[Unknown, Quansight]
Pull request opened by sangyx
Quansight-Labs/udiff#31
I have added the diff function of np.stack.
When we extend the code to higher-order derivatives, the situation becomes a bit more complicated. There are some places to improve:
• In some derivative functions, the program may enter an infinite loop. For example, np.sum is used when registering the gradient of np.sum. I plan to add depth attributes to DiffArray or to() to control the order of gradients.
• The function registered in nograd_functions will register the gradient when it is called in the form of ufunc.__call__.
• Consider a cleaner way to control whether the calculation should derive.
Do you have any idea about the above questions?
[Unknown, Quansight]
Pull request opened by joaosferreira
Quansight-Labs/unumpy#70
This pull request adds multimethods for statistical functions. The multimethods added are the following:
Order statistics
• percentile
• nanpercentile
• quantile
• nanquantile
Averages and variances
• median
• average
• mean
• nanmedian
• nanmean
• nanstd
• nanvar
Correlating
• corrcoef
• correlate
• cov
Histograms
• histogram
• histogram2d
• histogramdd
• bincount
• histogram_bin_edges
• digitize
A few things to discuss:
- Why is
dtypenot being dispatched instdandvar? Other multimethods that use_reduce_argreplaceralso don't dispatchdtype. - The previously added
minandmaxseem to be equal toaminandamaxthat this PR intends to add:
In [2]: onp.amin
Out[2]: <function numpy.amin(a, axis=None, out=None, keepdims=<no value>, initial=<no value>, where=<no value>)>
In [3]: onp.min
Out[3]: <function numpy.amin(a, axis=None, out=None, keepdims=<no value>, initial=<no value>, where=<no value>)>- I'm looking into implementing some defaults as well. Multimethods like
medianthat reduce array slices along an axis might be the easier ones for now. As I understand these require a for loop over the given axis, so in terms of complexity it would be O(n) where n is the length of that dimension.
*2 new commits pushed to
master*Quansight-Labs/unumpy@76a279f - Misc fixes found by the cupy backend
Quansight-Labs/unumpy@9a42b49 - Merge pull request #71 from Quansight-Labs/cupy-fixes
Pull request opened by joaosferreira
Quansight-Labs/unumpy#72
This fixes some issues regarding the correct dispatching of methods and classes like the one mentioned in #67 (review). It also allows proper dispatching of methods in modules such as
unumpy.linalg as demonstrated by the fixed tests.
*3 new commits pushed to
master*Quansight-Labs/unumpy@a52e773 - Add _get_from_name_domain to __ua_function__
Quansight-Labs/unumpy@986b492 - Fix tests for Linear Algebra routines
Quansight-Labs/unumpy@8f1d962 - Merge pull request #72 from joaosferreira/fix-backends
*15 new commits pushed to
master*Quansight-Labs/unumpy@0a168c9 - Add multimethods for indexing routines
Quansight-Labs/unumpy@0981d47 - Add defaut implementation for ravel_multi_index
Quansight-Labs/unumpy@bda0ab9 - Add default implementation for unravel_index
Quansight-Labs/unumpy@a30f2d4 - Add order kwarg to ravel_multi_index's and unravel_index's defaults
Quansight-Labs/unumpy@810e580 - Add more multimethods for indexing routines
Quansight-Labs/unumpy@ea0e0fe - Add default implementation for indices
Quansight-Labs/unumpy@57cf4de - Add c_, r_ and s_
Quansight-Labs/unumpy@b1e875e - Add lib.Arrayterator
Quansight-Labs/unumpy@eaefc2d - Add default implementation for ix_
Quansight-Labs/unumpy@5c46c50 - Add overriding metaclass for Arrayterator
Quansight-Labs/unumpy@3413b30 - Add dispatching of choices to choose
Quansight-Labs/unumpy@e7aa4f9 - Add default implementation for putmask
Quansight-Labs/unumpy@3833cc6 - Add backend fixes in #71 and #72 by merging branch 'master'
Quansight-Labs/unumpy@95321f9 - Add CClass, RClass and IndexExpression
Quansight-Labs/unumpy@f5a1d62 - Merge pull request #67 from joaosferreira/indexing-routines
[Unknown, Quansight]
Pull request opened by peterbell10
Quansight-Labs/uarray#247
Should fix the segfaults in Quansight-Labs/udiff#31
Returning selected_backend.get() instead of .release() from determine_backend meant the refcount was decreased without dropping the reference. The net result is decreasing the refcount of the backend once for each call to determine_backend. Eventually it gets GC'd and the python interpreter accesses invalid memory causing the segfault.
The GC issues are unrelated but were spotted while debugging this. The context managers had tp_traverse defined but the GC will only use these if the type has the Py_TPFLAGS_HAVE_GC flag set.
Reviewers: @hameerabbasi
[Hameer Abbasi, Quansight] Hello, Kartik. Are you looking for GSoC? If so, the selections are already over and have been for some time.
If not, head on over to https://labs.quansight.org/blog/2019/07/uarray-update-api-changes-overhead-and-comparison-to-array_function/ and read the blog. Play around with it, explore, and come back to us with any specific questions you may have.
*2 new commits pushed to
master*Quansight-Labs/uarray@33c6558 - Remove MyPy linting.
Quansight-Labs/uarray@dc9b68f - Merge pull request #248 from hameerabbasi/remove-mypy
*2 new commits pushed to
master*Quansight-Labs/uarray@60deba9 - Fix refcount and GC issues
Quansight-Labs/uarray@2e03d53 - Merge pull request #247 from Quansight-Labs/refcount-gc-errors
*1 new commit pushed to
master*Quansight-Labs/udiff@9cd95d9 - add np.stack (#31)
[Unknown, Quansight]
Issue opened by hameerabbasi
Quansight-Labs/udiff#32
@sangyx I noticed quite a bit of comments in Quansight-Labs/udiff#31 that contained commented out code.
If it's code for functionality you're working on, it's best to put it in a branch, that you'll merge later. If it's useless code, I'd just delete it.
Assignees: sangyx
[Unknown, Quansight]
Pull request opened by joaosferreira
Quansight-Labs/unumpy#73
This picks up the work started in Quansight-Labs/unumpy#46 that added multimethods for random routines. Most of the multimethods added in the previous PR where revised and hopefully some corrections were made. These were mostly changing the argument replacers and marking arguments for dispatching. This PR also adds two classes, RandomState and Generator. The multimethods added are manifold and are listed below:
Seeding and State
• get_state
• set_state
• seed
Simple random data
• rand
• randn
• randint
• random_integers
• random_sample
• choice
• bytes
Permutations
• shuffle
• permutation
Distributions
• beta
• binomial
• chisquare
• dirichlet
• exponential
• f
• gamma
• geometric
• gumbel
• hypergeometric
• laplace
• logistic
• lognormal
• logseries
• multinomial
• multivariate_normal
• negative_binomial
• noncentral_chisquare
• noncentral_f
• normal
• pareto
• poisson
• power
• rayleigh
• standard_cauchy
• standard_exponential
• standard_gamma
• standard_normal
• standard_t
• triangular
• uniform
• vonmises
• wald
• weibull
• zipf
Notes:
- Some aliases not included in the above list were also added, they are
random,ranfandsample. Although they are documented as aliases forrandom_samplethey all reference different objects and so they have their own multimethods. - The
Generatorclass is commented out in the tests because I'm not entirely sure on what argument I should pass it. - Most of CuPy's random methods should be working since its API follows NumPy's closely but we need to check the tests (some arguments might need to be unmarked).
Pull request opened by sangyx
Quansight-Labs/udiff#33
Clean code as Quansight-Labs/udiff#32
*2 new commits pushed to
master*Quansight-Labs/udiff@f800c34 - clean code & move unimplement code to dev
Quansight-Labs/udiff@938dc13 - Merge pull request #33 from sangyx/master
*12 new commits pushed to
master*Quansight-Labs/unumpy@c2b61fb - Add multimethods for statistical functions
Quansight-Labs/unumpy@7275f20 - Add default implementation for median
Quansight-Labs/unumpy@c7149fe - Remove _reduce_argreplacer
Quansight-Labs/unumpy@586a9e1 - Refactor default implementation for median
Quansight-Labs/unumpy@ddde1fd - Add default implementation for mean
Quansight-Labs/unumpy@764d78f - Add helper function for reduction methods
Quansight-Labs/unumpy@3b39aa7 - Add more default implementations for reduction methods
Quansight-Labs/unumpy@3059246 - Fix some default implementations
Quansight-Labs/unumpy@8fa8b52 - Merge branch 'master' and resolve conflicts with #67
Quansight-Labs/unumpy@3f10876 - Remove _reduce from most default implementations
Quansight-Labs/unumpy@f35a274 - Refactor median's default in terms of partition
Quansight-Labs/unumpy@1607d9b - Merge pull request #70 from joaosferreira/statistical-functions
*4 new commits pushed to
master*Quansight-Labs/udiff@5b4157e - implement jvp
Quansight-Labs/udiff@84d35ff - clean code
Quansight-Labs/udiff@9f7e3b5 - add test cases
Quansight-Labs/udiff@ffe037e - Merge pull request #34 from sangyx/master
[Unknown, Quansight]
Pull request opened by sangyx
Quansight-Labs/udiff#35
Hi, I have updated the docs. While I found my setting doctest_global_setup in docs/conf.py does not work. Could you help check? Many thanks!
Besides, I want to change the diff function form of JVP from lambda g, ans, x, y: to lambda ans, x, y: lambda g: , so as to separate the calculation of the result from the derivation process and solve the limitations of grad_variables, jacobian, and higher derivative. But this may not be done before September, so I will submit the GSoC result document first.
*5 new commits pushed to
master*Quansight-Labs/unumpy@e079d01 - Add multimethods for random routines
Quansight-Labs/unumpy@d93a45e - Add default_rng, BitGenerator and SeedSequence
Quansight-Labs/unumpy@9da7501 - Add multimethods for Generator's methods
Quansight-Labs/unumpy@ba6a968 - Add further tests to RandomState and Generator
Quansight-Labs/unumpy@5720e11 - Merge pull request #73 from joaosferreira/random-routines
*9 new commits pushed to
master*Quansight-Labs/udiff@73c3624 - split DIffArray into VJPDiffArray and JVPDiffArray
Quansight-Labs/udiff@1726b0d - update docs
Quansight-Labs/udiff@a4619fa - fix build docs
Quansight-Labs/udiff@e87aae6 - fix lint docs
Quansight-Labs/udiff@350b073 - add namespaces
Quansight-Labs/udiff@cd518f6 - remove unused namespaces
Quansight-Labs/udiff@56751d7 - update vjp doctest
Quansight-Labs/udiff@b4708c5 - fix broadcast
Quansight-Labs/udiff@0e40adb - Merge pull request #35 from sangyx/master