def test_perceptron(self):
pc = dy.ParameterCollection()
init = OrthogonalInitializer
mlp = MLP(pc, [10, 8, 5], init=init)
x = dy.random_normal((10,))
y = mlp(x, True)
assert y.dim() == ((5,), 1)
mlp_batch = MLP(pc, [10, 8, 5], p=0.5, init=init)
x = dy.random_normal((10,), batch_size=5)
y = mlp_batch(x, True)
assert y.dim() == ((5,), 5)
def test_linear(self):
pc = dy.ParameterCollection()
init = OrthogonalInitializer
affine = Linear(pc, in_dim=10, out_dim=5, bias=True, init=init)
x = dy.random_normal((10, ))
y = affine(x)
assert y.dim() == ((5, ), 1)
init = dy.ConstInitializer(1)
affine = Linear(pc, in_dim=10, out_dim=1, bias=False, init=init)
x = dy.random_normal((10, ))
y = affine(x)
assert math.fabs(np.sum(y.npvalue()) - np.sum(x.npvalue())) < 1e-6
def reparameterize(self, mu, logvar):
if self.training:
std = dy.exp(logvar * 0.5)
eps = dy.random_normal(dim=std.dim()[0], mean=0.0, stddev=1.0)
return dy.cmult(eps, std) + mu
else:
return mu
def reparameterize(mu, logvar):
# Get z by reparameterization.
d = mu.dim()[0][0]
eps = dy.random_normal(d)
std = dy.exp(logvar * 0.5)
return mu + dy.cmult(std, eps)
def reparameterize(self, mu, logvar):
if self.training:
std = dy.exp(logvar * 0.5)
eps = dy.random_normal(dim=std.dim()[0], mean=0.0, stddev=1.0)
return dy.cmult(eps, std) + mu
else:
return mu
def reparameterize(mu, logvar):
# Get z by reparameterization.
d = mu.dim()[0][0]
eps = dy.random_normal(d)
std = dy.exp(logvar * 0.5)
return mu + dy.cmult(std, eps)
def test_DeepBiRNNBuilder(self):
pc = dy.ParameterCollection()
ENC = DeepBiRNNBuilder(pc, 2, 50, 20, orthonormal_VanillaLSTMBuilder)
x = [dy.random_normal((50, )) for _ in range(10)]
y = ENC(x, p_x=0.33, p_h=0.33, train=True)
assert len(y) == 10
assert y[0].dim() == ((40, ), 1)
def reparameterize(self, mu, logvar): d = mu.dim()[0][0] eps = dy.random_normal(d) std = dy.exp(logvar * 0.5) return mu + dy.cmult(std, eps)
def reparameterize(mu, log_sigma_squared):
d = mu.dim()[0][0]
sample = dy.random_normal(d)
covar = dy.exp(log_sigma_squared * 0.5)
return mu + dy.cmult(covar, sample)