Question

在张量流中，训练modelA(including LSTMcells)之后，如何使用modelA的权重和偏差来初始化modelB(including LSTMcells)的权重和偏差。

这是model structures."()"是num_units。

modelA:
LSTM(20)-LSTM(3)※1

modelB:
LSTM(2)-LSTM(20)-LSTM(3)

([LSTM(2)]←added in modelB)

([LSTM(20)-LSTM(3)]←same in modelA※2)

我想使用※1的参数初始化※2的参数，然后重新训练模型B。

我的测试代码在这里。

def build_model_A():
    X = tf.placeholder(tf.float32, [None,None,1], name='X')

    lstm_2 = tf.nn.rnn_cell.BasicLSTMCell(20, name='blstmc2')#
    lstm_3 = tf.nn.rnn_cell.BasicLSTMCell(3, name='blstmc3')#

    lstm_out_2,lstm_out_2_state = tf.nn.dynamic_rnn(lstm_2,X,dtype=tf.float32,time_major=False)#

    lstm_out_3,lstm_out_3_state = tf.nn.dynamic_rnn(lstm_3,lstm_out_2,dtype=tf.float32,time_major=False)

    ##loss##
    p_loss = tf.square(tf.subtract(lstm_out_3, 0.5))

    # Optimization
    pred_loss = tf.reduce_sum(p_loss, name='pred_loss')

    pred_train_op = tf.train.AdamOptimizer().minimize(pred_loss, name='pred_train_op')

    return


def build_model_B():
    X = tf.placeholder(tf.float32, [None,None,1], name='X')

    lstm_1 = tf.nn.rnn_cell.BasicLSTMCell(2, name='blstmc1')#
    lstm_2 = tf.nn.rnn_cell.BasicLSTMCell(20, name='blstmc2')#
    lstm_3 = tf.nn.rnn_cell.BasicLSTMCell(3, name='blstmc3')#

    lstm_out_1,lstm_out_1_state = tf.nn.dynamic_rnn(lstm_1,X,dtype=tf.float32,time_major=False)#

    lstm_out_2,lstm_out_2_state = tf.nn.dynamic_rnn(lstm_2,lstm_out_1,dtype=tf.float32,time_major=False)

    lstm_out_3,lstm_out_3_state = tf.nn.dynamic_rnn(lstm_3,lstm_out_2,dtype=tf.float32,time_major=False)

    ##loss##
    p_loss = tf.square(tf.subtract(lstm_out_3, 0.5))

    # Optimization
    pred_loss = tf.reduce_sum(p_loss, name='pred_loss')


    pred_train_op = tf.train.AdamOptimizer().minimize(pred_loss, name='pred_train_op')

    return

def train_and_evaluate(sess,num_batches=None,input_data=None):

    pred_loss = sess.graph.get_tensor_by_name('pred_loss:0')

    sess.run(tf.global_variables_initializer())##

    saver = tf.train.Saver()###for model save

    for i in range(num_batches):

        train_optimizer = "pred_train_op"
        _, pa = sess.run([train_optimizer, pred_loss],
                           feed_dict={'X:0': input_data})

        if i%1000==0:
            [pa_A] = sess.run([pred_loss], feed_dict={'X:0': input_data})

            print("**Train_acc("+str(i)+")**")
            print("accuracy:"+str(pa_A))

            if i==10000:
                saver.save(sess, "model/test_sess_modelA/trained")

def train_fnc():

    tf.reset_default_graph()
    build_model_A()
    sess = tf.InteractiveSession()

    ##make noise inputdata X_a##
    x_sa = np.array([[1],[3],[5],[7],[9]])
    X_a = []
    for i in range(10):
        X_a.append(x_sa)
    X_a = np.array(X_a)

    train_batch_num = 10000

    train_and_evaluate(sess,num_batches=train_batch_num,input_data=X_a)

    return

def test_func():

    tf.reset_default_graph()
    build_model_B()
    saver = tf.train.Saver()
    sess = tf.Session()

    saver.restore(sess, "model/test_sess_modelA/trained")

def main_fnc():

    train_fnc()
    test_fnc()

我尝试使用“ tf.train.Saver（）”，但是在恢复模型B中的参数时，出现错误“ NotFoundError（请参阅上面的回溯）：关键rnn / blstmc1 / bias在检查点中找不到”。

而且，我还尝试使用“ saver = tf.train.Saver（{'blstmc2'：lstm_2，'blstmc3'：lstm_3}）“，但是我遇到了错误（无法将BasicLSTMCell转换为张量）。 / p>

在tensorflow中，如何将BasicLSTMCell从一个模型保存和恢复到另一个模型？

0 个答案: