Where communities thrive


  • Join over 1.5M+ people
  • Join over 100K+ communities
  • Free without limits
  • Create your own community
People
Activity
    Francois Lanusse
    @EiffL
    In particular we discuss authorship and membership
    Francois Lanusse
    @EiffL

    @/all We haven't had a lot of activity in the last two months but I'd like to pick things up where we left off. I've had recent discussions with several people (including @tilmantroester @modichirag @liuxx479 ) about baryonification of simulations, and I think there are several avenues for us to do some cool things. So I went ahead and created a Baryonification project and channel ( https://gitter.im/DifferentiableUniverseInitiative/Baryonification ) feel free to join to see what we are thinking about.

    I also think it might be time to have a first telecon to catch up on what everybody is doing, and discuss projects (including this one ^^^) Here is a poll to find a good timeslot:
    http://whenisgood.net/dui (results will be here: http://whenisgood.net/dui/results/cbngmre )

    And most importantly, I hope everybody is ok and coping well with the situation!

    Francois Lanusse
    @EiffL
    Anyone else wants to join this call ^^^ Looks like we are gearing up for this week, Wednesday, Thursday or Friday at 10am PST, 5pm UK, 6pm CEST
    Francois Lanusse
    @EiffL

    Alright @/all we'll have our first DUI telecon tomorrow Friday, at 10 am PST, 5pm UK, 6pm CEST. All are most welcome to attend, I'll post a zoom link here.
    Here is a slide deck if you want to have some slide to support the discussion, and keep a trace of the meeting: https://docs.google.com/presentation/d/16gHb4arCCks6Qch4fn4NHLTYpKXuom9SoI2Ypx675Cg/edit?usp=sharing

    On the agenda, we have:

    • Structure of the initiative, discussing the proposed terms and conditions, and ratification process
    • Individual updates on projects related to the initiatives (FlowPM, Differentiable HOD, ...)
    • Discussion on Baryonification
      But feel free to modify the slides and edit the agenda if you want to talk about something else.
    Francois Lanusse
    @EiffL

    And here is the connection information:

    Topic: DUI Meeting
    Time: Mar 27, 2020 06:00 PM Paris

    Join Zoom Meeting
    https://us04web.zoom.us/j/573671523

    Meeting ID: 573 671 523

    Francois Lanusse
    @EiffL
    @/all new link
    ChangHoon Hahn
    @changhoonhahn
    Francois is brutal about keeping time :P
    Ben Horowitz
    @bhorowitz
    Any plans for another meeting? Sad I missed the last one! I assume HyPhy was one of the things discussed (or to-be-discussed) under the "Baryonification" heading. I've also been working on a photo-z reconstruction framework that probably be of interest to those in DUI.
    Francois Lanusse
    @EiffL
    We haven't planned one yet, but I think we could do another one this Friday, it will have been 2 weeks since the last one. And I think we still had things to discuss last time ^^'
    Would that work for you Ben?
    Ben Horowitz
    @bhorowitz
    Yep!
    Francois Lanusse
    @EiffL
    Sweet, ok so I will give @/all a chance to raise their hand if they absolutely want to attend this telecon but can't make it, and let's schedule this next telecon on Friday April 10th, 9am PST, noon EST, 5 pm UK, 6pm CEST.
    On the agenda, we can continue the discussion on baryonification, get an update on HyPhy, and I really want to hear more about photo-z reconstruction :-D . I can also give an update on the good progress I made on the DHOD project. Any other updates, or discussion items welcome ;-)
    Francois Lanusse
    @EiffL
    @/all reminder that if you are free today, feel free to join the telecon at 9am PST, noon EST, 5 pm UK, 6pm CEST If there are plots you'd like to show I have setup a slide deck here: https://docs.google.com/presentation/d/16gHb4arCCks6Qch4fn4NHLTYpKXuom9SoI2Ypx675Cg/edit?usp=sharing
    On the agenda we will talk about differentiable HOD, Ben's HyPhy , Baryonification, and more. Feel free to add any topics you want to the agenda slide.
    Francois Lanusse
    @EiffL
    And here is the connection information:
    Francois Lanusse is inviting you to a scheduled Zoom meeting.
    Topic: DUI meeting
    Time: Apr 10, 2020 06:00 PM Paris
    Join Zoom Meeting
    https://us02web.zoom.us/j/741953111?pwd=M3RRUnFkaHJjUThWcFZ6RmVOUVRUZz09
    Francois Lanusse
    @EiffL
    @bhorowitz this photoz stuff is so cool, can you work out with the powers that be, if part of that project can be worked on in the open with the people here?
    Francois Lanusse
    @EiffL
    Dear @/all by popular demand we will have a telecon tomorrow Friday 24th at 9am PST, noon EST, 5 pm UK, 6pm CEST. We have a few things to talk about, I know I can give an update on the DHOD project, but if there are other things you would like to discuss, please add them to the agenda in the slides below:
    https://docs.google.com/presentation/d/16gHb4arCCks6Qch4fn4NHLTYpKXuom9SoI2Ypx675Cg/edit?usp=sharing
    I'll share the agenda again here tomorrow, and with any time we have left after discussions, we can do some live hacking ;-)
    Francois Lanusse
    @EiffL

    Hi everyone, here the connection information for ~1h from now:

    Topic: DUI Zoom Meeting
    Time: Apr 24, 2020 06:00 PM Paris
    
    Join Zoom Meeting
    https://us02web.zoom.us/j/82579967155?pwd=UDRBT1RvYVpIWkZ1bDI0aXBCeXlvdz09

    On the agenda, we'll talk abou the following things:

    • DHOD: It's working!
    • FlowPM and differentiable halo finding
    • Photo-z reconstruction
    • Plus whatever else you want to talk about

    See you soon

    ChangHoon Hahn
    @changhoonhahn
    Tabulated correlation function implemented by Johannes Lange: https://github.com/johannesulf/TabCorr
    Might be a quick temporary hack for differentiable 2pcf.
    Francois Lanusse
    @EiffL
    Dear all, I've had countless discussions about whether anyone would ever take the time to write a differentiable cosmology library and finally free us from the tyranny of MCMCs and/or emulators ^^' So becasuse we have a 3 day weekend here in France, I'm gonna go ahead and take a stab at it: https://github.com/DifferentiableUniverseInitiative/jax_cosmo
    It's gonna be a prototype just to see how painful it would actually be, and also to try out Jax. My goal modest goal is to be able to compute lensing Cls with simple Limber and Eisenstein & Hu power spectrum, that shouldn't be too hard.... If that sounds like fun to you, feel free to follow along :-) I'll be documenting things on issues
    Francois Lanusse
    @EiffL
    download (4).png
    and boom, looks like pure numpy code, but is differentiable and runs on GPU. Here is a toy demo where I fit a linear power spectrum for Omega_m
    Francois Lanusse
    @EiffL

    Hi everyone! Just wanted to share an update on the project mentioned above.... It works, and it's awesome! I have pushed a first release candidate on PyPi, but the API is still pretty much open for debate, and could use with some help to push it further.

    It can do basic 3x2pt right now, but there are many things that can be added like non-linear PS, or any differentiable emulator for any observable ( ;-) @aphearin @nesar )...
    Check it out here: https://github.com/DifferentiableUniverseInitiative/jax_cosmo

    Andrew Hearin
    @aphearin
    Nice!
    Got plans for a demo example where you highlight something that is difficult with standard methods that is improved with differentiability?
    toy example of how you get a Fisher matrix like that:
    F = - jax.hessian(likelihood)(fiducial_cosmology)
    Andrew Hearin
    @aphearin
    Right this is slick one-liner syntax for Fisher calculations, but I meant an interesting physics problem that is now made easy by the differentiability.
    I have been thinking the main big win in practical terms comes from improved ability to handle higher dimensions, so maybe a 3x2 model with a lot of nuisance params?
    Francois Lanusse
    @EiffL
    Andrew Hearin
    @aphearin
    Aha! Yes, that's an exciting example to use as demo
    Francois Lanusse
    @EiffL
    but I mean, yeah, in general, this can allow for efficient inference. And a whole bunch of survey parametrisation and optimization things.
    also.... just having a nice library in plain Python but still super fast is a big plus.
    Nesar Ramachandra
    @nesar
    This is really cool! @EiffL. Yep Autodiff by Jax makes things so much nicer when differentiations are involved.
    Nesar Ramachandra
    @nesar
    Adding to that fisher matrix bit. Stability of numerical derivatives is one problem that is usually encountered with standard fisher calculations. Autodiff resolves that bit.
    Andrew Hearin
    @aphearin
    Good point about numerical stable derivs, @nesar.
    The differentiability is the new thing here. Not the fast python implementation (colossus is pure python and very fast) or the slick one-liner (which ofc is possible without differentiability). I think that's why the photo-z example is so nice.
    Francois Lanusse
    @EiffL

    Hi everyone, made some progress on this JAX library and thought I would share :-) I have now reimplemented everything needed for the DES Y1 3x2pt in JAX, so everything is neatly differentiable. Background and angular power spectra are validated against the DESC Core Cosmology Library, so not only is it differentiable, but it gives the right results.

    Here are a few notebooks:

    I think it's all really cool so I'm gonna start a paper draft in that repo: https://github.com/DifferentiableUniverseInitiative/jax-cosmo-paper
    My plan is to compare to the cosmosis Y1 3-2pt and do like a Fisher comparison and MCMC chain versus HMC or VI. Maybe also a small survey optimization example, and linear compression with MOPED.

    The proof of concept is working but still could use with some help to implement additional tracers, real space correlation functions, and some accuracy/optimization improvements. Also very interested if anyone has a cool use case for differentiability that I haven't thought of and will welcome help and feedback on the paper draft :-)
    So feel free to join https://gitter.im/DifferentiableUniverseInitiative/jax_cosmo if you want to be kept in the loop

    Francois Lanusse
    @EiffL

    Hi @/all we haven't had a telecon in a while, let's catch up with a telecon tomorrow Friday 29th at 9am PST, noon EST, 5 pm UK, 6pm CEST Hopefully you can make it :-) but if you can't and there are any topics you want to discuss feel free to drop a message here or add a slide.

    Here is the slide deck: https://docs.google.com/presentation/d/16gHb4arCCks6Qch4fn4NHLTYpKXuom9SoI2Ypx675Cg/edit?usp=sharing

    On the agenda so far I've put:

    • I'd like to discuss (again ^^') the DUI structure, so that we adopt some guidelines and keep growing. For reference I got jealous of this project this week: https://scikit-hep.org and discovered this awesome project, which seems to be the same thing as what we are doing, but in HEP: https://gradhep.github.io/center
    • jax-cosmo and associated paper
    • Maybe start a universal PS emulator project: Like a simple Python code that anyone can easily download and use, and that emulates accurately CLASS and CAMB. So that noone ever needs to use CAMB anymore

    But please add more :-D

    15 replies
    Nesar Ramachandra
    @nesar
    unfortunately I have a conflict at that time, attending a zoom phd defense.
    But definitely interested in a differentiable CAMB emulator. I’ve some experience from emulator we constructed with gpflow for Cls.
    Francois Lanusse
    @EiffL
    so @nesar with @marius311 we discussed yesterday the prospect of instead of buildiing yet another emulator, we could build an "emulation framework", that would provide a standard for how emulators should be specified
    I had a first pass at a design document for that here: https://github.com/DifferentiableUniverseInitiative/zelos/blob/master/design.md
    PR to make suggestions/comments are most welcome. In particular, before diving into this, I'm interested in understanding why CAMB is still a thing when technically emulators already exist. Marius rightly pointed me to a Python CAMB emulator code he has from 6 years ago for instance, but it's not really a standard today, sorry Marius.
    Nesar Ramachandra
    @nesar
    I agree that there are plenty of emulators out there, and a flexible emulator framework is something the cosmo community could use too (statisticians use R packages like DiceKriging (https://cran.r-project.org/web/packages/DiceKriging/index.html)
    but I belive CAMB remains relevent because that will ultimately remains to be the ground truth that all emulators are trained on. Moreover, CAMB and it’s branches are developed for so many cosmological models (dynamical Dark energy, modified gravity …), and you’d still need to run them all over again to train individual emulators.
    3 replies
    Francois Lanusse
    @EiffL
    Topic: DUI Zoom Meeting
    Time: May 29, 2020 06:00 PM Paris
    
    Join Zoom Meeting
    https://us02web.zoom.us/j/82857698931?pwd=d3lNeG8vMytUWUtNRDd6QkxlV1NOQT09
    In 1h30m :-)
    Nesar Ramachandra
    @nesar
    i like the name Zelos :D
    I agree with the points on the slide, it would be great to have a nice framework for emulation alone. And in addition a bunch of choices of emulation validations (local and global) would be great too.