def __call__(self, x, y=None):
self.qw1 = nn.random.Gaussian(self.w1_m,
nn.softplus(self.w1_s),
p=nn.random.Gaussian(0, 1))
self.qb1 = nn.random.Gaussian(self.b1_m,
nn.softplus(self.b1_s),
p=nn.random.Gaussian(0, 1))
self.qw2 = nn.random.Gaussian(self.w2_m,
nn.softplus(self.w2_s),
p=nn.random.Gaussian(0, 1))
self.qb2 = nn.random.Gaussian(self.b2_m,
nn.softplus(self.b2_s),
p=nn.random.Gaussian(0, 1))
self.qw3 = nn.random.Gaussian(self.w3_m,
nn.softplus(self.w3_s),
p=nn.random.Gaussian(0, 1))
self.qb3 = nn.random.Gaussian(self.b3_m,
nn.softplus(self.b3_s),
p=nn.random.Gaussian(0, 1))
h = nn.tanh(x @ self.qw1.draw() + self.qb1.draw())
h = nn.tanh(h @ self.qw2.draw() + self.qb2.draw())
self.py = nn.random.Bernoulli(logit=h @ self.qw3.draw() +
self.qb3.draw(),
data=y)
return self.py.mu.value
def forward(self, x):
h = x
for i in range(self.n_unit - 1):
h = nn.tanh(
np.matmul(
h, self.parameter["w_encode{vari}".format(vari=i)] +
self.parameter["b_encode{vari}".format(vari=i)]))
for i in range(self.n_unit - 2, 0, -1):
h = nn.tanh(
np.matmul(
h, self.parameter["w_decode{vari}".format(vari=i)] +
self.parameter["b_decode{vari}".format(vari=i)]))
x_ = np.matmul(
h, self.parameter["w_decode0"] + self.parameter["b_decode0"])
self.px = nn.random.Gaussian(x_, 1., data=x)
def __call__(self, x, y=None):
h = nn.tanh(x @ self.w1 + self.b1)
coef = nn.softmax(h @ self.w2_c + self.b2_c)
mean = h @ self.w2_m + self.b2_m
std = nn.exp(h @ self.w2_s + self.b2_s)
self.py = nn.random.GaussianMixture(coef, mean, std, data=y)
return self.py
def transform(self, x):
h = x
for i in range(self.n_unit - 1):
h = nn.tanh(
np.matmul(
h, self.parameter["w_encode{vari}".format(vari=i)] +
self.parameter["b_encode{vari}".format(vari=i)]))
return h.value
def __call__(self, x, y=None):
self.pw1 = nn.random.Gaussian(0., 1., data=self.w1)
self.pb1 = nn.random.Gaussian(0., 1., data=self.b1)
self.pw2 = nn.random.Gaussian(0., 1., data=self.w2)
self.pb2 = nn.random.Gaussian(0., 1., data=self.b2)
h = nn.tanh(x @ self.w1 + self.b1)
self.py = nn.random.Gaussian(h @ self.w2 + self.b2, std=0.1, data=y)
return self.py.mu.value
def __call__(self, x, y=None):
h = nn.tanh(x @ self.w1 + self.b1)
self.py = nn.random.Bernoulli(logit=h @ self.w2 + self.b2, data=y)
return self.py.mu.value
def __call__(self, x, y=None):
h = nn.tanh(x @ self.w1 + self.b1)
self.py = nn.random.Gaussian(h @ self.w2 + self.b2, std=1.,
data=y) # 这里最后一步不是用激活函数，而是用高斯分布，什么作用
return self.py.mu.value
def test_tanh(self):
self.assertEqual(nn.tanh(0).value, 0)
def test_tanh(self):
self.assertEqual(nn.tanh(0).value, 0)