isn't this related to http://projects.scipy.org/numpy/ticket/626 percentile() and clamp() which was set to invalid -Sebastian On Sun, May 16, 2010 at 12:11 AM, wrote:

Author: oliphant Date: 2010-05-15 17:11:10 -0500 (Sat, 15 May 2010) New Revision: 8413

Modified:   trunk/numpy/lib/function_base.py Log: Add percentile function.

Modified: trunk/numpy/lib/function_base.py =================================================================== --- trunk/numpy/lib/function_base.py    2010-05-13 12:47:23 UTC (rev 8412) +++ trunk/numpy/lib/function_base.py    2010-05-15 22:11:10 UTC (rev 8413) @@ -1,6 +1,6 @@  __docformat__ = "restructuredtext en"  __all__ = ['select', 'piecewise', 'trim_zeros', -           'copy', 'iterable', +           'copy', 'iterable', 'percentile',            'diff', 'gradient', 'angle', 'unwrap', 'sort_complex', 'disp',            'extract', 'place', 'nansum', 'nanmax', 'nanargmax',            'nanargmin', 'nanmin', 'vectorize', 'asarray_chkfinite', 'average', @@ -2804,7 +2804,7 @@

    See Also     -------- -    mean +    mean, percentile

    Notes     ----- @@ -2863,6 +2863,132 @@     # and check, use out array.     return mean(sorted[indexer], axis=axis, out=out)

+def percentile(a, q, axis=None, out=None, overwrite_input=False): +    """ +    Compute the qth percentile of the data along the specified axis. + +    Returns the qth percentile of the array elements. + +    Parameters +    ---------- +    a : array_like +        Input array or object that can be converted to an array. +    q : float in range of [0,100] (or sequence of floats) +        percentile to compute which must be between 0 and 100 inclusive +    axis : {None, int}, optional +        Axis along which the percentiles are computed. The default (axis=None) +        is to compute the median along a flattened version of the array. +    out : ndarray, optional +        Alternative output array in which to place the result. It must +        have the same shape and buffer length as the expected output, +        but the type (of the output) will be cast if necessary. +    overwrite_input : {False, True}, optional +       If True, then allow use of memory of input array (a) for +       calculations. The input array will be modified by the call to +       median. This will save memory when you do not need to preserve +       the contents of the input array. Treat the input as undefined, +       but it will probably be fully or partially sorted. Default is +       False. Note that, if `overwrite_input` is True and the input +       is not already an ndarray, an error will be raised. + +    Returns +    ------- +    pcntile : ndarray +        A new array holding the result (unless `out` is specified, in +        which case that array is returned instead).  If the input contains +        integers, or floats of smaller precision than 64, then the output +        data-type is float64.  Otherwise, the output data-type is the same +        as that of the input. + +    See Also +    -------- +    mean, median + +    Notes +    ----- +    Given a vector V of length N, the qth percentile of V is the qth ranked +    value in a sorted copy of V.  A weighted average of the two nearest neighbors +    is used if the normalized ranking does not match q exactly. +    The same as the median if q is 0.5; the same as the min if q is 0; +    and the same as the max if q is 1 + +    Examples +    -------- +    >>> a = np.array([[10, 7, 4], [3, 2, 1]]) +    >>> a +    array([[10,  7,  4], +           [ 3,  2,  1]]) +    >>> np.percentile(a, 0.5) +    3.5 +    >>> np.percentile(a, 0.5, axis=0) +    array([ 6.5,  4.5,  2.5]) +    >>> np.percentile(a, 0.5, axis=1) +    array([ 7.,  2.]) +    >>> m = np.percentile(a, 0.5, axis=0) +    >>> out = np.zeros_like(m) +    >>> np.percentile(a, 0.5, axis=0, out=m) +    array([ 6.5,  4.5,  2.5]) +    >>> m +    array([ 6.5,  4.5,  2.5]) +    >>> b = a.copy() +    >>> np.percentile(b, 0.5, axis=1, overwrite_input=True) +    array([ 7.,  2.]) +    >>> assert not np.all(a==b) +    >>> b = a.copy() +    >>> np.percentile(b, 0.5, axis=None, overwrite_input=True) +    3.5 +    >>> assert not np.all(a==b) + +    """ +    if q == 0: +        return a.min(axis=axis, out=out) +    elif q == 100: +        return a.max(axis=axis, out=out) + +    if overwrite_input: +        if axis is None: +            sorted = a.ravel() +            sorted.sort() +        else: +            a.sort(axis=axis) +            sorted = a +    else: +        sorted = sort(a, axis=axis) +    if axis is None: +        axis = 0 + +    return _compute_qth_percentile(sorted, q, axis, out) + +# handle sequence of q's without calling sort multiple times +def _compute_qth_percentile(sorted, q, axis, out): +    if not isscalar(q): +        return [_compute_qth_percentile(sorted, qi, axis, out) +                    for qi in q] +    q = q / 100.0 +    if (q < 0) or (q > 1): +        raise ValueError, "percentile must be either in the range [0,100]" + +    indexer = [slice(None)] * sorted.ndim +    Nx = sorted.shape[axis] +    index = q*(Nx-1) +    i = int(index) +    if i == index: +        indexer[axis] = slice(i, i+1) +        weights = array(1) +        sumval = 1.0 +    else: +        indexer[axis] = slice(i, i+2) +        j = i + 1 +        weights = array([(j - index), (index - i)],float) +        wshape = [1]*sorted.ndim +        wshape[axis] = 2 +        weights.shape = wshape +        sumval = weights.sum() + +    # Use add.reduce in both cases to coerce data type as well as +    #   check and use out array. +    return add.reduce(sorted[indexer]*weights, axis=axis, out=out)/sumval +  def trapz(y, x=None, dx=1.0, axis=-1):     """     Integrate along the given axis using the composite trapezoidal rule.

_______________________________________________ Numpy-svn mailing list Numpy-svn@scipy.org http://mail.scipy.org/mailman/listinfo/numpy-svn