Question

我想加载预训练模型并继续使用此模型进行训练用于保存模型的标准代码段（pretrain.py）：

tf.reset_default_graph()

# tf Graph input
X = tf.placeholder("float", [None, n_input])
Y = tf.placeholder("float", [None, n_classes])

mlp_layer_name = ['h1', 'b1', 'h2', 'b2', 'h3', 'b3', 'w_o', 'b_o']
logits = multilayer_perceptron(X, n_input, n_classes, mlp_layer_name)

loss_op = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=Y), name='loss_op')
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
train_op = optimizer.minimize(loss_op, name='train_op')

saver = tf.train.Saver()

with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())

    # Training cycle
    for epoch in range(training_epochs):
        avg_cost = 0.

        # Loop over all batches
        for i in range(total_batch):
            batch_x, batch_y = next(train_generator)

            # Run optimization op (backprop) and cost op (to get loss value)
            _, c = sess.run([train_op, loss_op], feed_dict={X: batch_x,
                                                            Y: batch_y})
            # Compute average loss
            avg_cost += c / total_batch

        print("Epoch: {:3d}, cost = {:.6f}".format(epoch+1, avg_cost))

    print("Optimization Finished!")
    saver.save(sess, 'model')
    print("Model saved")

现在加载预训练模型并继续训练（continue.py）。

# tf Graph input
X = tf.placeholder("float", [None, n_input])
Y = tf.placeholder("float", [None, n_classes])
mlp_layer_name = ['h1', 'b1', 'h2', 'b2', 'h3', 'b3', 'w_o', 'b_o']
logits = multilayer_perceptron(X, n_input, n_classes, mlp_layer_name)
loss_op = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=Y), name='loss_op')
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
train_op = optimizer.minimize(loss_op, name='train_op')

with tf.Session() as sess:
    saver = tf.train.import_meta_graph('model.meta')
    saver.restore(sess, tf.train.latest_checkpoint('./')) # search for checkpoint file

    graph = tf.get_default_graph()

    for epoch in range(training_epochs):
        avg_cost = 0.

        # Loop over all batches
        for i in range(total_batch):
            batch_x, batch_y = next(train_generator)

            # Run optimization op (backprop) and cost op (to get loss value)
            _, c = sess.run([train_op, loss_op], feed_dict={X: batch_x,
                                                            Y: batch_y})
            # Compute average loss
            avg_cost += c / total_batch

        print("Epoch: {:3d}, cost = {:.6f}".format(epoch+1, avg_cost))

但它显示以下错误：

tensorflow.python.framework.errors_impl.FailedPreconditionError：试图使用未初始化的值h1 [[Node：h1 / read = IdentityT = DT_FLOAT，_ class = [＆＃34; loc：@ h1＆＃34;]，_ device =＆＃34; / job：localhost / replica：0 / task：0 / cpu：0＆＃34;]]

以下是我的问题：
1.在许多tensorflow的教程中，它使用get_tensor_by_name()来加载权重和偏差。在这里，我不想获得权重和偏见。我只想加载模型并继续训练。
错误表明张量是未初始化的。但是，我认为saver.restore(sess, tf.train.latest_checkpoint('./'))应该成功地加载了权重和偏见这是multilayer_perceptron()，如果它有助于说明我的questoins。

def multilayer_perceptron(x, n_input, n_classes, name):
    n_hidden_1 = 512
    n_hidden_2 = 256
    n_hidden_3 = 128
    # Store layers weight & bias
    weights = {
        'h1' : tf.get_variable(name[0], initializer=tf.random_normal([n_input, n_hidden_1])),
        'h2' : tf.get_variable(name[2], initializer=tf.random_normal([n_hidden_1, n_hidden_2])),
        'h3' : tf.get_variable(name[4], initializer=tf.random_normal([n_hidden_2, n_hidden_3])),
        'w_o': tf.get_variable(name[6], initializer=tf.random_normal([n_hidden_3, n_classes]))
    }
    biases = {
        'b1' : tf.get_variable(name[1], initializer=tf.random_normal([n_hidden_1])),
        'b2' : tf.get_variable(name[3], initializer=tf.random_normal([n_hidden_2])),
        'b3' : tf.get_variable(name[5], initializer=tf.random_normal([n_hidden_3])),
        'b_o': tf.get_variable(name[7], initializer=tf.random_normal([n_classes]))
    }

    layer_1 = tf.nn.relu(tf.add(tf.matmul(x      , weights['h1']), biases['b1']))
    layer_1 = tf.layers.dropout(layer_1, rate=0.5, training=True)
    layer_2 = tf.nn.relu(tf.add(tf.matmul(layer_1, weights['h2']), biases['b2']))
    layer_2 = tf.layers.dropout(layer_2, rate=0.3, training=True)
    layer_3 = tf.nn.relu(tf.add(tf.matmul(layer_2, weights['h3']), biases['b3']))
    layer_3 = tf.layers.dropout(layer_3, rate=0.1, training=True)
    out_layer = tf.matmul(layer_3, weights['w_o']) + biases['b_o']
    return out_layer

Answer 1

我想我找到了答案。关键是，如果tf.train.import_meta_graph()已使用saver.restore(sess, tf.train.latest_checkpoint('./'))，则无需致电# tf Graph input X = tf.placeholder("float", [None, n_input]) Y = tf.placeholder("float", [None, n_classes]) mlp_layer_name = ['h1', 'b1', 'h2', 'b2', 'h3', 'b3', 'w_o', 'b_o'] logits = multilayer_perceptron(X, n_input, n_classes, mlp_layer_name) loss_op = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=Y), name='loss_op') optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate) train_op = optimizer.minimize(loss_op, name='train_op') with tf.Session() as sess: saver = tf.train.Saver() saver.restore(sess, tf.train.latest_checkpoint('./')) # search for checkpoint file graph = tf.get_default_graph() for epoch in range(training_epochs): avg_cost = 0. # Loop over all batches for i in range(total_batch): batch_x, batch_y = next(train_generator) # Run optimization op (backprop) and cost op (to get loss value) _, c = sess.run([train_op, loss_op], feed_dict={X: batch_x, Y: batch_y}) # Compute average loss avg_cost += c / total_batch print("Epoch: {:3d}, cost = {:.6f}".format(epoch+1, avg_cost))。这是我的代码。

registerbackbuttonaction

如何加载张量流模型并继续训练

1 个答案: