def test_maxout_2d(self):
out, out_shp = A.Maxout(self.input_zeros,
self.maxout_size,
self.input_size,
type='maxout',
dimension=2)
self.assertTrue(out_shp, self.input_size)
def _activate(x, activation, input_size, verbose=2, **kwargs):
"""
This function is used to produce activations for the outputs of any type of layer.
Args:
x: input tensor.
activation: Refer to the ``add_layer`` method.
input_size: supply the size of the inputs.
verbose: typical toolbox verbose
dimension: used only for maxout. Give the dimension on which to maxout.
Returns:
tuple: ``(out, out_shp)``
"""
if verbose >= 3:
print("... Setting up activations")
# some activations like maxouts are supplied with special support parameters
if type(activation) is tuple:
if activation[0] == 'maxout':
maxout_size = activation[2]
maxout_type = activation[1]
out, out_shp = activations.Maxout(x=x,
maxout_size=maxout_size,
input_size=input_size,
type=maxout_type,
dimension=kwargs["dimension"])
if activation[0] == 'relu':
relu_leak = activation[1]
out = activations.ReLU(x=x, alpha=relu_leak)
out_shp = input_size
if activation[0] == 'softmax':
temperature = activation[1]
out = activations.Softmax(x=x, temp=temperature)
out_shp = input_size
else:
if activation == 'relu':
out = activations.ReLU(x=x)
elif activation == 'abs':
out = activations.Abs(x=x)
elif activation == 'sigmoid':
out = activations.Sigmoid(x=x)
elif activation == 'tanh':
out = activations.Tanh(x=x)
elif activation == 'softmax':
out = activations.Softmax(x=x)
elif activation == 'squared':
out = activations.Squared(x=x)
elif activation is None:
out = x
out_shp = input_size
if verbose >= 3:
print("... Activations are setup")
return (out, out_shp)
def test_meanout(self):
out, out_shp = A.Maxout(self.input_zeros,
self.maxout_size,
self.input_size_min,
type='meanout',
dimension=1)
self.assertTrue(np.allclose(out, self.input_zeros))
self.assertTrue(out_shp, self.input_size_min)