Remove all bn.input and bn.output

thu-ml · Apr 21, 2019 · 4e9230a · 4e9230a
1 parent eca61b7
commit 4e9230a
Show file tree

Hide file tree

Showing 4 changed files with 8 additions and 19 deletions.
diff --git a/examples/bayesian_neural_nets/variational_dropout.py b/examples/bayesian_neural_nets/variational_dropout.py
@@ -18,11 +18,11 @@
 
 
 @zs.meta_bayesian_net(scope="model", reuse_variables=True)
-def var_dropout(n, net_size, n_particles, is_training):
+def var_dropout(x, n, net_size, n_particles, is_training):
     normalizer_params = {'is_training': is_training,
                          'updates_collections': None}
     bn = zs.BayesianNet()
-    h = bn.input("x")
+    h = x
     for i, [n_in, n_out] in enumerate(zip(net_size[:-1], net_size[1:])):
         eps_mean = tf.ones([n, n_in])
         eps = bn.normal(
@@ -34,7 +34,7 @@ def var_dropout(n, net_size, n_particles, is_training):
         if i < len(net_size) - 2:
             h = tf.nn.relu(h)
     y = bn.categorical('y', h)
-    bn.output('y_logit', h)
+    bn.deterministic('y_logit', h)
     return bn
 
 
@@ -90,7 +90,7 @@ def q(n, net_size, n_particles):
     x_obs = tf.tile(tf.expand_dims(x, 0), [n_particles, 1, 1])
     y_obs = tf.tile(tf.expand_dims(y, 0), [n_particles, 1])
 
-    model = var_dropout(n, net_size, n_particles, is_training)
+    model = var_dropout(x_obs, n, net_size, n_particles, is_training)
     variational = q(n, net_size, n_particles)
 
     def log_joint(bn):
@@ -100,7 +100,7 @@ def log_joint(bn):
 
     model.log_joint = log_joint
 
-    lower_bound = zs.variational.elbo(model, {'x': x_obs, 'y': y_obs},
+    lower_bound = zs.variational.elbo(model, {'y': y_obs},
                                       variational=variational, axis=0)
     y_logit = lower_bound.bn["y_logit"]
     h_pred = tf.reduce_mean(tf.nn.softmax(y_logit), 0)

diff --git a/examples/toy_examples/toy2d_intractable.py b/examples/toy_examples/toy2d_intractable.py
@@ -24,8 +24,8 @@ def build_toy2d_intractable(n_particles):
 def build_mean_field_variational(n_particles):
     bn = zs.BayesianNet()
     for name in ["z1", "z2"]:
-        z_mean = bn.output(name + "_mean", tf.Variable(-2.))
-        z_logstd = bn.output(name + "_logstd", tf.Variable(-5.))
+        z_mean = bn.deterministic(name + "_mean", tf.Variable(-2.))
+        z_logstd = bn.deterministic(name + "_logstd", tf.Variable(-5.))
         bn.normal(name, z_mean, logstd=z_logstd, n_samples=n_particles)
     return bn
 

diff --git a/examples/variational_autoencoders/vae_conv.py b/examples/variational_autoencoders/vae_conv.py
@@ -68,7 +68,7 @@ def build_gen(n, x_dim, z_dim, n_particles, nf=16):
     lx_z = conv2d_transpose(lx_z, [28, 28, 1], kernel_size=(3, 3),
                             stride=(1, 1), activation_fn=None)
     x_logits = tf.reshape(lx_z, [n_particles, -1, x_dim])
-    bn.output("x_mean", tf.sigmoid(x_logits))
+    bn.deterministic("x_mean", tf.sigmoid(x_logits))
     bn.bernoulli("x", x_logits, group_ndims=1)
     return bn
 

diff --git a/zhusuan/framework/bn.py b/zhusuan/framework/bn.py
@@ -370,17 +370,6 @@ def stochastic(self, name, dist, **kwargs):
         self._nodes[name] = node
         return node
 
-    # TODO: Deprecate and prefer unbounded input as function arguments
-    def input(self, name):
-        input_tensor = tf.convert_to_tensor(self._get_observation(name))
-        self._nodes[name] = input_tensor
-        return input_tensor
-
-    # TODO: Deprecate and prefer deterministic
-    def output(self, name, input_tensor):
-        self._nodes[name] = input_tensor
-        return input_tensor
-
     def deterministic(self, name, input_tensor):
         """
         Add a named deterministic node in this :class:`BayesianNet`.