to have unbound variables that can be resolved by that callback... no need to change the syntax:
(m),(n)->(p),(q)
When registering such a gufunc, a callback function that resolves the missing dimensions would be required.
Extra niceties that could be built on top of that: - pass keyword arguments to that function so that stuff like full_matrices could be resolved inside the gufunc. Maybe even allowing to modify the number of results (harder) that would be needed to support stuff like "compute_uv" in svd as well.
- allow context to be created in that resolution that gets passed into the ufunc kernel itself (note that this might be *necessary*). If context is created another function would be needed to dispose that context.
In my experience when implementing the linalg gufunc, a very common pattern was needing some buffers for the actual LAPACK calls (as those functions are inplace, a tmp buffer was always needed). Some setup and buffer allocation was performed before looping. Every iteration in the inner loop will reuse that data and at the end of the loop the buffers will be released. That means the initialization/allocation/release is done once per inner loop call. If the hooks to allocate/dispose the context existed, that initialization/allocation/release could be done once per ufunc call. AFAIK, a ufunc call can involve several inner loop calls depending on outer dimensions and layout of the operands.
While you're messing around with the gufunc dimension matching logic,
any chance we can tempt you to implement the "optional dimensions" needed to handle '@', solve, etc. elegantly? The rule would be that you can write something like (n?,k),(k,m?)->(n?,m?) and the ? dimensions are allowed to take on an additional value "nothing at all". If there's no dimension available in the input, then we act like it was reshaped to add a dimension with shape 1, and then in the output we squeeze this dimension out again. I guess the rules would be that (1) in the input, you can have ? dimensions at the beginning or the end of your shape, but not both at the same time, (2) any dimension that has a ? in one place must have it in all places, (3) when checking argument conformity, "nothing at all" only matches against "nothing at all", not against 1; this is because if we allowed (n?,m),(n?,m)->(n?,m) to be applied to two arrays with shapes (5,) and (1, 5), then it would be ambiguous whether the output should have shape (5,) or (1, 5).
I definitely do not mind taking a look into it. I need to think a little more about the rules to convince myself that there is a consistent set of them that we can use. I also thought there may be a performance concern, that you may want to have different implementations when dimensions are missing, not automatically add a 1 and then remove it. It doesn't seem to be the case with neither `np.dot` nor `np.solve`, so maybe I am being overly cautious.
Thanks for your comments and ideas. I have a feeling there are some nice features hidden in here, but I can't seem to figure out what should they be on my own.
Jaime
-- (__/) ( O.o) ( > <) Este es Conejo. Copia a Conejo en tu firma y ayúdale en sus planes de dominación mundial.
NumPy-Discussion mailing list NumPy-Discussion@scipy.org http://mail.scipy.org/mailman/listinfo/numpy-discussion