def test_init_untie_biases(self, NINLayer_c01b, dummy_input_layer):
layer = NINLayer_c01b(
dummy_input_layer,
num_units=5,
untie_biases=True,
)
assert (layer.b.shape.eval() == (5, 4, 5)).all()
def test_get_output_for(self, dummy_input_layer, NINLayer_c01b,
extra_kwargs):
nonlinearity = Mock()
layer = NINLayer_c01b(
dummy_input_layer,
num_units=6,
nonlinearity=nonlinearity,
**extra_kwargs
)
input = theano.shared(np.random.uniform(-1, 1, (3, 4, 5, 2)))
result = layer.get_output_for(input)
assert result is nonlinearity.return_value
nonlinearity_arg = nonlinearity.call_args[0][0]
X = input.get_value()
W = layer.W.get_value()
out = np.dot(W, X.reshape(X.shape[0], -1))
out = out.reshape(W.shape[0], X.shape[1], X.shape[2], X.shape[3])
if layer.b is not None:
if layer.untie_biases:
out += layer.b.get_value()[..., None]
else:
out += layer.b.get_value()[:, None, None, None]
assert np.allclose(nonlinearity_arg.eval(), out)
def test_init_none_nonlinearity_bias(self, NINLayer_c01b,
dummy_input_layer):
layer = NINLayer_c01b(
dummy_input_layer,
num_units=3,
nonlinearity=None,
b=None,
)
assert layer.nonlinearity == lasagne.nonlinearities.identity
assert layer.b is None
def layer_vars(self, NINLayer_c01b, dummy_input_layer):
W = Mock()
b = Mock()
nonlinearity = Mock()
W.return_value = np.ones((5, 3))
b.return_value = np.ones((5, ))
layer = NINLayer_c01b(
dummy_input_layer,
num_units=5,
W=W,
b=b,
nonlinearity=nonlinearity,
)
return {
'W': W,
'b': b,
'nonlinearity': nonlinearity,
'layer': layer,
}