def test__parabolic_crossentropy(self):
"""when"""
categories = 11
loss_function = parabolic_crossentropy(categories, 1)
"""then"""
for truth in range(categories):
losses = []
for prediction in range(categories):
loss = K.eval(
loss_function(K.variable(one_hot(truth, categories)),
K.variable(one_hot(prediction, categories))))
losses.append(loss)
# all predictions left of truth need to increase
for i in range(1, truth):
self.assertGreater(losses[i - 1], losses[i])
# right of truth need to decrease
for i in range(truth, categories - 1):
self.assertLess(losses[i], losses[i + 1])
if truth > 0 and truth < categories - 1:
if truth > categories / 2:
# right tail:
self.assertGreater(losses[truth - 1], losses[truth + 1])
else:
# left tail
self.assertGreater(losses[truth + 1], losses[truth - 1])
def test__tailed_categorical_crossentropy(self):
"""when"""
categories = 11
loss = tailed_categorical_crossentropy(categories, 1)
"""then"""
l = K.eval(loss(K.variable(one_hot(3, 11)), K.variable(one_hot(6,
11))))
pickle
"""then"""
self.assertGreater(l, 0)
save_object(loss, '/tmp/test__tailed_categorical_crossentropy.dill')
def ta_one_hot_categories(df: _pd.DataFrame):
"""
Take a category column or a column of integers and turn them into a one hot encoded data frame
:param df: a series or a data frame which has category columns or integer columns
:return: a multi index data frame with one hot encoded integer columns
note: can be empty
"""
df = df.to_frame() if isinstance(df, _pd.Series) else df
res = None
for col in df.columns:
if hasattr(df[col], "cat"):
categories = [str(cat) for cat in df[col].cat.categories]
df_of_categories = index_of_categories(df[col])
elif df[col].dtype.kind in 'iu':
categories = sorted(set(df[col].values))
df_of_categories = df
else:
continue
number_of_categories = len(categories)
ohdf = df_of_categories[[col]].apply(lambda r: util.one_hot(r, number_of_categories), axis=1, result_type='expand')
ohdf.columns = _pd.MultiIndex.from_product([[col], categories])
res = ohdf if res is None else res.join(ohdf)
if res is None:
_log.warning(f'non of the {df.columns} are of type category index or integer value!\n'
f'You might want to call df.ta_one_hot_categories(df.ta_bucketize(3))')
return res
def test__differentiable_argmax(self):
"""given"""
args = 10
argmax = DifferentiableArgmax(args)
"""when"""
res = np.array([
K.eval(argmax(K.variable(one_hot(i, args)))) for i in range(args)
])
"""then"""
print(res)
np.testing.assert_array_almost_equal(res, np.arange(0, args))
def convoluted(df: _pd.Series) -> _np.ndarray:
min = df.min()
max = df.max()
interval_index = _np.linspace(min, max, buckets)
indexes = _np.digitize(df, interval_index) - 1
_np.array([util.one_hot(index, buckets) for index in indexes])
def test_normal_penalized_crossentropy(self):
"""when"""
loss = normal_penalized_crossentropy(11)
"""then"""
for i in range(11):
self.assertLess(
K.eval(
loss(K.variable(one_hot(i, 11)),
K.variable(one_hot(i, 11)))), 0.00001)
self.assertLess(
K.eval(loss(K.variable(one_hot(7, 11)), K.variable(one_hot(8,
11)))),
K.eval(loss(K.variable(one_hot(7, 11)), K.variable(one_hot(6,
11)))))
self.assertLess(
K.eval(loss(K.variable(one_hot(6, 11)), K.variable(one_hot(7,
11)))),
K.eval(loss(K.variable(one_hot(6, 11)), K.variable(one_hot(5,
11)))))
self.assertLess(
K.eval(loss(K.variable(one_hot(3, 11)), K.variable(one_hot(2,
11)))),
K.eval(loss(K.variable(one_hot(3, 11)), K.variable(one_hot(4,
11)))))