def __call__(self, pt):
"""
"""
pt = _matrix([[pt.x], [pt.y], [1]])
pt = _dot(self.matrix, pt)
x, y, w = pt[0,0], pt[1,0], pt[2,0]
return Point(x/w, y/w)
def __call__(self, pt):
"""
"""
pt = _matrix([[pt.x], [pt.y], [1]])
pt = _dot(self.matrix, pt)
x, y, w = pt[0, 0], pt[1, 0], pt[2, 0]
return Point(x / w, y / w)
def _pca_data(e_, model_keys, pc_filter):
if e_.shape[1] == 2:
return e_, model_keys, 1, 1, 1
#numpy.svd can't handle typical data size in UK Biobank. So we do PCA through the covariance matrix
# That is: we compute ths SVD of a covariance matrix, and use those coefficients to get the SVD of input data
# Shamelessly designed from https://stats.stackexchange.com/questions/134282/relationship-between-svd-and-pca-how-to-use-svd-to-perform-pca
# In numpy.cov, each row is a variable and each column an observation. Exactly opposite to standard PCA notation: it is transposed, then.
Xc_t = _get_pc_input(e_, model_keys)
k = numpy.cov(Xc_t)
u, s, vt = numpy.linalg.svd(k)
# we want to keep only those components with significant variance, to reduce dimensionality
selected = pc_filter(s)
Xc_t_ = _dot(vt[selected], Xc_t)
_data = {"pc{}".format(i): x for i, x in enumerate(Xc_t_)}
pca_keys = _data.keys()
_data["pheno"] = e_.pheno
pca_data = pandas.DataFrame(_data)
return pca_data, pca_keys, numpy.max(s), numpy.min(s), numpy.min(
s[selected])
def multiply(self, other):
"""
"""
return matrix2transform(_dot(other.matrix, self.matrix))
def multiply(self, other):
"""
"""
return matrix2transform(_dot(other.matrix, self.matrix))
