■

cell = tf.contrib.rnn.GRUCell(EMBEDDING_SIZE)
_, encoding = tf.contrib.rnn.static_rnn(cell, word_list, dtype=tf.float32 )
with tf.variable_scope('linear_regression'):
　 　 y_shape = y.get_shape()
　 　 output_shape = y_shape[1]
　　 weights = tf.get_variable('weights', [encoding.get_shape()[1], output_shape])
　 　 bias = tf.get_variable('bias', [output_shape])
　 　 with tf.op_scope([encoding, y], name, "mean_squared_error_regressor"):
　 　 　 predictions = tf.nn.xw_plus_b(encoding, weights, bias)
　　 　 predictions_dict = { "prediction": predictions }
　　 　 if len(y.get_shape()) == 1:
　 　 　 　 y = tf.reshape(y, [-1, 1])
　 　 　 loss = tf.losses.mean_squared_error( y , predictions)

　 eval_metric_ops = {
　 "rmse": tf.metrics.root_mean_squared_error( y, predictions)
　 }
　 train_op = tf.contrib.layers.optimize_loss( loss, tf.contrib.framework.get_global_step(), optimizer='Adam', learning_rate=0.01)

　 return model_fn_lib.ModelFnOps(
　 mode=mode,
　 predictions=predictions_dict,
　 loss=loss,
　 train_op=train_op,
　 eval_metric_ops=eval_metric_ops)