def test_002_parallel_layers(self):
l1 = nn.FeedForward(self._NFEATURES, self._SIZE, has_bias=True)
nn.init_(l1.linear.weight, self.W)
nn.init_(l1.linear.bias, self.b)
l2 = nn.FeedForward(self._NFEATURES, self._SIZE, has_bias=True)
nn.init_(l2.linear.weight, self.W)
nn.init_(l2.linear.bias, self.b)
network = nn.Parallel([l1, l2])
with torch.no_grad():
res = network(torch.tensor(self.x)).numpy()
np.testing.assert_almost_equal(res[:, :, :self._SIZE], res[:, :, self._SIZE:])
def test_005_poormans_birnn(self):
layer1 = nn.FeedForward(self._NFEATURES, self._SIZE, has_bias=True)
nn.init_(layer1.linear.weight, self.W)
nn.init_(layer1.linear.bias, self.b)
layer2 = nn.FeedForward(self._NFEATURES, self._SIZE, has_bias=True)
nn.init_(layer2.linear.weight, self.W)
nn.init_(layer2.linear.bias, self.b)
network = nn.birnn(layer1, layer2)
with torch.no_grad():
res = network(torch.tensor(self.x)).numpy()
np.testing.assert_almost_equal(res[:, :, :self._SIZE], res[:, :, self._SIZE:], decimal=5)
def test_003_simple_serial(self):
W2 = np.random.normal(size=(self._SIZE, self._SIZE)).astype(taiyaki_dtype)
res = self.res.dot(W2.transpose())
l1 = nn.FeedForward(self._NFEATURES, self._SIZE, has_bias=True,
fun=activation.linear)
nn.init_(l1.linear.weight, self.W)
nn.init_(l1.linear.bias, self.b)
l2 = nn.FeedForward(self._SIZE, self._SIZE, fun=activation.linear, has_bias=False)
nn.init_(l2.linear.weight, W2)
network = nn.Serial([l1, l2])
with torch.no_grad():
y = network(torch.tensor(self.x)).numpy()
np.testing.assert_almost_equal(y, res, decimal=4)
def test_001_single_layer_tanh(self):
network = nn.FeedForward(self._NFEATURES, self._SIZE, has_bias=True,
fun=activation.tanh)
nn.init_(network.linear.weight, self.W)
nn.init_(network.linear.bias, self.b)
with torch.no_grad():
y = network(torch.tensor(self.x)).numpy()
np.testing.assert_almost_equal(y, np.tanh(self.res), decimal=5)
def test_004_reverse(self):
network1 = nn.FeedForward(self._NFEATURES, self._SIZE, has_bias=True)
nn.init_(network1.linear.weight, self.W)
nn.init_(network1.linear.bias, self.b)
network2 = nn.Reverse(network1)
with torch.no_grad():
res1 = network1(torch.tensor(self.x)).numpy()
res2 = network2(torch.tensor(self.x)).numpy()
np.testing.assert_almost_equal(res1, res2, decimal=5)
def setUp(self):
sublayer = nn.FeedForward(12, 12, has_bias=True)
self.layer = nn.Residual(sublayer)
def setUp(self):
self.linear1 = nn.FeedForward(12, 8)
self.linear2 = nn.FeedForward(12, 8)
self.layer = nn.Product([self.linear1, self.linear2])
