# Licensed under a 3-clause BSD style license - see LICENSE.rst
"""
Classes that deal with computing intervals from arrays of values based on
various criteria.
"""
import abc
import numpy as np
from astropy.utils.decorators import deprecated_attribute, deprecated_renamed_argument
from .transform import BaseTransform
__all__ = [
"BaseInterval",
"ManualInterval",
"MinMaxInterval",
"AsymmetricPercentileInterval",
"PercentileInterval",
"ZScaleInterval",
]
[docs]class BaseInterval(BaseTransform):
"""
Base class for the interval classes, which, when called with an
array of values, return an interval computed following different
algorithms.
"""
[docs] @abc.abstractmethod
def get_limits(self, values):
"""
Return the minimum and maximum value in the interval based on
the values provided.
Parameters
----------
values : ndarray
The image values.
Returns
-------
vmin, vmax : float
The mininium and maximum image value in the interval.
"""
raise NotImplementedError("Needs to be implemented in a subclass.")
[docs] def __call__(self, values, clip=True, out=None):
"""
Transform values using this interval.
Parameters
----------
values : array-like
The input values.
clip : bool, optional
If `True` (default), values outside the [0:1] range are
clipped to the [0:1] range.
out : ndarray, optional
If specified, the output values will be placed in this array
(typically used for in-place calculations).
Returns
-------
result : ndarray
The transformed values.
"""
vmin, vmax = self.get_limits(values)
if out is None:
values = np.subtract(values, float(vmin))
else:
if out.dtype.kind != "f":
raise TypeError(
"Can only do in-place scaling for floating-point arrays"
)
values = np.subtract(values, float(vmin), out=out)
if (vmax - vmin) != 0:
np.true_divide(values, vmax - vmin, out=values)
if clip:
np.clip(values, 0.0, 1.0, out=values)
return values
[docs]class ManualInterval(BaseInterval):
"""
Interval based on user-specified values.
Parameters
----------
vmin : float, optional
The minimum value in the scaling. Defaults to the image
minimum (ignoring NaNs)
vmax : float, optional
The maximum value in the scaling. Defaults to the image
maximum (ignoring NaNs)
"""
def __init__(self, vmin=None, vmax=None):
self.vmin = vmin
self.vmax = vmax
[docs] def get_limits(self, values):
# Avoid overhead of preparing array if both limits have been specified
# manually, for performance.
if self.vmin is not None and self.vmax is not None:
return self.vmin, self.vmax
# Make sure values is a Numpy array
values = np.asarray(values).ravel()
# Filter out invalid values (inf, nan)
values = values[np.isfinite(values)]
vmin = np.min(values) if self.vmin is None else self.vmin
vmax = np.max(values) if self.vmax is None else self.vmax
return vmin, vmax
[docs]class MinMaxInterval(BaseInterval):
"""
Interval based on the minimum and maximum values in the data.
"""
[docs] def get_limits(self, values):
# Make sure values is a Numpy array
values = np.asarray(values).ravel()
# Filter out invalid values (inf, nan)
values = values[np.isfinite(values)]
return np.min(values), np.max(values)
[docs]class AsymmetricPercentileInterval(BaseInterval):
"""
Interval based on a keeping a specified fraction of pixels (can be
asymmetric).
Parameters
----------
lower_percentile : float
The lower percentile below which to ignore pixels.
upper_percentile : float
The upper percentile above which to ignore pixels.
n_samples : int, optional
Maximum number of values to use. If this is specified, and there
are more values in the dataset as this, then values are randomly
sampled from the array (with replacement).
"""
def __init__(self, lower_percentile, upper_percentile, n_samples=None):
self.lower_percentile = lower_percentile
self.upper_percentile = upper_percentile
self.n_samples = n_samples
[docs] def get_limits(self, values):
# Make sure values is a Numpy array
values = np.asarray(values).ravel()
# If needed, limit the number of samples. We sample with replacement
# since this is much faster.
if self.n_samples is not None and values.size > self.n_samples:
values = np.random.choice(values, self.n_samples)
# Filter out invalid values (inf, nan)
values = values[np.isfinite(values)]
# Determine values at percentiles
vmin, vmax = np.percentile(
values, (self.lower_percentile, self.upper_percentile)
)
return vmin, vmax
[docs]class PercentileInterval(AsymmetricPercentileInterval):
"""
Interval based on a keeping a specified fraction of pixels.
Parameters
----------
percentile : float
The fraction of pixels to keep. The same fraction of pixels is
eliminated from both ends.
n_samples : int, optional
Maximum number of values to use. If this is specified, and there
are more values in the dataset as this, then values are randomly
sampled from the array (with replacement).
"""
def __init__(self, percentile, n_samples=None):
lower_percentile = (100 - percentile) * 0.5
upper_percentile = 100 - lower_percentile
super().__init__(lower_percentile, upper_percentile, n_samples=n_samples)
[docs]class ZScaleInterval(BaseInterval):
"""
Interval based on IRAF's zscale.
https://iraf.net/forum/viewtopic.php?showtopic=134139
Original implementation:
https://github.com/spacetelescope/stsci.numdisplay/blob/master/lib/stsci/numdisplay/zscale.py
Licensed under a 3-clause BSD style license (see AURA_LICENSE.rst).
Parameters
----------
n_samples : int, optional
The number of points in the array to sample for determining
scaling factors. Defaults to 1000.
.. versionchanged:: 5.2
``n_samples`` replaces the deprecated ``nsamples`` argument,
which will be removed in the future.
contrast : float, optional
The scaling factor (between 0 and 1) for determining the minimum
and maximum value. Larger values increase the difference
between the minimum and maximum values used for display.
Defaults to 0.25.
max_reject : float, optional
If more than ``max_reject * npixels`` pixels are rejected, then
the returned values are the minimum and maximum of the data.
Defaults to 0.5.
min_npixels : int, optional
If there are less than ``min_npixels`` pixels remaining after
the pixel rejection, then the returned values are the minimum
and maximum of the data. Defaults to 5.
krej : float, optional
The number of sigma used for the rejection. Defaults to 2.5.
max_iterations : int, optional
The maximum number of iterations for the rejection. Defaults to
5.
"""
@deprecated_renamed_argument("nsamples", "n_samples", "5.2")
def __init__(
self,
n_samples=1000,
contrast=0.25,
max_reject=0.5,
min_npixels=5,
krej=2.5,
max_iterations=5,
):
self.n_samples = n_samples
self.contrast = contrast
self.max_reject = max_reject
self.min_npixels = min_npixels
self.krej = krej
self.max_iterations = max_iterations
# Mark `nsamples` as deprecated
nsamples = deprecated_attribute("nsamples", "5.2", alternative="n_samples")
[docs] def get_limits(self, values):
# Sample the image
values = np.asarray(values)
values = values[np.isfinite(values)]
stride = int(max(1.0, values.size / self.n_samples))
samples = values[::stride][: self.n_samples]
samples.sort()
npix = len(samples)
vmin = samples[0]
vmax = samples[-1]
# Fit a line to the sorted array of samples
minpix = max(self.min_npixels, int(npix * self.max_reject))
x = np.arange(npix)
ngoodpix = npix
last_ngoodpix = npix + 1
# Bad pixels mask used in k-sigma clipping
badpix = np.zeros(npix, dtype=bool)
# Kernel used to dilate the bad pixels mask
ngrow = max(1, int(npix * 0.01))
kernel = np.ones(ngrow, dtype=bool)
for _ in range(self.max_iterations):
if ngoodpix >= last_ngoodpix or ngoodpix < minpix:
break
fit = np.polyfit(x, samples, deg=1, w=(~badpix).astype(int))
fitted = np.poly1d(fit)(x)
# Subtract fitted line from the data array
flat = samples - fitted
# Compute the k-sigma rejection threshold
threshold = self.krej * flat[~badpix].std()
# Detect and reject pixels further than k*sigma from the
# fitted line
badpix[(flat < -threshold) | (flat > threshold)] = True
# Convolve with a kernel of length ngrow
badpix = np.convolve(badpix, kernel, mode="same")
last_ngoodpix = ngoodpix
ngoodpix = np.sum(~badpix)
if ngoodpix >= minpix:
slope, _ = fit
if self.contrast > 0:
slope = slope / self.contrast
center_pixel = (npix - 1) // 2
median = np.median(samples)
vmin = max(vmin, median - (center_pixel - 1) * slope)
vmax = min(vmax, median + (npix - center_pixel) * slope)
return vmin, vmax