def __call__(self, x):
h = F.crelu(self.qlin0(x))
h = F.crelu(self.qlin1(h))
qmu = self.qlin_mu(h)
qln_var = self.qlin_ln_var(h)
return qmu, qln_var
def decode(self, z):
# pdb.set_trace()
a = self.a_enc
# If this function is coming from the sampling call, the batch size of z and a won't match. Manually handle that here.
if (a.shape[0] != z.shape[0]):
a.volatile = 'ON'
batch_size = z.shape[0]
a.data = a.data[0:batch_size, :]
net_input = F.concat((z, a), axis=1)
h = self.plinx0(net_input)
h = self.plinx_batch_norm_0(h)
h = F.crelu(h)
for i in range(self.num_layers - 1):
layer_name = 'plinx' + str(i + 1)
h = self[layer_name](h)
layer_name = 'plinx_batch_norm_' + str(i + 1)
h = self[layer_name](h)
h = F.crelu(h)
self.p_ber_prob_logit = self.plinx_ber_prob(h)
return self.p_ber_prob_logit
def decode(self, z):
h = F.crelu(self.plin0(z))
for i in range(self.num_layers - 1):
layer_name = 'plin' + str(i + 1)
h = F.crelu(self[layer_name](h))
self.p_ber_prob_logit = self.plin_ber_prob(h)
def encode(self, x):
h = F.crelu(self.qlin0(x))
for i in range(self.num_layers-1):
layer_name = 'qlin' + str(i+1)
h = F.crelu(self[layer_name](h))
self.qmu = self.qlin_mu(h)
self.qln_var = self.qlin_ln_var(h)
def decode(self, z):
h = F.crelu(self.plin0(z))
for i in range(self.num_layers - 1):
layer_name = 'plin' + str(i + 1)
h = F.crelu(self[layer_name](h))
self.pmu = self.plin_mu(h)
self.pln_var = self.plin_ln_var(h)
def encode_a(self, x):
a_params = F.crelu(self.qlina0(x))
for i in range(self.num_layers - 1):
layer_name = 'qlina' + str(i + 1)
a_params = F.crelu(self[layer_name](a_params))
self.qmu_a = self.qlina_mu(a_params)
self.qln_var_a = self.qlina_ln_var(a_params)
return self.qmu_a, self.qln_var_a
def encode_z(self, x, a):
# a = F.gaussian(self.qmu_a, self.qln_var_a) # This should be outside the encoding function. Pass the function a.
net_input = F.concat((x, a), axis=1)
h = F.crelu(self.qlinz0(net_input))
for i in range(self.num_layers - 1):
layer_name = 'qlinz' + str(i + 1)
h = F.crelu(self[layer_name](h))
self.qmu_z = self.qlinz_mu(h)
self.qln_var_z = self.qlinz_ln_var(h)
return self.qmu_z, self.qln_var_z
def decode_a(self, z, x):
net_input = F.concat((x, z), axis=1)
h = F.crelu(self.plina0(net_input))
for i in range(self.num_layers - 1):
layer_name = 'plina' + str(i + 1)
h = F.crelu(self[layer_name](h))
self.pmu_a = self.plina_mu(h)
self.pln_var_a = self.plina_ln_var(h)
return self.pmu_a, self.pln_var_a
def decode(self, z):
h = self.plinx0(z)
h = self.plinx_batch_norm_0(h)
h = F.crelu(h)
for i in range(self.num_layers - 1):
layer_name = 'plinx' + str(i + 1)
h = self[layer_name](h)
layer_name = 'plinx_batch_norm_' + str(i + 1)
h = self[layer_name](h)
h = F.crelu(h)
self.p_ber_prob_logit = self.plinx_ber_prob(h)
return self.p_ber_prob_logit
def iaf(self, z, h, lin1, lin2):
ms = F.crelu(lin1(F.concat((z, h), axis=1)))
ms = lin2(ms)
m, s = F.split_axis(ms, 2, axis=1)
s = F.sigmoid(s)
z = s * z + (1 - s) * m
# pdb.set_trace()
return z, -F.sum(F.log(s), axis=1)
def encode(self, x):
h = self.qlin0(x)
h = self.qlin_batch_norm_0(h)
h = F.crelu(h)
for i in range(self.num_layers - 1):
layer_name = 'qlin' + str(i + 1)
h = self[layer_name](h)
layer_name = 'qlin_batch_norm_' + str(i + 1)
h = self[layer_name](h)
h = F.crelu(h)
self.qmu = self.qlin_mu(h)
self.qln_var = self.qlin_ln_var(h)
self.qh_vec_0 = self.qlin_h_vec_0(h)
return self.qmu, self.qln_var, self.qh_vec_0
def decode_a(self, z):
# net_input = F.concat((x,z), axis=1)
h = self.plina0(z)
h = self.plina_batch_norm_0(h)
h = F.crelu(h)
for i in range(self.num_layers - 1):
layer_name = 'plina' + str(i + 1)
h = self[layer_name](h)
layer_name = 'plina_batch_norm_' + str(i + 1)
h = self[layer_name](h)
h = F.crelu(h)
self.pmu_a = self.plina_mu(h)
self.pln_var_a = self.plina_ln_var(h)
return self.pmu_a, self.pln_var_a
def check_forward(self, x_data):
x = chainer.Variable(x_data)
y = functions.crelu(x, axis=self.axis)
self.assertEqual(y.data.dtype, self.dtype)
self.assertEqual(y.data.shape, self.y_shape)
expected_former = numpy.maximum(self.x, 0)
expected_latter = numpy.maximum(-self.x, 0)
expected = numpy.concatenate(
(expected_former, expected_latter), axis=self.axis)
testing.assert_allclose(expected, y.data)
def check_forward(self, x_data):
x = chainer.Variable(x_data)
y = functions.crelu(x, axis=self.axis)
self.assertEqual(y.data.dtype, numpy.float32)
self.assertEqual(y.data.shape, self.y_shape)
expected_former = self.x.copy()
expected_latter = self.x.copy()
for i in numpy.ndindex(self.x.shape):
expected_former[i] = max(0, self.x[i])
expected_latter[i] = max(0, -self.x[i])
expected = numpy.concatenate((expected_former, expected_latter),
axis=self.axis)
testing.assert_allclose(expected, y.data)
def check_forward(self, x_data):
x = chainer.Variable(x_data)
y = functions.crelu(x, axis=self.axis)
self.assertEqual(y.data.dtype, numpy.float32)
self.assertEqual(y.data.shape, self.y_shape)
expected_former = self.x.copy()
expected_latter = self.x.copy()
for i in numpy.ndindex(self.x.shape):
expected_former[i] = max(0, self.x[i])
expected_latter[i] = max(0, -self.x[i])
expected = numpy.concatenate(
(expected_former, expected_latter), axis=self.axis)
gradient_check.assert_allclose(expected, y.data)
def __call__(self, x):
return F.crelu(x, self.axis)
def forward(self, inputs, device):
x, = inputs
return functions.crelu(x, axis=self.axis),
def __call__(self, z):
h = F.crelu(self.plin0(z))
h = F.crelu(self.plin1(h))
ph = self.plin2(h)
return MNISTLikelihood(ph)
def forward(self, inputs, device):
x, = inputs
return functions.crelu(x, axis=self.axis),
def __call__(self, x): return F.crelu(x, self.axis)
def __call__(self, x): return functions.crelu(x, self.axis)
def __call__(self, x): return functions.crelu(x, self.axis)