Ecosyste.ms: Awesome

An open API service indexing awesome lists of open source software.

Awesome Lists | Featured Topics | Projects

https://github.com/esheldon/great3-config

config files for running on great3
https://github.com/esheldon/great3-config

Last synced: 2 days ago
JSON representation

config files for running on great3

Host: GitHub
URL: https://github.com/esheldon/great3-config
Owner: esheldon
Created: 2014-03-12T05:31:43.000Z (over 10 years ago)
Default Branch: master
Last Pushed: 2015-06-03T16:25:45.000Z (over 9 years ago)
Last Synced: 2024-10-19T21:24:37.769Z (19 days ago)
Language: Shell
Size: 363 KB
Stars: 0
Watchers: 2
Forks: 0
Open Issues: 0
Metadata Files:
- Readme: README.md

Awesome Lists containing this project

README

great3-config
=============

config files for running on great3

great3 run03
------------

Same as run02, but putting ring pairs together to help selections

rgc
====

- TODO
- currently using great-des prior, but should adapt to the galaxies
we have or try BA?

- maybe should use fracdev from the brightest sample? Maybe deep fields
still too noisy? Hmm... most between 200 and 1000 Maybe fit
from s/n > 400? Or limit on noise on fracdev 0.1?

- limiting error on fracdev makes the dist look quite different
even for the deep data

- maybe use TF-fracdev prior?
- other cuts?

- deep runs

- run-g303-rgc-deep01
- cm
- run-g303-rgc-deep02
- exp and dev for priors (not used yet)
- run-g303-rgc-deep03
- clipping fracdev [0,1]
- T distribution radically different

- regular runs

- run-g303-rgc-01
- F-fracdev prior
- not using separate priors from exp and dev
- looks good, but note accidentally used prior from g302 deep07!

- run-g303-rgc-02
- fracdev prior not flux dependent. Used correct prior!

- run-g303-rgc-03
- same as 01 with F-fracdev but now with deep01 prior
- this should not matter much
- also added continuous fracdev fitting

- run-g303-rgc-04
- same as 03 but now with exp/dev priors separate
- no better

- run-g303-rgc-05
- using TF prior from deep03, which clipped the fracdev
- Q,m looks good at low s/n c worse
- worse at high s/n

- run-g303-rgc-06
- using new F-fracdev-02 prior, in which I used a tighter
cut on fracdev error
- marginally better at high s/n

- run-g303-rgc-07
- same as 06 but BA prior, 0.3
- crap!

- run-g303-rgc-08
- erf2 prior
- not quite as good

- run-g303-rgc-09
- 1000 isamples
- looks pretty much the same as 06

- run-g303-rgc-10
- 500 isamples

- run-g303-rgc-11
- flat priors
- 500 isamples
- round s/n included

- run-g303-rgc-12
- 500,500 isamples
- round s/n included

- run-g303-rgc-13
- turb to see if round measure works better
- 500,500 isamples

- run-g303-rgc-e01
- an exp-only run

great3 run02
------------

Used psf parameters from Aaron and Mike

- run-g302-rgc-01
- cuts-sfit-01 (shear noise weighted)

Q_c = 172.2700
c+ = -0.00069 +/- 0.00008
cx = -0.00000 +/- 0.00006
m+ = +0.02172 +/- 0.00325
mx = +0.01682 +/- 0.00242

- cuts-sfit-02 (Ts2n weighted)

Q_c = 487.4262
c+ = -0.00060 +/- 0.00008
cx = +0.00003 +/- 0.00005
m+ = -0.00347 +/- 0.00296
mx = -0.00616 +/- 0.00202

# using 3 gauss psf
- run-g302-rgc-01
- looks better

# composite fit, priors from g302-rgc-deep02
- run-g302-rgc-03
- looks very bad, m=0.05 for both components. Must be a bug
- note no need to iterate in isample, just use [2000]

- could it be using BA as prior? In nsim run it didn't matter,
even with low s/n, although these were fairly large galaxies
- maybe try great-des style dist
- maybe run exp+dev sim for smaller sizes.
- Note dev T should be not equal to exp to be realistic!

# composite fit, priors from g302-rgc-deep02 with g fit prior
# using GreatDES g prior (new parameters)
- run-g302-rgc-04
- clip fracdev [0,1]

- about same with default cuts

- interestingly, s/n cut helps

for s/n > 40 I get Q=234 m ~ 0.015
for s/n > 45 I get Q=379 m ~ 0.009
for s/n > 50 I get Q=1030 m ~ 0.002
for s/n > 55 I get Q=668 m ~ -0.004
for s/n > 60 I get Q=405 m ~ -0.009

so there is some noise for sure here, hard to know if we should
believe this cut at exactly 50 is meaningful

- run-g302-rgc-05
- same as 04, just added saving some output

- run-g302-rgc-06
- computed prior from deep05, unclipped
- does great at low s/n now, but crappy
at high s/n
- so no clipping at high s/n is actually bad, or perhaps
the underlying reason just manifests that way
- prior without clipping is great at low s/n

- run-g302-rgc-12
- no prior but also no clip, not very good at any
s/n

- the following runs up to 12 explore various things, but
the conclusion is that the prior (unclipped) from the deep fields is
great at low s/n but not high. Maybe the prior itself is just crap
there, not sure, but have seen that doing the fit and clipping at high
s/n works well, and applying prior without clipping works great at low
s/n. Is that just particular to these fields?

Q for various parameters

all s/n high s/n (>50)
------------------------------
05 no prior, clipped | 230 | 1030
12 no prior | 340 | 160
13 prior, clipped | 120 | 750
06 prior | 600 | 130

- things to explore

- try cutting chi2per
- try using combined g_sens

- maybe the clipping of just high fracdev would help?
- prior clipped [-1.5,1.1]?

- maybe *not* fitting that spike at fracdev=0 is causing the problem
at high s/n?
- mike's plots show the fracdev should not evolve much, so
maybe confirms something is fishy?

- maybe cgc was crap *because* it fit the spike?

- or maybe *fitting* spike at fracdev=1 is a problem?
- it always gets fit, no way to remove it sensibly

- should have priors for exp and dev separately, not the one
from cm...! Maybe this is messing up the high s/n?

- ruled out
- best of exp/dev
- clipping to [0,1] after the fact with prior applied
- separate priors for different s/n cut. I tried s/n > 50 with new
prior based on psf flux cut and it looked like crap

- run-g302-rgc-13
- use prior but clip fracdev [0,1]
- still works quite well at high s/n, crap at low s/n

- run-g302-rgc-14
- clip [0,1] only for s/n > 35
- in matrix above, this would perform most consistently
- seems incorrect tough, as if it would not work on real
data
- worry it is some reflection on sample variance in
great3? Should create more low noise images
to reduce sample variance in prior?

- works pretty well only for high s/n... how can this be
consistent with 06? somehow breaking the ring test?

- run-g302-rgc-15
- realized I needed to update the TF prior when not
clipping!

- run-g302-rgc-16
- try to choose exp or dev, with prior

- run-g302-rgc-17
- same as 06 but using BA for g prior
- looks like crap!
- the only change I recall is resetting the center

cgc

- run-g302-cgc-deep01
- run-g302-cgc-01
- looks like crap!

run-g302-rgc-deep03
- another cm run but this time recording fracdev
run-g302-rgc-deep04
- another cm run this time recording unclipped fracdev
run-g302-rgc-deep05
- another cm run this time *using* unclipped fracdev

great3 run01
------------

rgc:
- run-g301-rgc-01, first run
- forgot to record efficiency
- using single psf

old stuff
---------
rgc

- rgc-deep01 expfit
- real\_galaxy ground constant
- run-rgc-01 edg
- run-rgc-02 edg i2 using 01 as expand shear, some minor bug fixes
- run-rgc-04 edg i3 using 02 as expand shear

- run-rgc-05 edg i3 using 04 as expand shear, adding psf info, and newer
fits to priors. lower score, 270

- run-rgc-06 similar to 04 but using g prior from lackner
- score 412. Ran another iteration from this and got 400; maybe noise
is about +/- 10

- rgc-deep02 bdf
- rgc-03 bdf trained from deep02

- rgc-deep03 exp
* exp with psf ngauss 3 more burnin slightly tighter flat priors run-rgc-05
* edg i4 using 04 as expand shear and deep03 for priors

- rgc-deep04 exp
- using cosmos g prior, all the rest is the same

run-cgc
---------
control ground constant

- cgc-deep01 sersic