def plot(self):
self.graph.clearPlot()
self.validindices = numpy.empty((0,), dtype=int)
self.current_selection = []
group, target_indices = self.selected_split()
self.warning([0, 1])
self.error(1)
if self.data and group is not None and target_indices:
X = self.data.X
I1 = grouputils.group_selection_mask(
self.data, group, target_indices)
I2 = ~I1
if isinstance(group, grouputils.RowGroup):
X = X.T
N1, N2 = numpy.count_nonzero(I1), numpy.count_nonzero(I2)
if not N1 or not N2:
self.error(
1, "Target labels most exclude/include at least one value."
)
if N1 < 2 and N2 < 2:
self.warning(
0, "Insufficient data to compute statistics. "
"More than one measurement per class should be provided"
)
X1, X2 = X[:, I1], X[:, I2]
if numpy.any(X1 < 0.0) or numpy.any(X2 < 0):
self.error(
"Negative values in the input. The inputs cannot be in "
"ratio scale."
)
X1 = numpy.full_like(X1, numpy.nan)
X2 = numpy.full_like(X2, numpy.nan)
with numpy.errstate(divide="ignore", invalid="ignore"):
fold = numpy.log2(numpy.mean(X1, axis=1) /
numpy.mean(X2, axis=1))
# TODO: handle missing values better (mstats)
_, P = scipy.stats.ttest_ind(X1, X2, axis=1, equal_var=True)
logP = numpy.log10(P)
if numpy.isscalar(logP):
# ttest_ind does not preserve output shape if either
# a or b is empty
logP = numpy.full(fold.shape, numpy.nan)
mask = numpy.isfinite(fold) & numpy.isfinite(logP)
self.validindices = numpy.flatnonzero(mask)
self.graph.setPlotData(numpy.array([fold[mask], -logP[mask]]).T)
self.infoLabel.setText("%i genes on input" % len(fold))
# ("{displayed} displayed, {undef} with undefined ratio "
# "or t-statistics.")
if not len(numpy.flatnonzero(mask)):
self.warning(1, "Could not compute statistics for any genes!")
def plot(self):
self.graph.clearPlot()
self.validindices = numpy.empty((0, ), dtype=int)
self.current_selection = []
group, target_indices = self.selected_split()
self.warning([0, 1])
self.error(1)
if self.data and group is not None and target_indices:
X = self.data.X
I1 = grouputils.group_selection_mask(self.data, group,
target_indices)
I2 = ~I1
if isinstance(group, grouputils.RowGroup):
X = X.T
N1, N2 = numpy.count_nonzero(I1), numpy.count_nonzero(I2)
if not N1 or not N2:
self.error(
1,
"Target labels most exclude/include at least one value.")
if N1 < 2 and N2 < 2:
self.warning(
0, "Insufficient data to compute statistics. "
"More than one measurement per class should be provided")
X1, X2 = X[:, I1], X[:, I2]
if numpy.any(X1 < 0.0) or numpy.any(X2 < 0):
self.error(
"Negative values in the input. The inputs cannot be in "
"ratio scale.")
X1 = numpy.full_like(X1, numpy.nan)
X2 = numpy.full_like(X2, numpy.nan)
with numpy.errstate(divide="ignore", invalid="ignore"):
fold = numpy.log2(
numpy.mean(X1, axis=1) / numpy.mean(X2, axis=1))
# TODO: handle missing values better (mstats)
_, P = scipy.stats.ttest_ind(X1, X2, axis=1, equal_var=True)
logP = numpy.log10(P)
if numpy.isscalar(logP):
# ttest_ind does not preserve output shape if either
# a or b is empty
logP = numpy.full(fold.shape, numpy.nan)
mask = numpy.isfinite(fold) & numpy.isfinite(logP)
self.validindices = numpy.flatnonzero(mask)
self.graph.setPlotData(numpy.array([fold[mask], -logP[mask]]).T)
self.infoLabel.setText("%i genes on input" % len(fold))
# ("{displayed} displayed, {undef} with undefined ratio "
# "or t-statistics.")
if not len(numpy.flatnonzero(mask)):
self.warning(1, "Could not compute statistics for any genes!")
def update_scores(self):
"""Compute the scores and update the histogram.
"""
self.__cancel_pending()
self.clear_plot()
self.scores = None
self.nulldist = None
self.error(0)
grp, split_selection = self.selected_split()
if not self.data or grp is None:
return
_, side, test_type, score_func = self.Scores[self.score_index]
def compute_scores(X, group_indices, warn=False):
arrays = [X[ind] for ind in group_indices]
ss = score_func(*arrays, axis=0)
return ss[0] if isinstance(ss, tuple) and not warn else ss
def permute_indices(group_indices, random_state=None):
assert all(ind.dtype.kind == "i" for ind in group_indices)
assert all(ind.ndim == 1 for ind in group_indices)
if random_state is None:
random_state = np.random
joined = np.hstack(group_indices)
random_state.shuffle(joined)
split_ind = np.cumsum([len(ind) for ind in group_indices])
return np.split(joined, split_ind[:-1])
if isinstance(grp, grouputils.RowGroup):
axis = 0
else:
axis = 1
if test_type == OWFeatureSelection.TwoSampleTest:
G1 = grouputils.group_selection_mask(self.data, grp,
split_selection)
G2 = ~G1
indices = [np.flatnonzero(G1), np.flatnonzero(G2)]
elif test_type == self.VarSampleTest:
indices = [
grouputils.group_selection_mask(self.data, grp, [i])
for i in range(len(grp.values))
]
indices = [np.flatnonzero(ind) for ind in indices]
else:
assert False
if not all(np.count_nonzero(ind) > 0 for ind in indices):
self.error(
0, "Target labels most exclude/include at least one "
"value.")
self.scores = None
self.nulldist = None
self.update_data_info_label()
return
X = self.data.X
if axis == 1:
X = X.T
# TODO: Check that each label has more than one measurement,
# raise warning otherwise.
def compute_scores_with_perm(X,
indices,
nperm=0,
rstate=None,
progress_advance=None):
warning = None
scores = compute_scores(X, indices, warn=True)
if isinstance(scores, tuple):
scores, warning = scores
if progress_advance is not None:
progress_advance()
null_scores = []
if nperm > 0:
if rstate is None:
rstate = np.random.RandomState(0)
for i in range(nperm):
p_indices = permute_indices(indices, rstate)
assert all(pind.shape == ind.shape
for pind, ind in zip(indices, p_indices))
pscore = compute_scores(X, p_indices)
assert pscore.shape == scores.shape
null_scores.append(pscore)
if progress_advance is not None:
progress_advance()
return scores, null_scores, warning
p_advance = concurrent.methodinvoke(self, "progressBarAdvance",
(float, ))
state = namespace(cancelled=False, advance=p_advance)
def progress():
if state.cancelled:
raise concurrent.CancelledError
else:
state.advance(100 / (nperm + 1))
self.progressBarInit()
set_scores = concurrent.methodinvoke(self, "__set_score_results",
(concurrent.Future, ))
nperm = self.permutations_count if self.compute_null else 0
self.__scores_state = state
self.__scores_future = self._executor.submit(compute_scores_with_perm,
X,
indices,
nperm,
progress_advance=progress)
self.__scores_future.add_done_callback(set_scores)
def update_scores(self):
"""Compute the scores and update the histogram.
"""
self.__cancel_pending()
self.clear_plot()
self.scores = None
self.nulldist = None
self.error(0)
grp, split_selection = self.selected_split()
if not self.data or grp is None:
return
_, side, test_type, score_func = self.Scores[self.score_index]
def compute_scores(X, group_indices, warn=False):
arrays = [X[ind] for ind in group_indices]
ss = score_func(*arrays, axis=0)
return ss[0] if isinstance(ss, tuple) and not warn else ss
def permute_indices(group_indices, random_state=None):
assert all(ind.dtype.kind == "i" for ind in group_indices)
assert all(ind.ndim == 1 for ind in group_indices)
if random_state is None:
random_state = np.random
joined = np.hstack(group_indices)
random_state.shuffle(joined)
split_ind = np.cumsum([len(ind) for ind in group_indices])
return np.split(joined, split_ind[:-1])
if isinstance(grp, grouputils.RowGroup):
axis = 0
else:
axis = 1
if test_type == OWFeatureSelection.TwoSampleTest:
G1 = grouputils.group_selection_mask(
self.data, grp, split_selection)
G2 = ~G1
indices = [np.flatnonzero(G1), np.flatnonzero(G2)]
elif test_type == self.VarSampleTest:
indices = [grouputils.group_selection_mask(self.data, grp, [i])
for i in range(len(grp.values))]
indices = [np.flatnonzero(ind) for ind in indices]
else:
assert False
if not all(ind.size > 0 for ind in indices):
self.error(0, "Target labels most exclude/include at least one "
"value.")
self.scores = None
self.nulldist = None
self.update_data_info_label()
return
X = self.data.X
if axis == 1:
X = X.T
# TODO: Check that each label has more than one measurement,
# raise warning otherwise.
def compute_scores_with_perm(X, indices, nperm=0, rstate=None,
progress_advance=None):
warning = None
scores = compute_scores(X, indices, warn=True)
if isinstance(scores, tuple):
scores, warning = scores
if progress_advance is not None:
progress_advance()
null_scores = []
if nperm > 0:
if rstate is None:
rstate = np.random.RandomState(0)
for i in range(nperm):
p_indices = permute_indices(indices, rstate)
assert all(pind.shape == ind.shape
for pind, ind in zip(indices, p_indices))
pscore = compute_scores(X, p_indices)
assert pscore.shape == scores.shape
null_scores.append(pscore)
if progress_advance is not None:
progress_advance()
return scores, null_scores, warning
p_advance = concurrent.methodinvoke(
self, "progressBarAdvance", (float,))
state = namespace(cancelled=False, advance=p_advance)
def progress():
if state.cancelled:
raise concurrent.CancelledError
else:
state.advance(100 / (nperm + 1))
self.progressBarInit()
set_scores = concurrent.methodinvoke(
self, "__set_score_results", (concurrent.Future,))
nperm = self.permutations_count if self.compute_null else 0
self.__scores_state = state
self.__scores_future = self._executor.submit(
compute_scores_with_perm, X, indices, nperm,
progress_advance=progress)
self.__scores_future.add_done_callback(set_scores)
