def build()

in product_matching/hyperboloid.py [0:0]

• 52 lines of code
• 2 McCabe index (conditional complexity)

    def build(self):
        """
        Build model structure.

        Hyperboloid use Joint learning arthitecture.
        """
        print("Query")
        query_input = Input(
            name='text_left',
            shape=self._params['input_shapes'][0],
        )
        asin_input = Input(
            name='text_right',
            shape=self._params['input_shapes'][1]
        )
        #query_input = Input(shape=(self.hyper_params["MAX_LEN"],), dtype='float32')
        print(query_input)
        query = ManifoldEmbedding(self._params["vocab_size"], self._params["emb_dim"], trainable = True, mask_zero=True,manifold=self.params['manifold'])(query_input)
        #query = Embedding(self._params["vocab_size"], self._params["emb_dim"], trainable = True, mask_zero=True)(query_input)
        print(query)
        box = TimeDistributed(Dense(self._params["box_dim"], activation='sigmoid'))(query)
        print(box)        
        box_bn = BatchNormalization()(box)
        box_dropout = Dropout(self._params["dropout"])(box_bn)
        print(box_dropout)
        intersection = Lambda(self.intersection_layer)([box_dropout,box_dropout])
        print(intersection)
        reshape_intersection = Reshape((intersection.shape[1]//self._params["box_dim"],self._params["box_dim"]))(intersection)
        print(reshape_intersection)
        query_attention = SelfAttention(return_sequences=True)(reshape_intersection)
        print(query_attention)
        query_attention_bn = BatchNormalization()(query_attention)
        query_attention_dropout = Dropout(self._params["dropout"])(query_attention_bn)
        print(query_attention_dropout)
        query_boxes = Dense(self._params["box_dim"])(query_attention_dropout)
        print(query_boxes)

        print("\nASIN")
        #asin_input = Input(shape=(self.hyper_params["MAX_LEN"],), dtype='float32')
        print(asin_input)
        asin = ManifoldEmbedding(self._params["vocab_size"], self._params["emb_dim"], trainable = True, mask_zero=True,manifold=self.params['manifold'])(asin_input)
        #asin = Embedding(self._params["vocab_size"], self._params["emb_dim"], trainable = True, mask_zero=True)(asin_input)
        print(asin)
        lstm = Bidirectional(LSTM(self._params["emb_dim"]//2, return_sequences=True))(asin)
        print(lstm)
        attention = SelfAttention(return_sequences=True)(lstm)
        print(attention)
        attention_bn = BatchNormalization()(attention)
        print(attention_bn)
        attention_dropout = Dropout(self._params["dropout"])(attention_bn)
        print(attention_dropout)
        dense_attention = Dense(self._params["box_dim"]//2)(attention_dropout)
        print(dense_attention)
        asin_embedding = self.hyperbolic_add()([self.crop_box(i,None,None)(dense_attention) for i in range(dense_attention.shape[1])])
        print(asin_embedding)
        print("\nJoin")
        merged_layer = Lambda(self.merge_layer)([query_boxes,asin_embedding])
        print(merged_layer)
    	merged_layer_bn = BatchNormalization()(merged_layer)
        print(merged_layer_bn)
        merged_layer_dropout = Dropout(self._params["dropout"])(merged_layer_bn) 
        print(merged_layer_dropout)