def __init__(self, condition_layout, block_layout, alphas=None):
super(FStat, self).__init__(condition_layout, block_layout)
self.validate_layouts(condition_layout, block_layout)
self.layout_full = intersect_layouts(block_layout, condition_layout)
self.alphas = alphas
def __init__(self, condition_layout, block_layout, alphas=None):
super(FStat, self).__init__(condition_layout, block_layout)
self.validate_layouts(condition_layout, block_layout)
self.layout_full = intersect_layouts(block_layout, condition_layout)
self.alphas = alphas
def validate_layouts(cls, condition_layout, block_layout):
full_layout = intersect_layouts(block_layout, condition_layout)
if min(map(len, full_layout)) < 2:
raise UnsupportedLayoutException(
"I can't use an FTest with the specified layouts, because " +
"the intersection between those layouts results in some " +
"groups that contain fewer than two samples.")
def validate_layouts(cls, condition_layout, block_layout):
full_layout = intersect_layouts(block_layout, condition_layout)
if min(map(len, full_layout)) < 2:
raise UnsupportedLayoutException(
"I can't use an FTest with the specified layouts, because " +
"the intersection between those layouts results in some " +
"groups that contain fewer than two samples.")
def __call__(self, data):
conds = self.condition_layout
blocks = self.block_layout
# Build two new layouts. c0 is a list of lists of indexes into
# the data that represent condition 0 for each block. c1 is
# the same for data that represent condition 1 for each block.
c0_blocks = intersect_layouts(blocks, [ conds[0] ])
c1_blocks = intersect_layouts(blocks, [ conds[1] ])
# Get the mean for each block for both conditions.
means0 = group_means(data, c0_blocks)
means1 = group_means(data, c1_blocks)
# If we have tuning params, add another dimension to the front
# of each ndarray to vary the tuning param.
if self.alphas is not None:
shape = (len(self.alphas),) + np.shape(means0)
old0 = means0
old1 = means1
means0 = np.zeros(shape)
means1 = np.zeros(shape)
for i, a in enumerate(self.alphas):
means0[i] = old0 + a
means1[i] = old1 + a
means0 /= means1
ratio = means0
# If we have more than one block, we combine their ratios
# using the geometric mean.
ratio = gmean(ratio, axis=-1)
# 'Symmetric' means that the order of the conditions does not
# matter, so we should always return a ratio >= 1. So for any
# ratios that are < 1, use the inverse.
if self.symmetric:
# Add another dimension to the front where and 1 is its
# inverse, then select the max across that dimension
ratio_and_inverse = np.array([ratio, 1.0 / ratio])
ratio = np.max(ratio_and_inverse, axis=0)
return ratio
def __init__(self, condition_layout, block_layout, alphas=None, family='gaussian', shrink=False):
super(GLMFStat, self).__init__(condition_layout, block_layout)
self.layout_full = intersect_layouts(block_layout, condition_layout)
self.alphas = alphas
ctor = GLM_FAMILIES[family]
self.family = ctor()
self.shrink = shrink
def __call__(self, data):
conds = self.condition_layout
blocks = self.block_layout
# Build two new layouts. c0 is a list of lists of indexes into
# the data that represent condition 0 for each block. c1 is
# the same for data that represent condition 1 for each block.
c0_blocks = intersect_layouts(blocks, [conds[0]])
c1_blocks = intersect_layouts(blocks, [conds[1]])
# Get the mean for each block for both conditions.
means0 = group_means(data, c0_blocks)
means1 = group_means(data, c1_blocks)
# If we have tuning params, add another dimension to the front
# of each ndarray to vary the tuning param.
if self.alphas is not None:
shape = (len(self.alphas), ) + np.shape(means0)
old0 = means0
old1 = means1
means0 = np.zeros(shape)
means1 = np.zeros(shape)
for i, a in enumerate(self.alphas):
means0[i] = old0 + a
means1[i] = old1 + a
means0 /= means1
ratio = means0
# If we have more than one block, we combine their ratios
# using the geometric mean.
ratio = gmean(ratio, axis=-1)
# 'Symmetric' means that the order of the conditions does not
# matter, so we should always return a ratio >= 1. So for any
# ratios that are < 1, use the inverse.
if self.symmetric:
# Add another dimension to the front where and 1 is its
# inverse, then select the max across that dimension
ratio_and_inverse = np.array([ratio, 1.0 / ratio])
ratio = np.max(ratio_and_inverse, axis=0)
return ratio
def __init__(self,
condition_layout,
block_layout,
alphas=None,
family='gaussian',
shrink=False):
super(GLMFStat, self).__init__(condition_layout, block_layout)
self.layout_full = intersect_layouts(block_layout, condition_layout)
self.alphas = alphas
ctor = GLM_FAMILIES[family]
self.family = ctor()
self.shrink = shrink
def __call__(self, data):
pairs = self.block_layout
conds = self.condition_layout
values = []
for i in [0, 1]:
# Make a new layout that is just the item for each pair
# from condition i. layout will be a list of sets, each
# with just one index, since there is only one item from
# each pair with condition i. So flatten it into a list of
# indexes, and grab the corresponding values from the
# data.
layout = intersect_layouts(self.block_layout, [conds[i]])
idxs = list(itertools.chain(*layout))
values.append(data[..., idxs])
# Now just get the differences between the two sets of values
# and call the child statistic on those values.
return self.child(values[0] - values[1])
def __call__(self, data):
pairs = self.block_layout
conds = self.condition_layout
values = []
for i in [ 0, 1 ]:
# Make a new layout that is just the item for each pair
# from condition i. layout will be a list of sets, each
# with just one index, since there is only one item from
# each pair with condition i. So flatten it into a list of
# indexes, and grab the corresponding values from the
# data.
layout = intersect_layouts(self.block_layout, [ conds[i] ])
idxs = list(itertools.chain(*layout))
values.append(data[..., idxs])
# Now just get the differences between the two sets of values
# and call the child statistic on those values.
return self.child(values[0] - values[1])
