def test_hourglass_calc_dims_check_dims(self):
"""
Test that hourglass_calc_dims implements the correct dimensions
"""
dims = hourglass_calc_dims(0.2, 4, 5)
self.assertEqual(dims, [4, 3, 2, 1])
dims = hourglass_calc_dims(0.5, 3, 10)
self.assertEqual(dims, [8, 7, 5])
dims = hourglass_calc_dims(0.5, 3, 3)
self.assertEqual(dims, [3, 2, 2])
dims = hourglass_calc_dims(0.3, 3, 10)
self.assertEqual(dims, [8, 5, 3])
dims = hourglass_calc_dims(1, 3, 10)
self.assertEqual(dims, [10, 10, 10])
dims = hourglass_calc_dims(0, 3, 100000)
self.assertEqual(dims, [66667, 33334, 1])
def feedforward_hourglass(
n_features: int,
encoding_layers: int = 3,
compression_factor: float = 0.5,
func: str = "tanh",
**kwargs,
) -> keras.models.Sequential:
"""
Builds an hourglass shaped neural network, with decreasing number of neurons
as one gets deeper into the encoder network and increasing number
of neurons as one gets out of the decoder network.
Parameters
----------
n_features: int
Number of input and output neurons
encoding_layers: int
Number of layers from the input layer (exclusive) to the
narrowest layer (inclusive). Must be > 0. The total nr of layers
including input and output layer will be 2*encoding_layers + 1.
compression_factor: float
How small the smallest layer is as a ratio of n_features
(smallest layer is rounded up to nearest integer). Must satisfy
0 <= compression_factor <= 1.
func: str
Activation function for the internal layers
Notes
-----
The resulting model will look like this when n_features = 10, encoding_layers= 3,
and compression_factor = 0.3::
* * * * * * * * * *
* * * * * * * *
* * * * *
* * *
* * *
* * * * *
* * * * * * * *
* * * * * * * * * *
Returns
-------
keras.models.Sequential
Examples
--------
>>> model = feedforward_hourglass(10)
>>> len(model.layers)
7
>>> [model.layers[i].units for i in range(len(model.layers))]
[8, 7, 5, 5, 7, 8, 10]
>>> model = feedforward_hourglass(5)
>>> [model.layers[i].units for i in range(len(model.layers))]
[4, 4, 3, 3, 4, 4, 5]
>>> model = feedforward_hourglass(10, compression_factor=0.2)
>>> [model.layers[i].units for i in range(len(model.layers))]
[7, 5, 2, 2, 5, 7, 10]
>>> model = feedforward_hourglass(10, encoding_layers=1)
>>> [model.layers[i].units for i in range(len(model.layers))]
[5, 5, 10]
"""
dims = hourglass_calc_dims(compression_factor, encoding_layers, n_features)
return feedforward_symmetric(n_features, dims, [func] * len(dims),
**kwargs)
def lstm_hourglass(
n_features: int,
lookback_window: int = 1,
encoding_layers: int = 3,
compression_factor: float = 0.5,
func: str = "relu",
out_func: str = "linear",
optimizer: Union[str, Optimizer] = "adam",
optimizer_kwargs: Dict[str, Any] = dict(),
loss: str = "mse",
**kwargs,
) -> keras.models.Sequential:
"""
Builds an hourglass shaped neural network, with decreasing number of neurons
as one gets deeper into the encoder network and increasing number
of neurons as one gets out of the decoder network.
Parameters
----------
n_features: int
Number of input and output neurons
encoding_layers: int
Number of layers from the input layer (exclusive) to the
narrowest layer (inclusive). Must be > 0. The total nr of layers
including input and output layer will be 2*encoding_layers + 1.
compression_factor: float
How small the smallest layer is as a ratio of n_features
(smallest layer is rounded up to nearest integer). Must satisfy
0 <= compression_factor <= 1.
func: str
Activation function for the internal layers
out_func: str
Activation function for the output Dense layer.
optimizer: Union[str, Optimizer]
If str then the name of the optimizer must be provided (e.x. "adam").
The arguments of the optimizer can be supplied in optimization_kwargs.
If a Keras optimizer call the instance of the respective
class (e.x. Adam(lr=0.01,beta_1=0.9, beta_2=0.999)). If no arguments are
provided Keras default values will be set.
optimizer_kwargs: Dict[str, Any]
The arguments for the chosen optimizer. If not provided Keras'
default values will be used.
loss: str
Keras' supported loss functions (e.x. "mse", "MSE", "mean_squared_error"
for mean squared error,
"mae", "MAE", "mean_absolute_error"
for mean absolute error,
for other supported loss functions
refer to https://keras.io/losses/).
Returns
-------
keras.models.Sequential
Examples
--------
>>> model = lstm_hourglass(10)
>>> len(model.layers)
7
>>> [model.layers[i].units for i in range(len(model.layers))]
[8, 7, 5, 5, 7, 8, 10]
>>> model = lstm_hourglass(5)
>>> [model.layers[i].units for i in range(len(model.layers))]
[4, 4, 3, 3, 4, 4, 5]
>>> model = lstm_hourglass(10, compression_factor=0.2)
>>> [model.layers[i].units for i in range(len(model.layers))]
[7, 5, 2, 2, 5, 7, 10]
>>> model = lstm_hourglass(10, encoding_layers=1)
>>> [model.layers[i].units for i in range(len(model.layers))]
[5, 5, 10]
"""
dims = hourglass_calc_dims(compression_factor, encoding_layers, n_features)
return lstm_symmetric(
n_features,
lookback_window,
dims,
[func] * len(dims),
out_func,
optimizer,
optimizer_kwargs,
loss,
**kwargs,
)