denis-bz

# interp1d != splev  random-uniform -> 0 1 2 ...
# big swings, different end conditions
# http://docs.scipy.org/doc/scipy/reference/generated/scipy.interpolate.interp1d.html
# http://docs.scipy.org/doc/scipy/reference/generated/scipy.interpolate.splev.html
# http://stackoverflow.com/questions/6216881/python-interp1d-vs-univariatespline
# 28mar 2013

from __future__ import division
import sys
import numpy as np

denis-bz

# from: griddata-intergrid.py
# run: 29 Jun 2013 19:09  in ~bz/py/interpol/git/intergrid/test  i486  mac 10.8.3

versions: numpy 1.7.1  scipy 0.12.0  python 2.7.4
--------------------------------------------------------------------------------
griddata vs. intergrid:  Nx 36 Ny 36 Nz 84  subsample 2  order 1  prefilter 0
func min av max: -1 0.0025 1
griddata: 121.1 sec
Intergrid: 2.82 msec  13608 points in a (36, 36, 84) grid  3 maps  order 1
av |f - griddata|: 0.44

denis-bz

""" Compare Savitzky-Golay filter f( x + xnoise )  ~  Savgol( f( x )) + ynoise
for f(x) = sin( 2 pi freq x )
to try to elucidate Press et al., Numerical Recipes p. 772:
    "for irregularly sampled data ... one can pretend that the data points *are* equally spaced
    ... if change in f across the full width ..." 

(Simpler: noise amplification of linear filters:
    |convolve( filter, x )| <= |filter| * |x|, || = rms
    |flat filter| = 1 / sqrt(N)
    |Savgol| ~ 2 * that

denis-bz

#!/bin/sh
# conda-requires: list "requires" of the 100-odd packages in anaconda,
#   http://docs.continuum.io/anaconda/pkgs.html
# gist.github.com/denis-bz

[[ $1 ]]  &&  cd $1  &&  shift  # .../anaconda/conda-meta

echo "# ${0##*/}  $PWD  `/bin/date '+%e %h %Y'` "  # 9 Aug 2013

for file in *.json

denis-bz

Links to html doc of a couple of Python-NumPy programs in github.com/denis-bz :

Intergrid: interpolate in an N-d box grid, uniform or non-uniform. This is just a wrapper for scipy.ndimage.map_coordinates and numpy.interp. It's fast: ~ 1M interpolations / sec in 4d.

Barypol: interpolate in an N-d box grid, using d + 1 corners of a simplex

denis-bz

#!/usr/bin/env python
""" nearsame.py: an easy and intuitive way to find outliers in data 2d to 20d.
    -> e.g.
        [50 24 15 10] % of the 38356 points have
        [ 3  2  1  0] of the 3 nearest neighbors in the same class.

    Here one could keep the 74 % for which the nearest 3 neighbors
    are mostly in the same class, and call the rest outliers.

    Intuitively, points in mixed neighborhoods, where classes overlap,

denis-bz

Test generator for sparse linear programming: m+n constraints, mn variables, 2mn non-zeros

Keywords: linear-programming, sparse, scipy, test-case-generator

Purpose

lpgen_2d generates test cases for linear programming, with parameters m and n: m+n constraints, mn variables, 2mn non-zeros. An example:

from scipy.optimize import linprog

denis-bz

One-line arg parse for flexible testing in Python

Many Python scripts with parameters say M and N look like this:

# parameters --
M = 10
N = 10
...

myfunc( M, N ... )

denis-bz

""" heapq_del.py: heapq.py + heapchange heapdel
    see http://docs.python.org/2/library/heapq.html
    c heap* but python _siftup _siftdown, slow
"""
    # cf. class Aheapq: heap --> A[ Arecs with .val .heappos ... ]

from heapq import heappush, heappop, heapify, heapreplace, _siftup, _siftdown
    # http://docs.python.org/2/library/heapq.html
    # https://github.com/python/cpython/blob/master/Modules/_heapqmodule.c
    # (note _siftup --> leaves, _siftdown --> root

denis-bz

""" trim outliers: smallest n values = next smallest, biggest n = next biggest
    X, smallest, biggest = winsorize( X, n )
        -> X trimmed *inplace*
        smallest: n+1 smallest
        biggest: n+1 biggest

    To winsorize the top 5 % and bottom 5 %,
        winsorize( X, X.size * .05 )
    X may be 2d, 3d ... (but winsorizing each dimension separately would make more sense)
    See also http://en.wikipedia.org/wiki/Winsorising