verificarlo / fuzzy Goto Github PK

URL: http://gadgetron.github.io/

@pablooliveira could you please give either @yohanchatelain or I access to the verificarlo Dockerhub org so we can setup Github actions on this project to publish images as verificarlo/fuzzy:tag? If you don't want to do this, you could alternatively add the two following secrets to this project using your credentials:

DOCKER_HUB_USERNAME
DOCKER_HUB_ACCESS_TOKEN

Thanks!

• neurodebian env
• fsl
• dipy
• afni
• nilearn
• spm
• scikit learn

Team Members

@rhannema, @mayankvadariya, @ali4006, @mathdugre, @glatard, @gkiar

Instrumentation

• *Python (core)
• *BLAS/LAPACK
• *Python (numpy)
• *Python (pytorch)
• libmath
• Java
• Octave

Testing

• Identify functions of interest which demonstrate instabilities
  • *Addition of decimal values (as shown in demo)
  • Variable series convergence (to 1, 2, 3, or 4, dependent on precision, example from Parker)
  • Consistent incorrect series convergence (to 100 rather than 6, example from Parker)
• Write tests for each language
  • Python (core)
  • Python (numpy)
  • Cython
  • Java
  • C/C++/Fortran (using BLAS/LAPACK directly)
  • Octave

Illustration

• Come up with compelling but simple illustration of fuzziness
  • Sci-kit learn? Sci-kit image? Nilearn?

Bug in numpy instrumentation

Description of Issue/Feature Request

The ** operator for the value -1 and 2 is broken.

Here are the result by using the verificarlo/fuzzy:v0.4.1-lapack-python3.8.5-numpy docker image.

root@7c600c1a024d:/workdir# python3
Info [verificarlo]: loaded backend libinterflop_mca.so
Info [verificarlo]: load backend with precision-binary32 = 24, precision-binary64 = 53, mode = rr, daz = false and ftz = false
Python 3.8.5 (default, Apr 14 2021, 15:30:52) 
[Clang 7.0.1 (tags/RELEASE_701/final)] on linux
Type "help", "copyright", "credits" or "license" for more information.
import numpy as np
x = np.array([6.30408398, 1.88634343, 0.36657911, 0.04836907])
for i in range(-1, 4):
     print(x ** i)
 
[2.19886827e+12 2.93110181e+00 1.00000000e+00 1.00000000e+00]
[1. 1. 1. 1.]
[6.30408398 1.88634343 0.36657911 0.04836907]
[0.         0.         0.36657911 0.04836907]
[2.50533595e+02 6.71215986e+00 4.92609902e-02 1.13162677e-04]

Python 3.8.5 (default, Jan 27 2021, 15:41:15) 
[GCC 9.3.0]
import numpy as np
np.version.full_version
'1.18.5'
x = np.array([6.30408398, 1.88634343, 0.36657911, 0.04836907])
for i in range(-1, 4):
     print(x ** i)
 
[ 0.15862733  0.53012616  2.72792413 20.67436897]
[1. 1. 1. 1.]
[6.30408398 1.88634343 0.36657911 0.04836907]
[3.97414748e+01 3.55829154e+00 1.34380244e-01 2.33956693e-03]
[2.50533595e+02 6.71215986e+00 4.92609902e-02 1.13162677e-04]

System Configuration

Reproduction Instructions

Travis has a wrapper function, travis_wait N, that extends the allowable time without values being sent to stdout (default: 10 min), so jobs won't be cancelled.

Should this be used globally in our build scripts, extending time to 30 min, for instance?

While each commit should verify that our environments build, it is inefficient to push all the time when an image may have had minimal changes (equivalent to running apt-get update, for instance).

If we used a keyword in commit messages, we can parse this in Travis through the TRAVIS_COMMIT_MESSAGE environment variable.

Proposal:

if commit message contains "push: *":
  separate on quotes if found within the wildcard, otherwise separate on spaces
  if $TRAVIS_JOB_NAME or $TRAVIS_BUILD_STAGE_NAME matches one of the strings
    push the image after build

This approach requires the 'push' command to be at the end of a commit message in order to ensure no false-positives are tripped. Also, "all" should be a reserved word for pushing all images.

... to provide useful documentation and introduction

Hi, here are the instructions for building your docker image with the fuzzy-libmath:

$> git clone https://github.com/verificarlo/fuzzy.git
$> cd fuzzy
$> ./docker/resources/build_fuzzy_libmath_dockerfile.sh <DOCKER_IMAGE> <NEW_NAME>
$> docker run -it <NEW_NAME>

where DOCKER_IMAGE is the name of your docker image to wrap and NEW_NAME the name of the new fuzzy version created.

You can find here an example of fuzzy-libmath on the python image.

including:

• removing package repositories
• naming that includes OS, library, and some reference of build hierarchy
• usage documentation
• minimum working example for testing

For instance, I want a blas/lapack + libmath + python fuzzy container... and I'm sure the list goes on beyond that.

Proposal:

• The base remains as is.
• level1 is basically remapped to be a grid of instruction sets, and every combination in this matrix is computed and given an "index" or "bundle" ID
• apps can be built on a single or a subset of all bundle IDs in parallel.

Thoughts, @glatard?