def test_lstm_symmetric_checks_dims(self):
"""
Test that lstm_symmetric validates parameter requirements
"""
# Raise an error if empty dimension and function layers defined.
with self.assertRaises(ValueError):
lstm_symmetric(4, dims=(), funcs=())
# Ensure failure with default 3 dimension layers when 2 function layers passed.
with self.assertRaises(ValueError):
lstm_symmetric(4, funcs=("tanh", "tanh"))
def test_lstm_symmetric_basic(n_features, n_features_out):
"""
Tests that lstm_symmetric implements the correct parameters
"""
model = lstm_symmetric(
n_features=n_features,
n_features_out=n_features_out,
lookback_window=3,
dims=(4, 3, 2, 1),
funcs=("relu", "relu", "tanh", "tanh"),
out_func="linear",
optimizer="sgd",
optimizer_kwargs={"lr": 0.01},
loss="mse",
)
# Ensure that the input dimension to Keras model matches the number of features.
assert model.layers[0].input_shape[2] == n_features
# Ensure that the dimension of each encoding layer matches the expected dimension.
assert [model.layers[i].input_shape[2] for i in range(1, 5)] == [4, 3, 2, 1]
# Ensure that the dimension of each decoding layer (excluding last decoding layer)
# matches the expected dimension.
assert [model.layers[i].input_shape[2] for i in range(5, 8)] == [1, 2, 3]
# Ensure that the dimension of last decoding layer matches the expected dimension.
assert model.layers[8].input_shape[1] == 4
# Ensure activation functions in the encoding part (layers 0-3)
# match expected activation functions.
assert [model.layers[i].activation.__name__ for i in range(0, 4)] == [
"relu",
"relu",
"tanh",
"tanh",
]
# Ensure activation functions in the decoding part (layers 4-7)
# match expected activation functions.
assert [model.layers[i].activation.__name__ for i in range(4, 8)] == [
"tanh",
"tanh",
"relu",
"relu",
]
# Ensure activation function for the output layer matches expected activation function.
assert model.layers[8].activation.__name__ == "linear"
# Assert that the expected Keras optimizer is used
assert model.optimizer.__class__ == optimizers.SGD
# Assert equality of difference up to 7 decimal places
# Note that AlmostEquality is used as Keras can use a value approximately equal
# to the given learning rate rather than the exact value.
assert np.isclose((K.eval(model.optimizer.lr),), (0.01,))
# Assert that the correct loss function is used.
assert model.loss == "mse"
def test_lstm_symmetric_basic(n_features, n_features_out):
"""
Tests that lstm_symmetric implements the correct parameters
"""
model = lstm_symmetric(
n_features=n_features,
n_features_out=n_features_out,
lookback_window=3,
dims=(4, 3, 2, 1),
funcs=("relu", "relu", "tanh", "tanh"),
out_func="linear",
optimizer="SGD",
optimizer_kwargs={"lr": 0.01},
loss="mse",
)
# Ensure that the input dimension to Keras model matches the number of features.
assert model.layers[0].input_shape[2] == n_features
# Ensure that the dimension of each encoding layer matches the expected dimension.
assert [model.layers[i].input_shape[2]
for i in range(1, 5)] == [4, 3, 2, 1]
# Ensure that the dimension of each decoding layer (excluding last decoding layer)
# matches the expected dimension.
assert [model.layers[i].input_shape[2] for i in range(5, 8)] == [1, 2, 3]
# Ensure that the dimension of last decoding layer matches the expected dimension.
assert model.layers[8].input_shape[1] == 4
# Ensure activation functions in the encoding part (layers 0-3)
# match expected activation functions.
assert [model.layers[i].activation.__name__ for i in range(0, 4)] == [
"relu",
"relu",
"tanh",
"tanh",
]
# Ensure activation functions in the decoding part (layers 4-7)
# match expected activation functions.
assert [model.layers[i].activation.__name__ for i in range(4, 8)] == [
"tanh",
"tanh",
"relu",
"relu",
]
# Ensure activation function for the output layer matches expected activation function.
assert model.layers[8].activation.__name__ == "linear"
# Assert that the expected Keras optimizer is used
assert model.optimizer.__class__ == optimizers.SGD
# Assert that the correct loss function is used.
assert model.loss == "mse"
def test_lstm_defaults(self):
"""
Test that models with all defaults are created without failure and are equal
"""
# Create models with default parameters
base = lstm_model(4)
symmetric = lstm_symmetric(4)
hourglass = lstm_hourglass(4)
# Ensure LSTM model layers are equal to base model layers
for i in range(len(base.layers)):
# Rename layers so as to not fail on names, only configuration
config = base.layers[i].get_config().update({"name": "test"})
symmetric_config = symmetric.layers[i].get_config().update(
{"name": "test"})
hourglass_config = hourglass.layers[i].get_config().update(
{"name": "test"})
assert config == symmetric_config
assert config == hourglass_config
def test_lstm_symmetric_checks_dims(self):
"""
Test that lstm_symmetric validates parameter requirements
"""
with self.assertRaises(ValueError):
lstm_symmetric(4, dims=[], funcs=[])