Project

General

Profile

Ideas Bulletin Board » History » Version 4

Version 3 (Andrew Vold, 01/23/2018 10:57 PM) → Version 4/5 (Andrew Vold, 01/23/2018 10:57 PM)

h1. Progress and Ideas Bulletin Board

h2. Cosmic rejection

* Preselection... use cosmic veto? Something different?
* Train on preselected (more challenging) cosmics
* Just try running current model and see which cosmics even pass
** Make pixelmap files for a bunch of cosmics on the grid.

h2. Software

* UPS products will soon go live on bluearc
* ANG updates for Caffe model analysis need to be committed
* We can soon deploy on grid
** NB: we'll have to be careful about how we ship model deploy files, they're too big for regular @cp@

* Tidy-up:
** We want to fix a few things before the next round of training
*** Calibration... PECorr -> GeV...
*** Vertex location/noise scrubbing... turn this off
** Do these things when the new files come out, or before we generate the next pixelmap files

* Implement ROOT2LevelDB in novasoft context

* Longer term
** A nicer, generic, labeling infrastructure might be nice
** This would make it easier to play with things like topology labeling (1-pi, 2-pi, etc.)

h2. Single purpose classifier models

* Fine-tuned models are attractive for analysis groups
* Numu already liked theirs
* Alex R. is working on a nue classifier... maybe NC too? Or was Adam going to do that?

h2. Systematics

* Bad channels
** Probably a tiny effect... but enhancing the actual effect might be good for regularization
** Can we do this with dropout? Does Caffe let you do it in a sensible way for input layers?
* Jitter
** Another good regularization technique
* Attenuation slopes, shifts
** Can we just slip in shifted samples?
** Or, is there a fancy way to make a custom layer which does this randomly for each training example?

h2. Prong classification

* Play with view merging in architecture... should it happen sooner?
* What size is appropriate

h2. Multimodal inputs

* Can we pass in other reconstructed information with the images?
** e.g. location, track length, etc.
** Some links:
*** Caffe says you can do it here: http://caffe.berkeleyvision.org/tutorial/data.html
*** Maybe a code snippet in here: https://github.com/karpathy/neuraltalk



h2. Minimalist ROOT I/O in pure Python and Numpy

* Read in branches from TTrees into numpy arrays without needing ROOT
** Github link: https://github.com/scikit-hep/uproot
** [[Sample code]] Sample code