rsnemmen / nmmn

Miscellaneous methods for Astronomy and Data Science. Array methods, statistical distributions, computing goodness-of-fit, numerical simulations and much more

About Rodrigo

Astrophysicist. Data scientist. Professor. Research Scientist. Team Leader. Curious person.

My website: rodrigonemmen.com
My Research Team’s website: blackholegroup.org

Trophy Case

Contact Me

• Email: [email protected] (this is an alias to minimize spam)
• LinkedIn: nemmen

Cool Stuff

TBD

Projects

• Awesome Project 1: A brief description of what this project does.
• Awesome Project 2: A brief description of what this project does.

Blog Posts

• Useful post 1
• Useful post 2

Badges

nmmn's Issues

Missing import evalfun in predbandnl

I was calculating some linear regression using BSCES and wanted to plot the prediction limits.
The module nmmn.stats.predbandnl raises a NameError exception saying that the function evalfun is not defined.

looking at the code under stats.py suggest that we are just missing a import section and import names

from . import misc

I have fixed this myself on my installation but this is just an FYI

Thanks for making these nice tools!

---

NameError Traceback (most recent call last)
in
23 lci_fb, uci_fb, xarray_fb = nmmn.stats.confbandnl(lir_fb, lx_fb, func, fitm_fb, covm_fb, 2, 0.95, xarray)
24
---> 25 lpi_sb, upi_sb, xarray_sb = nmmn.stats.predbandnl(lir_sb, lx_sb, func, fitm_sb, covm_sb, 2, 0.95, xarray)
26 #lpi_hb, upi_hb, xarray_hb = nmmn.stats.predbandnl(lir_hb, lx_hb, func, fitm_hb, covm_hb, 2, 0.95, xarray)
27 #lpi_fb, upi_fb, xarray_fb = nmmn.stats.predbandnl(lir_fb, lx_fb, func, fitm_fb, covm_fb, 2, 0.95, xarray)

~/.local/lib/python3.8/site-packages/nmmn/stats.py in predbandnl(xd, yd, fun, par, varcov, deg, conf, x)
540
541 # Residual sum of squares
--> 542 rss=residual(yd, evalfun(fun,xd,par) )
543
544 grad,p=[],[]

NameError: name 'evalfun' is not defined

Question about stats.confband vs confbandl

I have been using BCES in conjunction with nmmn for weighted linear regression and bootstrapping. It has been working fantastically so far. I have a question on usage between confband and confbandl. The documentation mentioned confband is for linear, while confbandl is nonlinear, the latter of which is used in the BCES example and modified for a weighted LR based on the function. I have been testing both stats modules, and I was surprised to see such a difference even when using BCES as my LR in both for the same dataset. Am I using one of these incorrectly? I originally thought to use stats.confbandl, but it seems from the documentation that stats.confband is more appropriate.

Edit: I also did this test with a random Gaussian distribution dataset, where the confbandl gave CI that were 2 orders of mag larger than expected, and the confband ones were within %5 of expectations. That's why I ultimately decided on utilizing confband. https://github.com/jrcooper91/van_der_Wel_size-mass/blob/master/Testing_bootstraps_BCES%20.ipynb
stats.confband

x = np.array([  0.0630552 ,0.138341,0.150978,1.509820,2.83197654,6.79663764,10.445637])
y = np.array([ 1.48,  1.14,  1.16,  1.15,  1.43,  1.43,  1.22])
yer = np.array([ 1.395,  0.1  ,  0.275,  0.12 ,  0.235,  0.135,  0.155])
sort = np.argsort(x)
x = x[sort]
x = np.array(x)
y = y[sort] 
y = np.array(y)
yer = yer[sort]
yer = np.array(yer)
xer=zeros(len(x))
cov=zeros(len(x))   # no correlation between error measurements
i=0 
nboot = 10000
a,b,erra,errb,covab=bces.bces.bcesp(x,xer,y,yer,cov,nboot)
ybces=a[3]*x+b[3]  
    # Gets lower and upper bounds on the confidence band 
lcb1,ucb1,x=nmmn.stats.confband(x, y, a[i], b[i], 0.68, x)
lcb2,ucb2,x2=nmmn.stats.confband(x, y, a[i], b[i], 0.95, x)
lcb3,ucb3,x3=nmmn.stats.confband(x, y, a[i], b[i], 0.997, x)
errorbar(x,y,yerr=yer,fmt='o')
ax = plot(x,ybces,'-k')
ax = fill_between(x, lcb1, ucb1, alpha=0.6, facecolor='purple')
ax = fill_between(x, lcb2, ucb2, alpha=0.3, facecolor='blue')
ax = fill_between(x, lcb3, ucb3, alpha=0.4, facecolor='grey')
ax = xlabel('x')
ax = ylabel('y')
plt.show()

stats.confbandl

x = np.array([ 0.0630552 , 0.13834198, 0.15097889, 1.50982026, 2.83197654, 6.79663764, 10.44563709])
y = np.array([ 1.48, 1.14, 1.16, 1.15, 1.43, 1.43, 1.22])
yer = np.array([ 1.395, 0.1 , 0.275, 0.12 , 0.235, 0.135, 0.155])
sort = np.argsort(x)
x = x[sort]
x = np.array(x)
y = y[sort]
y = np.array(y)
yer = yer[sort]
yer = np.array(yer)
xer=zeros(len(x))
cov=zeros(len(x)) # no correlation between error measurements
i=0
nboot = 10000
def func(x): return x[1]*x[0]+x[2]
a,b,erra,errb,covab=bces.bces.bcesp(x,xer,y,yer,cov,nboot)
ybces=a[3]*x+b[3] # the integer corresponds to the desired BCES method for plotting (3-ort, 0-y|x, 1-x|y, don't use bissector)
# array with best-fit parameters
fitm=np.array([ a[i],b[i] ])
# covariance matrix of parameter uncertainties
covm=np.array([ (erra[i]**2,covab[i]), (covab[i],errb[i]**2) ])
# Gets lower and upper bounds on the confidence band
lcb1,ucb1,xcb1=nmmn.stats.confbandnl(x,y,func,fitm,covm,2,0.68,x)
lcb2,ucb2,xcb2=nmmn.stats.confbandnl(x,y,func,fitm,covm,2,0.95, x)
lcb3,ucb3,xcb3=nmmn.stats.confbandnl(x,y,func,fitm,covm,2,0.997, x)
errorbar(xcb1,y,yerr=yer,fmt='o')
ax = plot(x,ybces,'-k')
ax = fill_between(x, lcb1, ucb1, alpha=0.6, facecolor='purple')
ax = fill_between(x, lcb2, ucb2, alpha=0.3, facecolor='blue')
ax = fill_between(x, lcb3, ucb3, alpha=0.4, facecolor='grey')
ax = xlabel('x')
ax = ylabel('y')
plt.show()

nmmn.dsp and 'stats example wavelet.ipynb' inconsistency

I recently updated my version of nmmn,
and found that the wa.reconstruction() method called from
https://gist.github.com/rsnemmen/baea0541c6cc23ae2fbe48421ffc838e
is now inconsistent.