Add randn for arrays of standard gaussian distributed random numbers. #146

daniel-vainsencher · 2016-03-12T20:10:30Z

This time trying to rebase, rather than merge, let me know if I've done it correctly.

We don't want to compile crate ndarray twice for benchmarks and tests; only once. Tests and benchmarks are outside of the crate.

The new ShapeError is a struct, so that we can stuff more information into it without changing the API.

A .push() loop is not efficient enough; we can do this manually using our knowledge that we will produce exactly as many elements as was preallocated for.

The former will optimize out in some cases where the other won't. Improves `dot` bench in fact.

`Debug` enables much better error messages. We control all implementors, so this is a very safe change.

We're in quick development, so we are sticking with the latest stable rust.

Previous for loop used for the epilogue expanded into an unrolled loop, even if it's a very short loop of 0 to maximally 7 elements.

These should just cut the loop one index earlier. No impact on the result.

…inputs

bluss · 2016-03-12T20:53:23Z

Thanks. I'm not promising that we will integrate rand directly like this. It could for example belong to a crate of its own "ndarray-rand" or so.
Something is definitely amiss when there are so many commits listed in the PR. Would be great if you could reset your branch to just contain one patch on top of master and force push again.

daniel-vainsencher · 2016-03-12T21:24:17Z

This is just a first cut. I agree we need to get the dependencies and design right. And I'll have no choice but to read up on git rebasing... out of time for now, will get back to this, as I need it. BTW, is there a good alternative to depending on rand?

bluss · 2016-03-12T21:30:57Z

It's just one commit, so you can use cherry-pick to move it to the correct branch. It's simple, like this #124 (comment)

I don't know a better crate than rand.

rand could be an optional dependency, but even better than that (also what crates.io advises) is to make a separate crate if the feature is easy to split out.

A new crate ndarray-rand could be hosted by me, it would depend on ndarray and rand, and it provides these constructors.

bluss · 2016-03-12T21:47:39Z

Maybe using rand for the gaussian samples is the wrong thing? I wonder if we can do this much more efficiently if we're generating a big set.

daniel-vainsencher · 2016-03-13T04:46:56Z

I am not sure what is the significant advantage of avoiding rand as a dependency of ndarray.

If we create our own code that takes advantage of making a whole array of them at a time, that's fine, but currently a hypothetical.
Otherwise random numbers are not esoteric in numerical computing, they are a fundamental tool used all the time. randn is one of my most common numpy imports in python.
While cargo is good enough that having users add another crate is not difficult, this only works if they are aware it exists; discovery is still made harder. I haven't thought through any ergonomic differences for users of randn.

daniel-vainsencher · 2016-03-13T05:10:17Z

Created a new pull request with clean history and addressing the points above.

bluss and others added 30 commits March 12, 2016 15:08

Simplify arr2, arr3

8b6101a

Bump to 0.4.0-alpha.7

5017e3b

Add forward compat note to the section with .iadd() and similar methods

899524b

Disable bench, test for main library

c951776

We don't want to compile crate ndarray twice for benchmarks and tests; only once. Tests and benchmarks are outside of the crate.

Add more benchmarks using dot

b0b2a72

Move errors into a common file

ccd23ef

Merge StrideError and ShapeError

0ef91dc

The new ShapeError is a struct, so that we can stuff more information into it without changing the API.

Rename two error kinds

fe100b6

Implement better Debug for ShapeError

02d812c

Rm repr() for ErrorKind

cca093c

Edit docs

138d5d9

Make .map() autovectorizable

ccb2d76

A .push() loop is not efficient enough; we can do this manually using our knowledge that we will produce exactly as many elements as was preallocated for.

Edit docs for constructors

616a3ce

Edit docs for data traits

4e679f7

Edit docs for traits

c0a98d2

Use assert!() instead of assert_eq!() in .dot()

584e811

The former will optimize out in some cases where the other won't. Improves `dot` bench in fact.

Use += in examples/life.rs

3845d74

Fix printout in convo.rs to be padded better

9602cb8

Require Debug for Dimension trait

8eeee5a

`Debug` enables much better error messages. We control all implementors, so this is a very safe change.

Print a more detailed index out of bounds message in debug build

e8d801d

arrayformat: simplify closure passing

6281a6a

Enable assign_ops by default in Rust 1.8

54e33d5

Update readme to say we require Rust 1.7

d170071

We're in quick development, so we are sticking with the latest stable rust.

Update doc for assign_ops

7783abf

enable indexing Dimension by Axis

60ddbcb

make dimension-indexing methods belong to the trait and doc-hidden

031024f

Impl From for types that can be converted to array views

c5fa8a5

Back out ArrayView::from_slice in favour of ArrayView::from

6c5e45c

Add convenience trait AsArray

bfdd135

Make ArrayView::from and friends visible in the documentation

86e7822

bluss and others added 22 commits March 12, 2016 15:08

numeric_util: Simplify epilogue in unrolled_sum

4ddb8e7

Previous for loop used for the epilogue expanded into an unrolled loop, even if it's a very short loop of 0 to maximally 7 elements.

numeric_util: Simplify epilogue in unrolled_dot

1bd6e75

Fix off-by-ones in numeric_util

27c7666

These should just cut the loop one index earlier. No impact on the result.

Add test for zeros_f and alternate memory order

6ac9e0f

Add shape & stride info to Debug for arrays

c1985fc

Add detailed matrix multiply benchmarks

f000d11

Add test for matrix multiplication vs memory order

ee91de1

Use blas-sys directly for BLAS integration in mat_mul and dot

bc9c6d3

General matrix multiply also returns f-order result from two f-order …

94bb33b

…inputs

Add test for mat_mul return value's memory order

7af8160

Use openblas in travis build

dbccf39

Remove unused helpers from linalg.rs

ce07e7f

Add comment for gemm

afc5736

Rename c_int -> blas_index

794e019

Update docs Makefile

b95990e

Put a deprecation notice on rblas -- it will move to its own crate

b879cd7

Create subdirectory for ndarray-rblas

d164cfb

Add getters .as_ptr(), .as_mut_ptr() to ArrayBase

39af789

Add ShapeError constructor

fac059e

Deprecate ndarray::blas in favour of ndarray-rblas crate

8ec2f25

Fix ndarray-rblas license and description

631bcd5

Add randn to create an array of standard gaussian numbers.

65bbc61

daniel-vainsencher closed this Mar 13, 2016

bluss mentioned this pull request Mar 13, 2016

Add randn to create an array of standard gaussian numbers. #147

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Add randn for arrays of standard gaussian distributed random numbers. #146

Add randn for arrays of standard gaussian distributed random numbers. #146

Uh oh!

daniel-vainsencher commented Mar 12, 2016

Uh oh!

bluss commented Mar 12, 2016

Uh oh!

daniel-vainsencher commented Mar 12, 2016

Uh oh!

bluss commented Mar 12, 2016

Uh oh!

bluss commented Mar 12, 2016

Uh oh!

daniel-vainsencher commented Mar 13, 2016

Uh oh!

daniel-vainsencher commented Mar 13, 2016

Uh oh!

Uh oh!

Add randn for arrays of standard gaussian distributed random numbers. #146

Add randn for arrays of standard gaussian distributed random numbers. #146

Uh oh!

Conversation

daniel-vainsencher commented Mar 12, 2016

Uh oh!

bluss commented Mar 12, 2016

Uh oh!

daniel-vainsencher commented Mar 12, 2016

Uh oh!

bluss commented Mar 12, 2016

Uh oh!

bluss commented Mar 12, 2016

Uh oh!

daniel-vainsencher commented Mar 13, 2016

Uh oh!

daniel-vainsencher commented Mar 13, 2016

Uh oh!

Uh oh!