def test_empty():
np.testing.assert_array_equal(
one_hot([]), np.zeros((0, 0))
)
np.testing.assert_array_equal(
one_hot([], dim=2), np.zeros((0, 2))
)
def one_hot_probs(value):
if not multitarget:
return one_hot(value)
max_card = max(len(c.values) for c in self.domain.class_vars)
probs = np.zeros(value.shape + (max_card,), float)
for i in range(len(self.domain.class_vars)):
probs[:, i, :] = one_hot(value[:, i])
return probs
def _get_bin_distributions(self, bin_indices):
"""Compute the distribution of instances within bins.
Parameters
----------
bin_indices : np.ndarray
An array with same shape as `x` but containing the bin index of the
instance.
Returns
-------
np.ndarray
A 2d array; the first dimension represents different bins, the
second - the counts of different target values.
"""
if self.target_var and self.target_var.is_discrete:
y = self.y
# TODO This probably also isn't the best handling of sparse data...
if sp.issparse(y):
y = np.squeeze(np.array(y.todense()))
y = one_hot(y)
bins = np.arange(self.n_bins)[:, np.newaxis]
mask = bin_indices == bins
distributions = np.zeros((self.n_bins, y.shape[1]))
for bin_idx in range(self.n_bins):
distributions[bin_idx] = y[mask[bin_idx]].sum(axis=0)
else:
distributions, _ = ut.bincount(bin_indices.astype(np.int64))
# To keep things consistent across different variable types, we
# want to return a 2d array where the first dim represent different
# bins, and the second the distributions.
distributions = distributions[:, np.newaxis]
return distributions
def test_dim(self):
np.testing.assert_array_equal(
one_hot(self.values, dim=4),
[[1, 0, 0, 0],
[0, 1, 0, 0],
[0, 0, 1, 0],
[0, 1, 0, 0]]
)
def test_one_hot(self):
np.testing.assert_array_equal(
one_hot(self.values),
[[1, 0, 0],
[0, 1, 0],
[0, 0, 1],
[0, 1, 0]]
)
def _get_bin_distributions(self, bin_indices):
"""Compute the distribution of instances within bins.
Parameters
----------
bin_indices : np.ndarray
An array with same shape as `x` but containing the bin index of the
instance.
Returns
-------
np.ndarray
A 2d array; the first dimension represents different bins, the
second - the counts of different target values.
"""
if self.target_var and self.target_var.is_discrete:
y = self.y
# TODO This probably also isn't the best handling of sparse data...
if sp.issparse(y):
y = np.squeeze(np.array(y.todense()))
# Since y can contain missing values, we need to filter them out as
# well as their corresponding `x` values
y_nan_mask = np.isnan(y)
y, bin_indices = y[~y_nan_mask], bin_indices[~y_nan_mask]
y = one_hot(y)
# In the event that y does not take up all the values and the
# largest discrete value does not appear at all, one hot encoding
# will produce too few columns. This causes problems, so we need to
# pad y with zeros to properly compute the distribution
if y.shape[1] != len(self.target_var.values):
n_missing_columns = len(self.target_var.values) - y.shape[1]
y = np.hstack((y, np.zeros((y.shape[0], n_missing_columns))))
bins = np.arange(self.n_bins)[:, np.newaxis]
mask = bin_indices == bins
distributions = np.zeros((self.n_bins, y.shape[1]))
for bin_idx in range(self.n_bins):
distributions[bin_idx] = y[mask[bin_idx]].sum(axis=0)
else:
distributions, _ = ut.bincount(bin_indices.astype(np.int64))
# To keep things consistent across different variable types, we
# want to return a 2d array where the first dim represent different
# bins, and the second the distributions.
distributions = distributions[:, np.newaxis]
return distributions
def _get_bin_distributions(self, bin_indices):
"""Compute the distribution of instances within bins.
Parameters
----------
bin_indices : np.ndarray
An array with same shape as `x` but containing the bin index of the
instance.
Returns
-------
np.ndarray
A 2d array; the first dimension represents different bins, the
second - the counts of different target values.
"""
if self.target_var and self.target_var.is_discrete:
y = self.y
# TODO This probably also isn't the best handling of sparse data...
if sp.issparse(y):
y = np.squeeze(np.array(y.todense()))
# Since y can contain missing values, we need to filter them out as
# well as their corresponding `x` values
y_nan_mask = np.isnan(y)
y, bin_indices = y[~y_nan_mask], bin_indices[~y_nan_mask]
y = one_hot(y)
# In the event that y does not take up all the values and the
# largest discrete value does not appear at all, one hot encoding
# will produce too few columns. This causes problems, so we need to
# pad y with zeros to properly compute the distribution
if y.shape[1] != len(self.target_var.values):
n_missing_columns = len(self.target_var.values) - y.shape[1]
y = np.hstack((y, np.zeros((y.shape[0], n_missing_columns))))
bins = np.arange(self.n_bins)[:, np.newaxis]
mask = bin_indices == bins
distributions = np.zeros((self.n_bins, y.shape[1]))
for bin_idx in range(self.n_bins):
distributions[bin_idx] = y[mask[bin_idx]].sum(axis=0)
else:
distributions, _ = ut.bincount(bin_indices.astype(np.int64))
# To keep things consistent across different variable types, we
# want to return a 2d array where the first dim represent different
# bins, and the second the distributions.
distributions = distributions[:, np.newaxis]
return distributions
def __call__(self, data, ret=Value):
def fix_dim(x):
return x[0] if one_d else x
if not 0 <= ret <= 2:
raise ValueError("invalid value of argument 'ret'")
if ret > 0 and any(v.is_continuous for v in self.domain.class_vars):
raise ValueError("cannot predict continuous distributions")
# Call the predictor
one_d = False
if isinstance(data, np.ndarray):
one_d = data.ndim == 1
prediction = self.predict(np.atleast_2d(data))
elif isinstance(data, scipy.sparse.csr.csr_matrix):
prediction = self.predict(data)
elif isinstance(data, (Table, Instance)):
if isinstance(data, Instance):
data = Table(data.domain, [data])
one_d = True
if data.domain != self.domain:
if self.original_domain.attributes != data.domain.attributes \
and data.X.size \
and not np.isnan(data.X).all():
data = data.transform(self.original_domain)
if np.isnan(data.X).all():
raise DomainTransformationError(
"domain transformation produced no defined values")
data = data.transform(self.domain)
prediction = self.predict_storage(data)
elif isinstance(data, (list, tuple)):
if not isinstance(data[0], (list, tuple)):
data = [data]
one_d = True
data = Table.from_list(self.original_domain, data)
data = data.transform(self.domain)
prediction = self.predict_storage(data)
else:
raise TypeError("Unrecognized argument (instance of '{}')"
.format(type(data).__name__))
# Parse the result into value and probs
multitarget = len(self.domain.class_vars) > 1
if isinstance(prediction, tuple):
value, probs = prediction
elif prediction.ndim == 1 + multitarget:
value, probs = prediction, None
elif prediction.ndim == 2 + multitarget:
value, probs = None, prediction
else:
raise TypeError("model returned a %i-dimensional array",
prediction.ndim)
# Ensure that we have what we need to return
if ret != Model.Probs and value is None:
value = np.argmax(probs, axis=-1)
if ret != Model.Value and probs is None:
if multitarget:
max_card = max(len(c.values)
for c in self.domain.class_vars)
probs = np.zeros(value.shape + (max_card,), float)
for i in range(len(self.domain.class_vars)):
probs[:, i, :] = one_hot(value[:, i])
else:
probs = one_hot(value)
if ret == Model.ValueProbs:
return fix_dim(value), fix_dim(probs)
else:
return fix_dim(probs)
# Return what we need to
if ret == Model.Probs:
return fix_dim(probs)
if isinstance(data, Instance) and not multitarget:
value = Value(self.domain.class_var, value[0])
if ret == Model.Value:
return fix_dim(value)
else: # ret == Model.ValueProbs
return fix_dim(value), fix_dim(probs)
def __call__(self, data, ret=Value):
if not 0 <= ret <= 2:
raise ValueError("invalid value of argument 'ret'")
if ret > 0 and any(v.is_continuous for v in self.domain.class_vars):
raise ValueError("cannot predict continuous distributions")
# Call the predictor
if isinstance(data, np.ndarray):
prediction = self.predict(np.atleast_2d(data))
elif isinstance(data, scipy.sparse.csr.csr_matrix):
prediction = self.predict(data)
elif isinstance(data, (Table, Instance)):
if isinstance(data, Instance):
data = Table(data.domain, [data])
if data.domain != self.domain:
data = data.transform(self.domain)
prediction = self.predict_storage(data)
elif isinstance(data, (list, tuple)):
if not isinstance(data[0], (list, tuple)):
data = [data]
data = Table(self.original_domain, data)
data = data.transform(self.domain)
prediction = self.predict_storage(data)
else:
raise TypeError("Unrecognized argument (instance of '{}')".format(
type(data).__name__))
# Parse the result into value and probs
multitarget = len(self.domain.class_vars) > 1
if isinstance(prediction, tuple):
value, probs = prediction
elif prediction.ndim == 1 + multitarget:
value, probs = prediction, None
elif prediction.ndim == 2 + multitarget:
value, probs = None, prediction
else:
raise TypeError("model returned a %i-dimensional array",
prediction.ndim)
# Ensure that we have what we need to return
if ret != Model.Probs and value is None:
value = np.argmax(probs, axis=-1)
if ret != Model.Value and probs is None:
if multitarget:
max_card = max(len(c.values) for c in self.domain.class_vars)
probs = np.zeros(value.shape + (max_card, ), float)
for i, cvar in enumerate(self.domain.class_vars):
probs[:, i, :] = one_hot(value[:, i])
else:
probs = one_hot(value)
if ret == Model.ValueProbs:
return value, probs
else:
return probs
# Return what we need to
if ret == Model.Probs:
return probs
if isinstance(data, Instance) and not multitarget:
value = Value(self.domain.class_var, value[0])
if ret == Model.Value:
return value
else: # ret == Model.ValueProbs
return value, probs
def __call__(self, data, ret=Value):
if not 0 <= ret <= 2:
raise ValueError("invalid value of argument 'ret'")
if ret > 0 and any(v.is_continuous for v in self.domain.class_vars):
raise ValueError("cannot predict continuous distributions")
# Call the predictor
if isinstance(data, np.ndarray):
prediction = self.predict(np.atleast_2d(data))
elif isinstance(data, scipy.sparse.csr.csr_matrix):
prediction = self.predict(data)
elif isinstance(data, Instance):
if data.domain != self.domain:
data = Instance(self.domain, data)
data = Table(data.domain, [data])
prediction = self.predict_storage(data)
elif isinstance(data, Table):
if data.domain != self.domain:
data = data.transform(self.domain)
prediction = self.predict_storage(data)
elif isinstance(data, (list, tuple)):
if not isinstance(data[0], (list, tuple)):
data = [data]
data = Table(self.original_domain, data)
data = data.transform(self.domain)
prediction = self.predict_storage(data)
else:
raise TypeError("Unrecognized argument (instance of '{}')"
.format(type(data).__name__))
# Parse the result into value and probs
multitarget = len(self.domain.class_vars) > 1
if isinstance(prediction, tuple):
value, probs = prediction
elif prediction.ndim == 1 + multitarget:
value, probs = prediction, None
elif prediction.ndim == 2 + multitarget:
value, probs = None, prediction
else:
raise TypeError("model returned a %i-dimensional array",
prediction.ndim)
# Ensure that we have what we need to return
if ret != Model.Probs and value is None:
value = np.argmax(probs, axis=-1)
if ret != Model.Value and probs is None:
if multitarget:
max_card = max(len(c.values)
for c in self.domain.class_vars)
probs = np.zeros(value.shape + (max_card,), float)
for i, cvar in enumerate(self.domain.class_vars):
probs[:, i, :] = one_hot(value[:, i])
else:
probs = one_hot(value)
if ret == Model.ValueProbs:
return value, probs
else:
return probs
# Return what we need to
if ret == Model.Probs:
return probs
if isinstance(data, Instance) and not multitarget:
value = Value(self.domain.class_var, value[0])
if ret == Model.Value:
return value
else: # ret == Model.ValueProbs
return value, probs
def test_one_hot(self):
np.testing.assert_equal(
one_hot([0, 1, 2, 1], int), [[1, 0, 0],
[0, 1, 0],
[0, 0, 1],
[0, 1, 0]])
def test_one_hot(self):
np.testing.assert_equal(
one_hot([0, 1, 2, 1], int), [[1, 0, 0],
[0, 1, 0],
[0, 0, 1],
[0, 1, 0]])
def test_one_hot(self):
np.testing.assert_equal(
one_hot([0, 1, 2, 1], int), [[1, 0, 0], [0, 1, 0], [0, 0, 1], [0, 1, 0]]
)
np.testing.assert_equal(one_hot([], int), np.zeros((0, 0), dtype=int))
def test_dim_too_low(self):
with self.assertRaises(ValueError):
one_hot(self.values, dim=2)
def test_dtype(self):
res = one_hot(self.values)
self.assertEqual(res.dtype, float)
res = one_hot(self.values, dtype=int)
self.assertEqual(res.dtype, int)
