HEP.TrackX Weekly

Friday 16 Sept 2016, 13:30 → 14:30 US/Pacific

vidyo

Paolo Calafiura, Steven Farrell

Please join the meeting by clicking this link:
https://vidyoportal.cern.ch/flex.html?roomdirect.html&key=UoHjbOgfcsrBsDfBwOlOsCObR3E 

If you want to join by phone, please use one of the phone numbers listed in the link below: http://information-technology.web.cern.ch/services/fe/howto/users-join-vidyo-meeting-phone and enter the meeting extension 103958871 in order to join.

Present: Giuseppe, Jim K, Maria, Paolo, Mayur, Steve, Pietro, Prabhat

News:

Paolo:

• we have submitted FWPs for LBL and FNAL (includes CalTech). We should receive all funds we asked for in our proposal in a single payment in FY16 (two weeks left).

• Early “technical” goals:

  • Name the project - currently x-cut-hep-tracking : (MS suggestion) HEP-TrkX  or HEP.TrkX  (X standing for exploration and cross-cut!) 23.09.2016

  • Common sw repository - currently bitbucket  → Maria to have a look and make a proposal -- Created repository https://github.com/cmscaltech/HEP.TrkX

    • (MS) Need everyone’s git account userid to add you

  • Shared dropbox/cernbox - currently google drive

    • Caltech box  (MS) made an are in Box (50GB) and added Paolo to test how this works

      • Prabhat: also for data?

      • Good question, would work up to small sizes. Also we need to address data curation (see our proposal).

        • → Jim K to propose solution for data

  • Project web page. Volunteers?

• • → Maria undergrad will look into this

    • (MS) Will have the web page ready this weekend

• Shared platform to run our codes?

  • Like lxplus, NERSC?(would need to apply for time this weekend), …

    • Prabhat recommends that Paolo put in an ERCAP allocation

      • 1M core hours on Cori Phase II and/or Edison

      • Call out DOE grant and sponsoring PM

      • Call out that you will utilize NERSC production ML/DL libraries

  • Prabhat: What software frameworks is the team planning on using? (Theano, Caffe, TensorFlow, CNTK, Torch)

    • Crucial point.

    • Also having the ability to scale-up to big learning apps is crucial

    • Suggest to put in a small NERSC ERCAP request this weekend.

  • Maria: helped Brazilian group setup early-access to Intel tools. Wonder if it is the same access NERSC has. (the “Early Adopter Program” to receive servers with advanced tech) -- In negotiation to find out more about this deal soon

    • Maria is preparing a system with 16 GTx1080 on SuperMicro servers-- ready in Oct and for SC16 (MS)

Kickoff Workshop:

• Two/three days before the end of the year

  • Proposed dates, not yet vetoed:

    • After CHEP:  Oct 17-19 (@LBL?)

      • Tue afternoon?

      • Wed afternoon during visit workshop

    • Before/after SC16(Nov 12-18),

      • Thursday 17 - Friday 18?

    • Standalone  (Caltech offered to host, FNAL would prefer to host).

(http://dolcit.cms.caltech.edu/scmls/ another ML workhosp at Caltech, Nov 18)
Some weeks that should work for Paolo:

• Oct 17

• Oct 24

• Nov 4-5 @ Caltech would for Maria, may help with travel restrictions.

  • Also at Fermi

• Nov 7 (I don’t get to vote…NOOOOO)

• Nov 28

• Dec 5 (if need be)

• Dec 12 (too late)

JBK cannot do Nov 4-5,  Nov 1-3 works, FNAL would be a great place for the workshop (MS lets do it at FNAL Nov 1-3!)

• Goals:

  • Get to know the team!

  • Review/update long-term goals listed in proposal also in view of reviewers feedback (when we get it!)

  • Define short-term goals (three-six months) in some detail. Probably include:

    • Generate both toy and “realistic” HL-LHC tracking data samples

    • Put together baseline and/or “state-of-the-art” solution(s) for seeding and tracking

      • Both could rely on aCTS infrastructure

      • Agree on quality metrics

    • Start experimenting with several ML techniques on toy samples.

    • Possible presentations/publications

  • Agree on software process (code management, testing, issue management)

  • Hackathons/teach-ins. For example:

    • aCTS simulation

    • Parallel CKF

    • NN training and testing with tensorflow (caffe, theano, you name it...)

  • Invite external collaborators (PCKF, aCTS)

Progress Reports:

Steve:

• ACTS workflow setup for producing track data of varying complexity (e.g., no mag field, fewer layers) in order to explore limitations of our RNN models

Mayur:

• Fitting LSTMs to tracks that are 1-Dimensional before moving on to 3D (LHC) data

• Generating various trajectories as shown below

• Understanding LSTM mechanics, limitations and features

• Mayur will link his working document on this

• Prabhat: is there LSTM expertise in our group

  • Maria: Jean Roch and students have done a project with LSTMs this summer.

  • Jim: mostly with summer students, using RNNs for signal processing.

Maria: Caltech will start in October. Will meet with Pietro and other to coord Caltech activities/situation. MS Have not met with Pietro and Yisong yet bc of DOE matters this week

Jim K: met with Lindsay and Giuseppe last week

Other news from Caltech (09.23.2016): one of MS grad students, Dustin Anderson has decided to become a data scientist  after graduation. He is at CERN and will get back to Caltech in Dec. He will start spending time on ML and he has already done some v serious technical work on setting up distributive NPI architecture with JR (who spoke with Jim abotu this). I foresee engagement in our project. Also Javier Duarte (former student who just defended)  has started as Lederman Fellow at FNAl and he is terrific -- we discussed he would join from FNAL side. On the side of Caltech I some have administrative issues but Panagiotis is helping.