def test_sum_reduction(self):
mae_obj = losses.MeanAbsoluteError(
reduction=losses_utils.Reduction.SUM)
y_true = K.constant([1, 9, 2, -5, -2, 6], shape=(2, 3))
y_pred = K.constant([4, 8, 12, 8, 1, 3], shape=(2, 3))
loss = mae_obj(y_true, y_pred, sample_weight=2.3)
assert np.isclose(K.eval(loss), 25.29999, atol=1e-3)
def test_invalid_sample_weight(self):
mae_obj = losses.MeanAbsoluteError()
y_true = K.constant([1, 9, 2, -5, -2, 6], shape=(2, 3, 1))
y_pred = K.constant([4, 8, 12, 8, 1, 3], shape=(2, 3, 1))
sample_weight = K.constant([3, 6, 5, 0], shape=(2, 2))
with pytest.raises(Exception):
mae_obj(y_true, y_pred, sample_weight=sample_weight)
def test_no_reduction(self):
mae_obj = losses.MeanAbsoluteError(
reduction=losses_utils.Reduction.NONE)
y_true = K.constant([1, 9, 2, -5, -2, 6], shape=(2, 3))
y_pred = K.constant([4, 8, 12, 8, 1, 3], shape=(2, 3))
loss = mae_obj(y_true, y_pred, sample_weight=2.3)
assert np.allclose(K.eval(loss), [10.7333, 14.5666], atol=1e-3)
def test_sample_weighted(self):
mae_obj = losses.MeanAbsoluteError()
y_true = K.constant([1, 9, 2, -5, -2, 6], shape=(2, 3))
y_pred = K.constant([4, 8, 12, 8, 1, 3], shape=(2, 3))
sample_weight = K.constant([1.2, 3.4], shape=(2, 1))
loss = mae_obj(y_true, y_pred, sample_weight=sample_weight)
assert np.isclose(K.eval(loss), 81.4 / 6, atol=1e-3)
def test_timestep_weighted(self):
mae_obj = losses.MeanAbsoluteError()
y_true = K.constant([1, 9, 2, -5, -2, 6], shape=(2, 3, 1))
y_pred = K.constant([4, 8, 12, 8, 1, 3], shape=(2, 3, 1))
sample_weight = K.constant([3, 6, 5, 0, 4, 2], shape=(2, 3))
loss = mae_obj(y_true, y_pred, sample_weight=sample_weight)
assert np.isclose(K.eval(loss), 13.833, atol=1e-3)
def train(self):
# Getting the untrained model
empty_model = self.model
# Initiating the optimizer
optimizer = keras.optimizers.Adam(learning_rate=self.lr)
# Compiling the model
empty_model.compile(loss=losses.MeanAbsoluteError(),
optimizer=optimizer)
if (self.Xval is not None) & (self.Yval is not None) & (self.Xauxval
is not None):
history = empty_model.fit(
[self.X, self.Xaux],
self.Y,
epochs=self.epochs,
batch_size=self.batch,
validation_data=([self.Xval, self.Xauxval], self.Yval),
shuffle=False)
else:
history = empty_model.fit([self.X, self.Xaux],
self.Y,
epochs=self.epochs,
batch_size=self.batch,
shuffle=False)
# Saving to original model attribute in the class
self.model = empty_model
# Returning the training history
return history
def _get_multi_io_model():
inp_1 = layers.Input(shape=(1, ), name='input_1')
inp_2 = layers.Input(shape=(1, ), name='input_2')
d = testing_utils.Bias(name='output')
out_1 = d(inp_1)
out_2 = d(inp_2)
return keras.Model([inp_1, inp_2], [out_1, out_2])
@keras_parameterized.run_all_keras_modes
@parameterized.named_parameters([
dict(testcase_name='string', value='mae'),
dict(testcase_name='built_in_fn', value=losses.mae),
dict(testcase_name='built_in_class', value=losses.MeanAbsoluteError()),
dict(testcase_name='custom_fn', value=my_mae),
dict(testcase_name='custom_class', value=MyMeanAbsoluteError()),
dict(testcase_name='list_of_strings', value=['mae', 'mae']),
dict(testcase_name='list_of_built_in_fns', value=[losses.mae, losses.mae]),
dict(testcase_name='list_of_built_in_classes',
value=[losses.MeanAbsoluteError(),
losses.MeanAbsoluteError()]),
dict(testcase_name='list_of_custom_fns', value=[my_mae, my_mae]),
dict(testcase_name='list_of_custom_classes',
value=[MyMeanAbsoluteError(),
MyMeanAbsoluteError()]),
dict(testcase_name='dict_of_string',
value={
'output': 'mae',
'output_1': 'mae',
def test_scalar_weighted(self):
mae_obj = losses.MeanAbsoluteError()
y_true = K.constant([1, 9, 2, -5, -2, 6], shape=(2, 3))
y_pred = K.constant([4, 8, 12, 8, 1, 3], shape=(2, 3))
loss = mae_obj(y_true, y_pred, sample_weight=2.3)
assert np.isclose(K.eval(loss), 12.65, atol=1e-3)
def test_all_correct_unweighted(self):
mae_obj = losses.MeanAbsoluteError()
y_true = K.constant([4, 8, 12, 8, 1, 3], shape=(2, 3))
loss = mae_obj(y_true, y_true)
assert np.isclose(K.eval(loss), 0.0, atol=1e-3)
def test_config(self):
mae_obj = losses.MeanAbsoluteError(
reduction=losses_utils.Reduction.SUM, name='mae_1')
assert mae_obj.name == 'mae_1'
assert mae_obj.reduction == losses_utils.Reduction.SUM
def _get_multi_io_model():
inp_1 = layers.Input(shape=(1,), name="input_1")
inp_2 = layers.Input(shape=(1,), name="input_2")
d = test_utils.Bias(name="output")
out_1 = d(inp_1)
out_2 = d(inp_2)
return keras.Model([inp_1, inp_2], [out_1, out_2])
@test_combinations.run_all_keras_modes
@parameterized.named_parameters(
[
dict(testcase_name="string", value="mae"),
dict(testcase_name="built_in_fn", value=losses.mae),
dict(testcase_name="built_in_class", value=losses.MeanAbsoluteError()),
dict(testcase_name="custom_fn", value=my_mae),
dict(testcase_name="custom_class", value=MyMeanAbsoluteError()),
dict(testcase_name="list_of_strings", value=["mae", "mae"]),
dict(
testcase_name="list_of_built_in_fns", value=[losses.mae, losses.mae]
),
dict(
testcase_name="list_of_built_in_classes",
value=[losses.MeanAbsoluteError(), losses.MeanAbsoluteError()],
),
dict(testcase_name="list_of_custom_fns", value=[my_mae, my_mae]),
dict(
testcase_name="list_of_custom_classes",
value=[MyMeanAbsoluteError(), MyMeanAbsoluteError()],
),
