level1_model.py 文件源码

def _get_initial_lstm(self, features):
        with tf.variable_scope('level1/initial_lstm'):
            features_mean = tf.reduce_mean(features, 1)

            w2_init = np.transpose(self.model_load['/init_network/weight2'][:], (1, 0))
            b2_init = self.model_load['/init_network/bias2'][:]

            w_1_ = np.transpose(self.model_load['/init_network/weight1'][:], (1, 0))
            w_1 = tf.get_variable('w_w1', initializer=w_1_)
            b_1 = tf.get_variable('w_b1', initializer=self.model_load['/init_network/bias1'][:])
            h1 = tf.nn.relu(tf.matmul(features_mean, w_1) + b_1)
            # todo: this dropout can be added later
            # if self.dropout:
            # h1 = tf.nn.dropout(h1, 0.5)

            w_h = tf.get_variable('w_h', initializer=w2_init[:, self.H:])
            b_h = tf.get_variable('b_h', initializer=b2_init[self.H:])
            h = tf.nn.tanh(tf.matmul(h1, w_h) + b_h)

            w_c = tf.get_variable('w_c', initializer=w2_init[:, :self.H])
            b_c = tf.get_variable('b_c', initializer=b2_init[:self.H])
            c = tf.nn.tanh(tf.matmul(h1, w_c) + b_c)

            return c, h