def __init__(self, mean=None, std=None, with_std=True, preprocess=None):
if mean is not None:
self._in_mean = (mean, ) if is_scalar(mean) else tuple(mean)
else: # mean is None
self._in_mean = None
if std is not None:
self._in_std = (std, ) if is_scalar(std) else tuple(std)
else: # std is None
self._in_std = None
# Input validation
if with_std is True:
if (mean is None and std is not None) or \
(mean is not None and std is None) or \
(mean is not None and std is not None and
len(self._in_mean) != len(self._in_std)):
raise ValueError("if `with_std` is True, `mean` and `std` "
"should be both None or both scalar or "
"both tuple of the same length")
self._mean = None
self._std = None
self._with_std = with_std
# Properties of the linear normalizer
self._w = None
self._b = None
super(CNormalizerMeanStd, self).__init__(preprocess=preprocess)
def test_init_reshape(self):
"""Test CArray reshape during initialization."""
arrays = [[[2, 3], [22, 33]], [2, 3], [[2], [3]], 3]
for a in arrays:
for sparse in (False, True):
out_def = CArray(a)
size = out_def.size # Expected size
in_shape = out_def.shape # Expected input_shape
for shape in [size, (size, ), (1, size), (size, 1)]:
out_res = CArray(a, tosparse=sparse, shape=shape)
# Resulting shape will always be (1, n) for sparse
if is_scalar(shape):
shape = (1, shape) if out_res.issparse else (shape, )
if out_res.issparse and len(shape) < 2:
shape = (1, shape[0])
self.logger.info("Expected 'shape' {:}, got {:}".format(
shape, out_res.shape))
self.assertEqual(out_res.shape, shape)
# The input_shape should not be altered by reshaping
self.logger.info(
"Expected 'input_shape' {:}, got {:}".format(
in_shape, out_res.input_shape))
self.assertEqual(out_res.input_shape, in_shape)
with self.assertRaises(ValueError):
# Shape with wrong size, expect error
CArray(a, tosparse=sparse, shape=(2, size))
def check_init_builtin(totest_list):
for totest_elem in totest_list:
for tosparse in [False, True]:
init_array = CArray(totest_elem, tosparse=tosparse)
self.assertTrue(init_array.issparse == tosparse)
if is_list_of_lists(totest_elem):
self.assertTrue(
init_array.shape[0] == len(totest_elem))
self.assertTrue(
init_array.shape[1] == len(totest_elem[0]))
elif is_list(totest_elem):
if init_array.issparse is True:
self.assertTrue(
init_array.shape[1] == len(totest_elem))
elif init_array.isdense is True:
self.assertTrue(init_array.ndim == 1)
self.assertTrue(
init_array.shape[0] == len(totest_elem))
elif is_scalar(totest_elem) or is_bool(totest_elem):
self.assertTrue(init_array.size == 1)
else:
raise TypeError("test_init_builtin should not be used "
"to test {:}".format(
type(totest_elem)))
def _cmp_kernel(self, k_fun, a1, a2):
k = k_fun(a1, a2)
if isinstance(k, CArray):
self.logger.info("k shape with inputs {:} {:} is: {:}"
"".format(a1.shape, a2.shape, k.shape))
self.assertEqual(k.shape, (CArray(a1).atleast_2d().shape[0],
CArray(a2).atleast_2d().shape[0]))
else:
self.assertTrue(is_scalar(k))
def _test_fun_result(self, fun, x, res_expected):
"""Test if function returns the correct value.
Parameters
----------
fun : CFunction
x : CArray
res_expected : scalar
"""
self.logger.info("Checking value of {:} @ {:}".format(
fun.class_type, x))
res = fun.fun(x)
self.logger.info("Correct result: {:}".format(res_expected))
self.logger.info("Function result: {:}".format(res))
self.assertTrue(is_scalar(res))
self.assertAlmostEqual(res_expected, res, places=4)
def check_init_builtin(totest_elem):
for tosparse in [False, True]:
init_array = CArray(totest_elem, tosparse=tosparse)
self.assertEqual(init_array.issparse, tosparse)
if is_list_of_lists(totest_elem):
if not is_list_of_lists(totest_elem[0]):
self.assertEqual(
init_array.shape[0], len(totest_elem))
self.assertEqual(
init_array.shape[1], len(totest_elem[0]))
else: # N-Dimensional input
in_shape = init_array.input_shape
self.assertEqual(in_shape[0], len(totest_elem))
self.assertEqual(in_shape[1], len(totest_elem[0]))
self.assertEqual(
init_array.shape[0], len(totest_elem))
self.assertEqual(
init_array.shape[1], sum(in_shape[1:]))
elif is_list(totest_elem):
if init_array.issparse is True:
self.assertEqual(
init_array.shape[1], len(totest_elem))
elif init_array.isdense is True:
self.assertTrue(init_array.ndim == 1)
self.assertEqual(
init_array.shape[0], len(totest_elem))
self.assertEqual(
init_array.input_shape, (len(totest_elem), ))
elif is_scalar(totest_elem) or is_bool(totest_elem):
self.assertEqual(init_array.size, 1)
self.assertEqual(init_array.input_shape, (1, ))
else:
raise TypeError(
"test_init_builtin should not be used "
"to test {:}".format(type(totest_elem)))
def _get_best_params(self, res_vect, params, params_matrix, pick='first'):
"""Returns the best parameters given input performance scores.
The best parameters have the closest associated performance score
to the metric's best value.
Parameters
----------
res_vect : CArray
Array with the performance results associated
to each parameters combination.
params : dict
Dictionary with the parameters to be evaluated.
params_matrix : CArray
Indices of each combination of parameters to evaluate.
pick : {'first', 'last', 'random'}, optional
Defines which of the best parameters set pick.
Usually, 'first' (default) correspond to the smallest
parameters while 'last' correspond to the biggest.
The order is consistent to the parameters dict passed as input.
Returns
-------
best_params_dict : dict
Dictionary with the parameters that have obtained
the best performance score.
best_value : any
Performance value associated with the best parameters.
"""
if not is_scalar(self.metric.best_value):
raise TypeError(
"XVal only works with metric with the best value as scalar")
# Get the index of the results closest to the best value
diff = abs(res_vect - self.metric.best_value)
best_params_list = []
best_score = []
# Get the best parameters for each binary classifier
for i in range(res_vect.shape[1]):
# diff has one row for each parameters combination and
# one column for each binary classifier
condidates_idx = diff[:, i].find_2d(diff[:, i] == diff[:,
i].min())[0]
# Get the value of the result closest to the best value
best_score.append(res_vect[condidates_idx[0], i])
# Get the index of the corresponding parameters
best_params_idx = params_matrix[condidates_idx, :]
# Build the list of candidate parameters for binary clf
clf_best_params_list = []
for c_idx in range(best_params_idx.shape[0]):
# For each candidate get corresponding parameters
best_params_dict = dict()
for j, par in enumerate(params):
par_idx = best_params_idx[c_idx, j].item()
best_params_dict[par] = params[par][par_idx]
clf_best_params_list.append(best_params_dict)
# Chose which candidate parameters assign to classifier
if pick == 'first': # Usually the smallest
clf_best_params_dict = clf_best_params_list[0]
elif pick == 'last': # Usually the biggest
clf_best_params_dict = clf_best_params_list[-1]
elif pick == 'random':
import random
clf_best_params_dict = random.choice(clf_best_params_list)
else:
raise ValueError("pick strategy '{:}' not known".format(pick))
best_params_list.append(clf_best_params_dict)
# For each param, built the tuple of the best value for each binary clf
best_params_dict = dict()
for par in params:
this_param_list = []
for params_dict in best_params_list:
this_param_list.append(params_dict[par])
best_params_dict[par] = tuple(this_param_list)
return best_params_dict, best_score
def _check_repeat(array):
self.logger.info("Array:\n{:}".format(array))
for axis in (None, 0, 1):
if axis is None or array.ndim < 2:
repeats_add = CArray.randint(2, shape=array.size)
elif axis == 0:
repeats_add = CArray.randint(2, shape=array.shape[0])
elif axis == 1:
repeats_add = CArray.randint(2, shape=array.shape[1])
else:
repeats_add = None
for repeats in (0, 1, 2, repeats_add):
with self.assertRaises(TypeError):
array.repeat(repeats=np.array([1, 2]), axis=axis)
if axis == 1 and array.ndim < 2:
# No columns to repeat
with self.assertRaises(ValueError):
array.repeat(repeats=repeats, axis=axis)
continue
res = array.repeat(repeats=repeats, axis=axis)
self.logger.info("array.repeat({:}, axis={:}):"
"\n{:}".format(repeats, axis, res))
self.assertIsInstance(res, CArray)
self.assertEqual(res.isdense, array.isdense)
self.assertEqual(res.issparse, array.issparse)
self.assertEqual(res.dtype, array.dtype)
if axis is None or array.ndim < 2:
# A flat array is always returned
if is_scalar(repeats):
repeats_mul = array.size * repeats
else:
repeats_mul = repeats.sum()
self.assertEqual(res.shape, (repeats_mul, ))
elif axis == 0:
if is_scalar(repeats):
repeats_mul = array.shape[0] * repeats
else:
repeats_mul = repeats.sum()
self.assertEqual(res.shape,
(repeats_mul, array.shape[1]))
elif axis == 1:
if is_scalar(repeats):
repeats_mul = array.shape[1] * repeats
else:
repeats_mul = repeats.sum()
self.assertEqual(res.shape,
(array.shape[0], repeats_mul))
if is_scalar(repeats):
repeats_size = array.size * repeats
else:
if axis is None or array.ndim < 2:
repeats_size = repeats.sum()
elif axis == 0:
repeats_size = repeats.sum() * array.shape[1]
elif axis == 1:
repeats_size = repeats.sum() * array.shape[0]
else:
repeats_size = None
self.assertEqual(res.size, repeats_size)
if not is_scalar(repeats):
repeats = repeats.tondarray()
np_res = array.tondarray().repeat(repeats=repeats,
axis=axis)
self.assertFalse((res.tondarray() != np_res).any())
def _get_best_params(self, res_vect, params, params_matrix, pick='first'):
"""Returns the best parameters given input performance scores.
The best parameters have the closest associated performance score
to the metric's best value.
Parameters
----------
res_vect : CArray
Array with the performance results associated
to each parameters combination.
params : dict
Dictionary with the parameters to be evaluated.
params_matrix : CArray
Indices of each combination of parameters to evaluate.
pick : {'first', 'last', 'random'}, optional
Defines which of the best parameters set pick.
Usually, 'first' (default) correspond to the smallest
parameters while 'last' correspond to the biggest.
The order is consistent to the parameters dict passed as input.
Returns
-------
best_params_dict : dict
Dictionary with the parameters that have obtained
the best performance score.
best_value : any
Performance value associated with the best parameters.
"""
if not is_scalar(self.metric.best_value):
raise TypeError(
"XVal only works with metric with the best value as scalar")
# Get the index of the results closest to the best value
diff = abs(res_vect - self.metric.best_value)
condidates_idx = diff.find(diff == diff.nanmin())
if len(condidates_idx) < 1:
raise ValueError("all metric outputs are equal to Nan!")
# Get the value of the result closest to the best value
best_score = res_vect[condidates_idx[0]]
# Get the index of the corresponding parameters
best_params_idx = params_matrix[condidates_idx, :]
# Build the list of candidate parameters
best_params_list = []
for c_idx in range(best_params_idx.shape[0]):
# For each candidate get corresponding parameters
best_params_dict = dict()
for j, par in enumerate(params):
value_idx = best_params_idx[c_idx, j].item()
best_params_dict[par] = params[par][value_idx]
best_params_list.append(best_params_dict)
# Chose which candidate parameters assign to classifier
if pick == 'first': # Usually the smallest
best_params_dict = best_params_list[0]
elif pick == 'last': # Usually the biggest
best_params_dict = best_params_list[-1]
elif pick == 'random':
import random
best_params_dict = random.choice(best_params_list)
else:
raise ValueError("pick strategy '{:}' not known".format(pick))
return best_params_dict, best_score
