def find_cont(fl, fwhm1=300, fwhm2=200, nchunks=4):
""" Given the flux, estimate the continuum. fwhm values are
smoothing lengths.
"""
# smooth flux, with smoothing length much longer than expected
# emission line widths.
fl = nan2num(fl.astype(float), replace="mean")
co = convolve_psf(fl, fwhm1, edge=10)
npts = len(fl)
indices = np.arange(npts)
# throw away top and bottom 2% of data that deviates from
# continuum and re-fit new continuum. Go chunk by chunk so that
# points are thrown away evenly across the spectrum.
nfl = fl / co
step = npts // nchunks + 1
ind = range(0, npts, step) + [npts]
igood = []
for i0, i1 in zip(ind[:-1], ind[1:]):
isort = nfl[i0:i1].argsort()
len_isort = len(isort)
j0, j1 = int(0.05 * len_isort), int(0.95 * len_isort)
igood.extend(isort[j0:j1] + i0)
good = np.in1d(indices, igood)
sfl = fl.copy()
sfl[~good] = np.interp(indices[~good], indices[good], sfl[good])
co = convolve_psf(sfl, fwhm2, edge=10)
return co
def convolve_psf(a, fwhm, edge="invert", replace_nan=True, debug=False):
""" Convolve an array with a gaussian window.
Given an array of values `a` and a gaussian full width at half
maximum `fwhm` in pixel units, returns the convolution of the
array with the normalised gaussian.
Parameters
----------
a : array, shape(N,)
Array to convolve
fwhm : float
Gaussian full width at half maximum in pixels. This should be > 2
to sample the gaussian PSF properly.
Returns
-------
convolved_a : array, shape (N,)
Notes
-----
The Gaussian kernel is calculated for as many pixels required
until it drops to 1% of its peak value. The data will be spoiled
at distances `n`/2 (rounding down) from the edges, where `n` is
the width of the Gaussian in pixels.
"""
const2 = 2.354820046 # 2*sqrt(2*ln(2))
const100 = 3.034854259 # sqrt(2*ln(100))
sigma = fwhm / const2
# gaussian drops to 1/100 of maximum value at x =
# sqrt(2*ln(100))*sigma, so number of pixels to include from
# centre of gaussian is:
n = np.ceil(const100 * sigma)
if replace_nan:
a = nan2num(a, replace="interp")
if debug:
print "First and last %s pixels of output array will be invalid" % n
x = np.linspace(-n, n, 2 * n + 1) # total no. of pixels = 2n+1
gauss = np.exp(-0.5 * (x / sigma) ** 2)
return convolve_window(a, gauss, edge=edge)
