更改优化器参数后在 Tensorflow 中恢复模型时出错

Question

我在 Tensorflow 中训练了一个模型。在训练过程中，我在优化器中设置了一个 var_list，换句话说，我在 CNN 之上训练了一个 GRU。这是优化器的代码：

with tf.name_scope('optimizer'):
    update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
    with tf.control_dependencies(update_ops):
        optimizer = tf.train.AdamOptimizer(0.0001).minimize(MSE, var_list=gru_output_var_list)

然后，在训练并保存检查点中的变量之后，我试图从优化器中删除var_list，以便能够微调整个网络，使用 GRU 进行卷积层。但是，这会引发错误：

Key weight_fc_sig/Adam_1 not found in checkpoint

其中weight_fc_sig 是模型中变量之一的名称。

我通读了github，发现优化器的状态以及检查点中的变量都被保存了。所以，我想知道如何解决这个问题，换句话说，我需要知道如何重置优化器的状态。

非常感谢任何帮助！

Answer 1

首先，我在 tensorflow 中构建了一个模型，然后通过以下方式将带有变量的图形保存到检查点中：

saver = tf.train.Saver()
saver.save(sess, model_path + "ckpt")

所以当我检查通过以下方式存储的变量列表时：

from tensorflow.python import pywrap_tensorflow
model_path = 'C:/Users/user/PycharmProjects/TensorflowDifferentProjects/MNIStDataset/tensorlogs/ckpt'
reader = pywrap_tensorflow.NewCheckpointReader(model_path)
var_to_shape_map = reader.get_variable_to_shape_map()

for key in sorted(var_to_shape_map):
    print("tensor_name: ", key)

我得到了以下变量列表：

tensor_name:  Adam_optimizer/beta1_power
tensor_name:  Adam_optimizer/beta2_power
tensor_name:  conv1/biases
tensor_name:  conv1/biases/Adam
tensor_name:  conv1/biases/Adam_1
tensor_name:  conv1/weights
tensor_name:  conv1/weights/Adam
tensor_name:  conv1/weights/Adam_1
tensor_name:  conv2/biases
tensor_name:  conv2/biases/Adam
tensor_name:  conv2/biases/Adam_1
tensor_name:  conv2/weights
tensor_name:  conv2/weights/Adam
tensor_name:  conv2/weights/Adam_1
tensor_name:  fc1/biases
tensor_name:  fc1/biases/Adam
tensor_name:  fc1/biases/Adam_1
tensor_name:  fc1/weights
tensor_name:  fc1/weights/Adam
tensor_name:  fc1/weights/Adam_1
tensor_name:  fc2/biases
tensor_name:  fc2/biases/Adam
tensor_name:  fc2/biases/Adam_1
tensor_name:  fc2/weights
tensor_name:  fc2/weights/Adam
tensor_name:  fc2/weights/Adam_1

当我再次训练相同的模型时，但这一次，我只将权重和偏差列表传递给保护程序： saver = tf.train.Saver(var_list=lst_vars)，然后我打印出保存的权重和偏差列表， 我得到了以下列表：

tensor_name:  conv1/biases
tensor_name:  conv1/weights
tensor_name:  conv2/biases
tensor_name:  conv2/weights
tensor_name:  fc1/biases
tensor_name:  fc1/weights
tensor_name:  fc2/biases
tensor_name:  fc2/weights

现在，当我尝试再次运行相同的模型，但删除了要恢复的变量列表时，我现在有了这个保护程序：

saver = tf.train.Saver(), 

我遇到了以下错误：

Key fc2/weights/Adam_1 not found in checkpoint.

因此，解决方案是明确提及我需要恢复的变量列表。换句话说，即使我只 保存了我需要存储的权重和偏差列表，在导入它们时，我应该特别提到它们，所以我应该 说：

saver = tf.train.Saver(var_list=lst_vars) 

lst_vars 是我需要恢复的变量列表，与之前的相同 打印在上面。

所以总的来说，每当我们尝试恢复图形时，如果我没有提到要恢复的变量列表， tensorflow 会查看并看到一些变量尚未存储，换句话说，只要没有列表， tensorflow 假设我们正在尝试恢复整个图，这是不正确的。我只恢复负责的部分 权重和偏差。因此，一旦提到这一点，tensorflow 就会知道我不是在初始化整个图，而是在初始化部分

现在即使我真的提到了我需要恢复的变量列表，如下所示：

saver = tf.train.Saver(var_list=lst_vars)

这不会造成任何问题

另外，我可以添加/修改传递给优化器的 var_list，这也不会造成任何问题。

同时，即使我将变量列表传递给优化器以进行处理：

with tf.name_scope('Adam_optimizer'):
    train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy, var_list=lst_vars[:3])

saver = tf.train.Saver(var_list=lst_vars)

然后我可以再次运行相同的模型，但优化器中没有 var_list 参数。这就是微调的情况。

---

现在，为了更进一步，我可以修改模型，添加更多层，但我应该记住，因为我只有 以下变量存储在检查点中：

tensor_name:  conv1/biases
tensor_name:  conv1/weights
tensor_name:  conv2/biases
tensor_name:  conv2/weights
tensor_name:  fc1/biases
tensor_name:  fc1/weights
tensor_name:  fc2/biases
tensor_name:  fc2/weights
    

我应该向保护者提及这些是我正在恢复的变量。于是我说了如下：

saver = tf.train.Saver(var_list=[lst_vars[0], lst_vars[1], lst_vars[2], lst_vars[3],
                             lst_vars[6], lst_vars[7], lst_vars[8], lst_vars[9]])

在这种情况下，没有问题，代码运行良好！！！ 我也可以要求优化器训练新模型，也许训练某些参数，我的意思是权重和偏差等等......

另外，请注意我可以将整个模型保存为：

saver = tf.train.Saver()

然后恢复部分模型（通过再次运行模型，然后通过：

saver = tf.train.Saver(var_list=lst_vars))

另外，我可以修改模型并添加更多的卷积层。所以我可以微调模型，只要我确切地提到什么 我正在恢复的变量。例如：

saver = tf.train.Saver(var_list=[lst_vars[0], lst_vars[1], lst_vars[2], lst_vars[3],
                             lst_vars[6], lst_vars[7], lst_vars[8], lst_vars[9]])

所有这些解释都是因为我认为优化器可能存在一些问题，我需要知道如何休息。在 github 上提出了一个问题，确切地说是如何让优化器休息，这也是我得出所有这些结论的原因

---

下面是代码供感兴趣的人参考：

from tensorflow.examples.tutorials.mnist import input_data
import tensorflow as tf
import os


def conv2d(x, w):
    return tf.nn.conv2d(x, w, strides=[1, 1, 1, 1], padding='SAME')


def max_pool_2x2(x):
    return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')


def weight_variable(shape, name):
    initial = tf.truncated_normal(shape, stddev=0.1)
    return tf.Variable(initial, name=name)


def bias_variable(shape, name):
    initial = tf.constant(0.1, shape=shape)
    return tf.Variable(initial, name=name)


def deepnn(x):
    with tf.name_scope('reshape'):
        x_image = tf.reshape(x, [-1, 28, 28, 1])

    # First convolutional layer, maps one grayscale image to 32 feature maps.
    with tf.name_scope('conv1'):
        w_conv1 = weight_variable([5, 5, 1, 32], name='weights')
        b_conv1 = bias_variable([32], name='biases')
        h_conv1 = tf.nn.relu(conv2d(x_image, w_conv1) + b_conv1)

    # Pooling layer, downsampling by 2X
    with tf.name_scope('pool1'):
        h_pool1 = max_pool_2x2(h_conv1)

    # Second convolutional layer -- maps 32 feature maps to 64
    with tf.name_scope('conv2'):
        w_conv2 = weight_variable([5, 5, 32, 64], name='weights')
        b_conv2 = bias_variable([64], name='biases')
        h_conv2 = tf.nn.relu(conv2d(h_pool1, w_conv2) + b_conv2)

    # Second pooling layer
    with tf.name_scope('pool2'):
        h_pool2 = max_pool_2x2(h_conv2)

    # Fully connected layer 1 -- after 2 round of downsampling, our 28 x 28 image is
    # down to 7 x 7 x 64 feature maps -- maps this to 1024 features.
    with tf.name_scope('fc1'):
        w_fc1 = weight_variable([7 * 7 * 64, 1024], name='weights')
        b_fc1 = bias_variable([1024], name='biases')

        h_pool2_flat = tf.reshape(h_pool2, [-1, 7 * 7 * 64])
        h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, w_fc1) + b_fc1)

    # Dropout - control the complexity of the model, prevents co-adaptation of features
    with tf.name_scope('dropout'):
        keep_prob = tf.placeholder(tf.float32)
        h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)

    # Map the 1024 features to 10 classes, one for each digit.
    with tf.name_scope('fc2'):
        w_fc2 = weight_variable([1024, 10], name='weights')
        b_fc2 = bias_variable([10], name='biases')

        y_conv = tf.matmul(h_fc1_drop, w_fc2) + b_fc2

    return y_conv, keep_prob


mnist = input_data.read_data_sets('MNIST_data', one_hot=True)

# Create the model
x = tf.placeholder(tf.float32, [None, 784])

# Define loss and optimizer
y_ = tf.placeholder(tf.float32, [None, 10])

# Build the graph for the deep net
y_conv, keep_prob = deepnn(x)

with tf.name_scope('loss'):
    cross_entropy = tf.nn.softmax_cross_entropy_with_logits(labels=y_, logits=y_conv)

cross_entropy = tf.reduce_mean(cross_entropy)

# Note that this list of variables only include the weights and biases in the model.
lst_vars = []
for v in tf.global_variables():
    lst_vars.append(v)
    print(v.name, '....')

with tf.name_scope('Adam_optimizer'):
    train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)

with tf.name_scope('accuracy'):
    correct_prediction = tf.equal(tf.arg_max(y_conv, 1), tf.arg_max(y_, 1))
    correct_prediction = tf.cast(correct_prediction, tf.float32)

accuracy = tf.reduce_mean(correct_prediction)

model_path = 'C:/Users/user/PycharmProjects/TensorflowDifferentProjects/MNIStDataset/tensorlogs/'

saver = tf.train.Saver(var_list=lst_vars)

train_writer = tf.summary.FileWriter(model_path + "EventsFile/")
train_writer.add_graph(tf.get_default_graph())

for v in tf.global_variables():
    print(v.name)

# Note that a session is created within a with block so that it is destroyed
# once the block has been exited.
with tf.Session() as sess:
    print('all variables initialized!!')

    sess.run(tf.global_variables_initializer())

    ckpt = tf.train.get_checkpoint_state(
        os.path.dirname(model_path))
    if ckpt and ckpt.model_checkpoint_path:
        saver.restore(sess, ckpt.model_checkpoint_path)
        print('checkpoints are saved!!!')
    else:
        print('No stored checkpoints')

    for i in range(700):
        batch = mnist.train.next_batch(50)
        if i % 100 == 0:
            train_accuracy = accuracy.eval(feed_dict={x: batch[0], y_: batch[1], keep_prob: 1.0})
            print('step %d, training accuracy %g' % (i, train_accuracy))

        train_step.run(feed_dict={x: batch[0], y_: batch[1], keep_prob: 0.5})

    print('test accuracy %g' % accuracy.eval(feed_dict={
        x: mnist.test.images, y_: mnist.test.labels, keep_prob: 1.0}))

    save_path = saver.save(sess, model_path + "ckpt")

以及修改后的模型（我在其中添加了另一个卷积层）：

from tensorflow.examples.tutorials.mnist import input_data
import tensorflow as tf
import os

def conv2d(x, w):
    return tf.nn.conv2d(x, w, strides=[1, 1, 1, 1], padding='SAME')

def max_pool_2x2(x):
    return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')

def weight_variable(shape, name):
    initial = tf.truncated_normal(shape, stddev=0.1)
    return tf.Variable(initial, name=name)

def bias_variable(shape, name):
    initial = tf.constant(0.1, shape=shape)
    return tf.Variable(initial, name=name)

def deepnn(x):

    with tf.name_scope('reshape'):
        x_image = tf.reshape(x, [-1, 28, 28, 1])

    # First convolutional layer, maps one grayscale image to 32 feature maps.
    with tf.name_scope('conv1'):
        w_conv1 = weight_variable([5, 5, 1, 32], name='weights')
        b_conv1 = bias_variable([32], name='biases')
        h_conv1 = tf.nn.relu(conv2d(x_image, w_conv1) + b_conv1)

    # Pooling layer, downsampling by 2X
    with tf.name_scope('pool1'):
        h_pool1 = max_pool_2x2(h_conv1)

    # Second convolutional layer -- maps 32 feature maps to 64
    with tf.name_scope('conv2'):
        w_conv2 = weight_variable([5, 5, 32, 64], name='weights')
        b_conv2 = bias_variable([64], name='biases')
        h_conv2 = tf.nn.relu(conv2d(h_pool1, w_conv2) + b_conv2)

    # Second pooling layer
    with tf.name_scope('pool2'):
        h_pool2 = max_pool_2x2(h_conv2)

    with tf.name_scope('conv3'):
        w_conv3 = weight_variable([1, 1, 64, 64], name='weights')
        b_conv3 = bias_variable([64], name='biases')
        h_conv3 = tf.nn.relu(conv2d(h_pool2, w_conv3) + b_conv3)

    # Fully connected layer 1 -- after 2 round of downsampling, our 28 x 28 image is
    # down to 7 x 7 x 64 feature maps -- maps this to 1024 features.
    with tf.name_scope('fc1'):
        w_fc1 = weight_variable([7 * 7 * 64, 1024], name='weights')
        b_fc1 = bias_variable([1024], name='biases')

        h_conv3_flat = tf.reshape(h_conv3, [-1, 7 * 7 * 64])
        h_fc1 = tf.nn.relu(tf.matmul(h_conv3_flat, w_fc1) + b_fc1)

    # Dropout - control the complexity of the model, prevents co-adaptation of features
    with tf.name_scope('dropout'):
        keep_prob = tf.placeholder(tf.float32)
        h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)

    # Map the 1024 features to 10 classes, one for each digit.
    with tf.name_scope('fc2'):
        w_fc2 = weight_variable([1024, 10], name='weights')
        b_fc2 = bias_variable([10], name='biases')

        y_conv = tf.matmul(h_fc1_drop, w_fc2) + b_fc2

    return y_conv, keep_prob

mnist = input_data.read_data_sets('MNIST_data', one_hot=True)

# Create the model
x = tf.placeholder(tf.float32, [None, 784])

# Define loss and optimizer
y_ = tf.placeholder(tf.float32, [None, 10])

# Build the graph for the deep net
y_conv, keep_prob = deepnn(x)

with tf.name_scope('loss'):
    cross_entropy = tf.nn.softmax_cross_entropy_with_logits(labels=y_, logits=y_conv)

cross_entropy = tf.reduce_mean(cross_entropy)

# Note that this list of variables only include the weights and biases in the model.
lst_vars = []
for v in tf.global_variables():
    lst_vars.append(v)
    print(v.name, '....')

with tf.name_scope('Adam_optimizer'):
    train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)

with tf.name_scope('accuracy'):
    correct_prediction = tf.equal(tf.arg_max(y_conv, 1), tf.arg_max(y_, 1))
    correct_prediction = tf.cast(correct_prediction, tf.float32)

accuracy = tf.reduce_mean(correct_prediction)

model_path = 'C:/Users/user/PycharmProjects/TensorflowDifferentProjects/MNIStDataset/tensorlogs/'

saver = tf.train.Saver(var_list=[lst_vars[0], lst_vars[1], lst_vars[2], lst_vars[3],
                                 lst_vars[6], lst_vars[7], lst_vars[8], lst_vars[9]])

train_writer = tf.summary.FileWriter(model_path + "EventsFile/")
train_writer.add_graph(tf.get_default_graph())

for v in tf.global_variables():
    print(v.name)

# Note that a session is created within a with block so that it is destroyed
# once the block has been exited.
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())

    print('all variables initialized!!')

    ckpt = tf.train.get_checkpoint_state(
        os.path.dirname(model_path))
    if ckpt and ckpt.model_checkpoint_path:
        saver.restore(sess, ckpt.model_checkpoint_path)
        print('checkpoints are saved!!!')
    else:
        print('No stored checkpoints')

    for i in range(700):
        batch = mnist.train.next_batch(50)
        if i % 100 == 0:
            train_accuracy = accuracy.eval(feed_dict={x: batch[0], y_: batch[1], keep_prob: 1.0})
            print('step %d, training accuracy %g' % (i, train_accuracy))

        train_step.run(feed_dict={x: batch[0], y_: batch[1], keep_prob: 0.5})

    print('test accuracy %g' % accuracy.eval(feed_dict={
        x: mnist.test.images, y_: mnist.test.labels, keep_prob: 1.0}))

    save_path = saver.save(sess, model_path + "ckpt")