def test_binary_crossentropy():
fn = 'binary_crossentropy'
y_pred = np.array([0.28, 0.68])
y_true = np.array([0, 1])
Loss = losses.get(fn)
Input = InputLayer(shape=[2])
Input.ForwardPropagation(feed=y_pred)
Input.outbound.append(Loss)
Loss.inbound.append(Input)
loss_value = Loss.ForwardPropagation(y_true=y_true)
# print(loss_value)
assert 0.357 == round(loss_value, 3)
def test_categorical_crossentropy():
fn = 'categorical_crossentropy'
y_pred = np.array([[0.28, 0.68, 0.04], [0.03, 0.95, 0.02]])
y_true = np.array([[0, 1, 0], [0, 1, 0]])
Loss = losses.get(fn)
Input = InputLayer(shape=[3])
Input.ForwardPropagation(feed=y_pred)
Input.outbound.append(Loss)
Loss.inbound.append(Input)
loss_value = Loss.ForwardPropagation(y_true=y_true)
# print(loss_value)
assert 0.44 / y_true.shape[0] == round(loss_value, 2)
def test_softmax_categorical_crossentropy():
fn = 'softmax_categorical_crossentropy'
y_pred = np.array([[0.9, 4.2, 0.3], [0.9, 4.2, 0.3]])
y_true = np.array([[0., 1., 0.], [0., 1., 0.]])
Loss = losses.get(fn)
assert isinstance(Loss, losses.SoftmaxCategoricalCrossentropy)
Input = InputLayer(shape=[3])
Input.ForwardPropagation(feed=y_pred)
Input.outbound.append(Loss)
Loss.inbound.append(Input)
loss_value = Loss.ForwardPropagation(y_true=y_true)
Loss.BackwardPropagation()
# print(loss_value)
assert round(loss_value, 3) == round(0.027776516853274126, 3)
def test_mae():
fn = 'mae'
y_pred = np.array([[3., 4.], [7., 6.]])
y_true = np.array([[2, 1], [0, 0]])
Loss = losses.get(fn)
Input = InputLayer(shape=[2])
Input.ForwardPropagation(feed=y_pred)
Input.outbound.append(Loss)
Loss.inbound.append(Input)
loss_value = Loss.ForwardPropagation(y_true=y_true)
# print(loss_value)
# Todo check this test method
assert loss_value[0] == 4 / y_true.shape[0] and loss_value[
1] == 4.5 / y_true.shape[0]
def test_mse():
fn = 'mse'
y_pred = np.array([[1, 2], [3, 4]])
y_true = np.zeros((2, 2))
Loss = losses.get(fn)
Input = InputLayer(shape=[2])
Input.ForwardPropagation(feed=y_pred)
Input.outbound.append(Loss)
Loss.inbound.append(Input)
loss_value = Loss.ForwardPropagation(y_true=y_true)
# print(loss_value)
# Todo check this test method
assert loss_value[0] == 5 / y_true.shape[0] and loss_value[
1] == 10 / y_true.shape[0]
def test_none():
try:
none = losses.get(None)
except Exception as e:
assert e.args[0] == 'You should specify a loss function'