dssm.py 文件源码

def build(self):
        query = Input(name='query', shape=(self.config['vocab_size'],))#, sparse=True)
        show_layer_info('Input', query)
        doc = Input(name='doc', shape=(self.config['vocab_size'],))#, sparse=True)
        show_layer_info('Input', doc)

        def mlp_work(input_dim):
            seq = Sequential()
            #seq.add(SparseFullyConnectedLayer(self.config['hidden_sizes'][0], input_dim=input_dim, activation='relu'))
            num_hidden_layers = len(self.config['hidden_sizes'])
            if num_hidden_layers == 1:
                seq.add(Dense(self.config['hidden_sizes'][0], input_shape=(input_dim,), activity_regularizer=regularizers.l2(self.config['reg_rate'])))
            else:
                seq.add(Dense(self.config['hidden_sizes'][0], activation='tanh', input_shape=(input_dim,), activity_regularizer=regularizers.l2(self.config['reg_rate'])))
                for i in range(num_hidden_layers-2):
                    seq.add(Dense(self.config['hidden_sizes'][i+1], activation='tanh', activity_regularizer=regularizers.l2(self.config['reg_rate'])))
                    seq.add(Dropout(rate=self.config['dropout_rate']))
                seq.add(Dense(self.config['hidden_sizes'][num_hidden_layers-1], activity_regularizer=regularizers.l2(self.config['reg_rate'])))
                seq.add(Dropout(rate=self.config['dropout_rate']))
            return seq

        mlp = mlp_work(self.config['vocab_size'])
        rq = mlp(query)
        show_layer_info('MLP', rq)
        rd = mlp(doc)
        show_layer_info('MLP', rd)

        '''
        rep = Concatenate(axis=1) ([rq, rd])
        show_layer_info('Concatenate', rep)
        rep = Dropout(rate=self.config['dropout_rate'])(rep)
        show_layer_info('Dropout', rep)
        if self.config['target_mode'] == 'classification':
            out_ = Dense(2, activation='softmax')(rep)
        elif self.config['target_mode'] in ['regression', 'ranking']:
            out_ = Dense(1)(rep)
        show_layer_info('Dense', out_)
        '''
        out_ = Dot( axes= [1, 1], normalize=True)([rq, rd])
        show_layer_info('Dot', out_)
        if self.config['target_mode'] == 'classification':
            out_ = Dense(2, activation='softmax')(out_)
            show_layer_info('Dense', out_)

        model = Model(inputs=[query, doc], outputs=[out_])
        return model