tucker first version

models/tucker.py

@@ -0,0 +1,60 @@
+import tensorflow as tf
+
+class TuckerModel(tf.keras.Model):
+  def __init__(self, config):
+    super(TuckerModel, self).__init__()
+
+    self.config = config
+
+    self.entity_embeddings = tf.Variable(
+        tf.random.normal(
+            [config.n_entities, config.d_entity],
+            0.0, 0.1))  
+
+    self.bias = tf.Variable(tf.zeros([config.n_entities,]))  
+
+    ''' ##
+    self.relation_embeddings = tf.Variable(
+        tf.random.normal(
+            [config.n_relations, config.d_relation],
+            0.0, 0.1))  
+    
+    ''' ##
+    self.core_tensor = tf.Variable(
+        tf.random.normal(
+            [config.d_relation, config.d_entity * config.d_entity],
+            0.0, 0.1)) 
+
+    self.input_dropout = tf.keras.layers.Dropout(config.input_d)
+    self.hidden_dropout1 = tf.keras.layers.Dropout(config.hidden_d1)
+    self.hidden_dropout2 = tf.keras.layers.Dropout(config.hidden_d2)
+
+  def call(self, input, training=False):
+    batch_sub_embedding = tf.nn.embedding_lookup(
+        self.entity_embeddings, input[:,0])
+    #batch_sub_embedding = tf.expand_dims(batch_sub_embedding, 1)
+    batch_sub_embedding = self.input_dropout(
+      batch_sub_embedding,
+      training=training)
+
+    '''
+    batch_rel_embedding = tf.nn.embedding_lookup(
+        self.relation_embeddings, input[:,1])
+    w_mat = tf.matmul(batch_rel_embedding, self.core_tensor)
+    w_mat = tf.reshape(w_mat, [-1, config.d_entity, config.d_entity])
+    w_mat = self.hidden_dropout1(w_mat, training=training)
+    '''
+    ####
+    w_mat = tf.reshape(self.core_tensor, [self.config.d_entity, self.config.d_entity])
+    ####
+
+    sub_rel_representation = tf.matmul(batch_sub_embedding, w_mat)
+    sub_rel_representation = tf.squeeze(sub_rel_representation)
+    sub_rel_representation = self.hidden_dropout2(
+      sub_rel_representation, training=training)
+
+    scores = tf.matmul(
+      sub_rel_representation,
+      self.entity_embeddings,
+      transpose_b = True) + self.bias
+    return scores