def test_008_rnn_no_input(self):
iW = np.zeros((self._SIZE, self._NFEATURES), dtype=sloika_dtype)
sW = np.random.normal(size=(self._SIZE,
self._SIZE)).astype(sloika_dtype)
network = nn.Recurrent(self._NFEATURES, self._SIZE)
network.set_params({'iW': iW, 'sW': sW})
f = network.compile()
res = f(self.x)
np.testing.assert_almost_equal(res, 0.0)
def test_007_rnn_no_state(self):
sW = np.zeros((self._SIZE, self._SIZE), dtype=sloika_dtype)
network = nn.Recurrent(self._NFEATURES,
self._SIZE,
has_bias=True,
fun=activation.linear)
network.set_params({'iW': self.W, 'sW': sW, 'b': self.b})
f = network.compile()
res = f(self.x)
np.testing.assert_almost_equal(res, self.res, decimal=5)
def test_010_birnn_no_input_with_bias(self):
iW = np.zeros((self._SIZE, self._NFEATURES), dtype=sloika_dtype)
sW = np.random.normal(size=(self._SIZE,
self._SIZE)).astype(sloika_dtype)
layer1 = nn.Recurrent(self._NFEATURES,
self._SIZE,
has_bias=True,
fun=activation.linear)
layer1.set_params({'iW': iW, 'sW': sW, 'b': self.b})
layer2 = nn.Recurrent(self._NFEATURES,
self._SIZE,
has_bias=True,
fun=activation.linear)
layer2.set_params({'iW': iW, 'sW': sW, 'b': self.b})
network = nn.birnn(layer1, layer2)
f = network.compile()
res = f(self.x)
np.testing.assert_almost_equal(res[:, :, :self._SIZE],
res[::-1, :, self._SIZE:])
def test_009_rnn_no_input_with_bias(self):
iW = np.zeros((self._SIZE, self._NFEATURES), dtype=sloika_dtype)
sW = rvs(self._SIZE).astype(sloika_dtype)
network = nn.Recurrent(self._NFEATURES,
self._SIZE,
has_bias=True,
fun=activation.linear)
network.set_params({'iW': iW, 'sW': sW, 'b': self.b})
f = network.compile()
res = f(self.x)
res2 = np.zeros((self._NBATCH, self._SIZE), dtype=sloika_dtype)
for i in range(self._NSTEP):
res2 = res2.dot(sW.transpose()) + self.b
np.testing.assert_almost_equal(res[i], res2)
def setUp(self):
self.layer = nn.Recurrent(12, 64, has_bias=True)
def setUp(self):
self.layer = nn.Recurrent(12, 64)