issue-43b.diff

diff --git a/nw2vec/ae.py b/nw2vec/ae.py
index 0488548..1fbb469 100644
--- a/nw2vec/ae.py
+++ b/nw2vec/ae.py
@@ -483,18 +483,19 @@ def build_p_builder(dims, use_bias=False):
         p_adj = layers.Bilinear(0, use_bias=use_bias,
                                 kernel_regularizer='l2', bias_regularizer='l2',
                                 name='p_adj')([p_layer1, p_layer1])
-        p_v_μ_flat = keras.layers.Dense(dim_data, use_bias=use_bias,
-                                        kernel_regularizer='l2', bias_regularizer='l2',
-                                        name='p_v_mu_flat')(p_layer1)
-        p_v_logD_flat = keras.layers.Dense(dim_data, use_bias=use_bias,
-                                           kernel_regularizer='l2', bias_regularizer='l2',
-                                           name='p_v_logD_flat')(p_layer1)
-        p_v_u_flat = keras.layers.Dense(dim_data, use_bias=use_bias,
-                                        kernel_regularizer='l2', bias_regularizer='l2',
-                                        name='p_v_u_flat')(p_layer1)
-        p_v_μlogDu_flat = keras.layers.Concatenate(name='p_v_mulogDu_flat')(
-            [p_v_μ_flat, p_v_logD_flat, p_v_u_flat])
-        return ([p_adj, p_v_μlogDu_flat], ('SigmoidBernoulliAdjacency', 'Gaussian'))
+        # p_v_μ_flat = keras.layers.Dense(dim_data, use_bias=use_bias,
+        #                                 kernel_regularizer='l2', bias_regularizer='l2',
+        #                                 name='p_v_mu_flat')(p_layer1)
+        # p_v_logD_flat = keras.layers.Dense(dim_data, use_bias=use_bias,
+        #                                    kernel_regularizer='l2', bias_regularizer='l2',
+        #                                    name='p_v_logD_flat')(p_layer1)
+        # p_v_u_flat = keras.layers.Dense(dim_data, use_bias=use_bias,
+        #                                 kernel_regularizer='l2', bias_regularizer='l2',
+        #                                 name='p_v_u_flat')(p_layer1)
+        # p_v_μlogDu_flat = keras.layers.Concatenate(name='p_v_mulogDu_flat')(
+        #     [p_v_μ_flat, p_v_logD_flat, p_v_u_flat])
+        # return ([p_adj, p_v_μlogDu_flat], ('SigmoidBernoulliAdjacency', 'Gaussian'))
+        return ([p_adj], ('SigmoidBernoulliAdjacency',))
 
     return p_builder
 
diff --git a/nw2vec/codecs.py b/nw2vec/codecs.py
index 9551357..7fe8c51 100644
--- a/nw2vec/codecs.py
+++ b/nw2vec/codecs.py
@@ -151,12 +151,14 @@ class SigmoidBernoulliAdjacency(Codec):
         # `adj` now has shape (batch or 1, sampling, batch, batch)
         assert len(adj.shape) == 4  # If this fails, change it to a dynamic check
 
-        # #38 showed that weighing links and non-links so that they each contribute 1/2
-        # gives a worse result on BlogCatalog than not weighing them flat. See
-        # https://github.com/ixxi-dante/nw2vec/commit/3bba9c8fdd5afd66a593a61bf8473b8f794e28fb
-        # for the change which led to this version.
-        weighted_sigmoid_cross_entropies = \
-            tf.nn.sigmoid_cross_entropy_with_logits(labels=adj, logits=self.logits)
+        # Temporary replacement for #43
+        density = K.mean(adj)
+        weighted_sigmoid_cross_entropies = (
+            .5
+            * tf.nn.weighted_cross_entropy_with_logits(
+                targets=adj, logits=self.logits, pos_weight=(1 / density) - 1)
+            / (1 - density)
+        )
 
         return - K.mean(K.sum(weighted_sigmoid_cross_entropies, axis=-1), axis=-1)
 
diff --git a/projects/scale/blogcatalog.py b/projects/scale/blogcatalog.py
index bec7341..7153829 100644
--- a/projects/scale/blogcatalog.py
+++ b/projects/scale/blogcatalog.py
@@ -26,7 +26,8 @@ dim_l1, dim_ξ = 10, 10
 use_bias = False
 
 # Training
-loss_weights = [1.0, 1.0, 1.0]  # q, p_adj, p_v
+# loss_weights = [1.0, 1.0, 1.0]  # q, p_adj, p_v
+loss_weights = [1.0, 1.0]  # q, p_adj
 n_epochs = 10000
 # seeds_per_batch = len(nodes) -> defined below
 max_walk_length = 1
@@ -113,7 +114,7 @@ dim_data = len(groups)
 dims = (dim_data, dim_l1, dim_ξ)
 DATA_PARAMETERS = 'crop={crop}'.format(crop=crop)
 VAE_PARAMETERS = (
-    'no_adj_cross_entropy_weighing'
+    'no_feature_reconstruction'
     '-n_ξ_samples={n_ξ_samples}'
     '-dims={dims}'
     '-bias={use_bias}').format(n_ξ_samples=n_ξ_samples,
@@ -173,7 +174,7 @@ def target_func(batch_adj, required_nodes, final_nodes):
                                    0, n_ξ_samples),
             0, 1
         ),
-        utils.expand_dims_tile(features[final_nodes], 1, n_ξ_samples),
+        # utils.expand_dims_tile(features[final_nodes], 1, n_ξ_samples),
     ]