python - tensorflow 模型精度没有增加

Question

我目前正在 Udacity 上学习深度学习类(class)，并正在尝试完成第四项作业，您应该创建自己的模型，并了解在 noMINST 数据集上可以实现的最佳准确度。

我尝试实现VGG 16模型，但遇到了一些问题，最初，损失直接变为nan，因此我将最后一个激活函数从relu更改为sigmoid，但现在精度不高改进并停留在 0-6% 左右，所以我猜测我的实现是错误的，但我似乎看不到错误，我将非常感谢任何帮助或建议!

贝娄是我的完整代码，除了在数据集中读取之外，因为该代码是由 提供的，所以我猜它是正确的。

pickle_file = 'notMNIST.pickle'

with open(pickle_file, 'rb') as f:
    save = pickle.load(f)
    train_dataset = save['train_dataset']
    train_labels = save['train_labels']
    valid_dataset = save['valid_dataset']
    valid_labels = save['valid_labels']
    test_dataset = save['test_dataset']
    test_labels = save['test_labels']
    del save  # hint to help gc free up memory
    print('Training set', train_dataset.shape, train_labels.shape)
    print('Validation set', valid_dataset.shape, valid_labels.shape)
    print('Test set', test_dataset.shape, test_labels.shape)

image_size = 28
num_labels = 10
num_channels = 1  # grayscale

import numpy as np


def reformat(dataset, labels):
    dataset = dataset.reshape(
        (-1, image_size, image_size, num_channels)).astype(np.float32)
    labels = (np.arange(num_labels) == labels[:, None]).astype(np.float32)
    return dataset, labels


train_dataset, train_labels = reformat(train_dataset, train_labels)
valid_dataset, valid_labels = reformat(valid_dataset, valid_labels)
test_dataset, test_labels = reformat(test_dataset, test_labels)
print('Training set', train_dataset.shape, train_labels.shape)
print('Validation set', valid_dataset.shape, valid_labels.shape)
print('Test set', test_dataset.shape, test_labels.shape)


def accuracy(predictions, labels):
    return (100.0 * np.sum(np.argmax(predictions, 1) == np.argmax(labels, 1))
            / predictions.shape[0])


batch_size = 16
patch_size = 5
depth = 16
num_hidden = 64

graph = tf.Graph()

with graph.as_default():
    # Input data.
    tf_train_dataset = tf.placeholder(
        tf.float32, shape=(batch_size, image_size, image_size, num_channels))
    tf_train_labels = tf.placeholder(tf.float32, shape=(batch_size, num_labels))
    tf_valid_dataset = tf.constant(valid_dataset)
    tf_test_dataset = tf.constant(test_dataset)

    # Variables.
    l1_w = tf.Variable(tf.truncated_normal([3, 3, 1, 64], stddev=0.1))
    l1_b = tf.Variable(tf.zeros([64]))
    l2_w = tf.Variable(tf.truncated_normal([3, 3, 64, 64], stddev=0.1))
    l2_b = tf.Variable(tf.zeros([64]))

    l3_w = tf.Variable(tf.truncated_normal([3, 3, 64, 128], stddev=0.1))
    l3_b = tf.Variable(tf.zeros([128]))
    l4_w = tf.Variable(tf.truncated_normal([3, 3, 128, 128], stddev=0.1))
    l4_b = tf.Variable(tf.zeros([128]))

    l5_w = tf.Variable(tf.truncated_normal([3, 3, 128, 256], stddev=0.1))
    l5_b = tf.Variable(tf.zeros([256]))
    l6_w = tf.Variable(tf.truncated_normal([3, 3, 256, 256], stddev=0.1))
    l6_b = tf.Variable(tf.zeros([256]))
    l7_w = tf.Variable(tf.truncated_normal([3, 3, 256, 256], stddev=0.1))
    l7_b = tf.Variable(tf.zeros([256]))

    l8_w = tf.Variable(tf.truncated_normal([3, 3, 256, 512], stddev=0.1))
    l8_b = tf.Variable(tf.zeros([512]))
    l9_w = tf.Variable(tf.truncated_normal([3, 3, 512, 512], stddev=0.1))
    l9_b = tf.Variable(tf.zeros([512]))
    l10_w = tf.Variable(tf.truncated_normal([3, 3, 512, 512], stddev=0.1))
    l10_b = tf.Variable(tf.zeros([512]))

    l11_w = tf.Variable(tf.truncated_normal([3, 3, 512, 512], stddev=0.1))
    l11_b = tf.Variable(tf.zeros([512]))
    l12_w = tf.Variable(tf.truncated_normal([3, 3, 512, 512], stddev=0.1))
    l12_b = tf.Variable(tf.zeros([512]))
    l13_w = tf.Variable(tf.truncated_normal([3, 3, 512, 512], stddev=0.1))
    l13_b = tf.Variable(tf.zeros([512]))

    l14_w = tf.Variable(tf.truncated_normal([512, num_hidden], stddev=0.1))
    l14_b = tf.Variable(tf.constant(1.0, shape=[num_hidden]))

    l15_w = tf.Variable(tf.truncated_normal([num_hidden, num_labels], stddev=0.1))
    l15_b = tf.Variable(tf.constant(1.0, shape=[num_labels]))


    # Model.
    def model(data):
        conv_1 = tf.nn.relu(tf.nn.conv2d(data, l1_w, [1, 1, 1, 1], padding='SAME') + l1_b)
        conv_1 = tf.nn.relu(tf.nn.conv2d(conv_1, l2_w, [1, 1, 1, 1], padding='SAME') + l2_b)
        max_pool_1 = tf.nn.max_pool(conv_1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')

        conv_2 = tf.nn.relu(tf.nn.conv2d(max_pool_1, l3_w, [1, 1, 1, 1], padding='SAME') + l3_b)
        conv_2 = tf.nn.relu(tf.nn.conv2d(conv_2, l4_w, [1, 1, 1, 1], padding='SAME') + l4_b)
        max_pool_2 = tf.nn.max_pool(conv_2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')

        conv_3 = tf.nn.relu(tf.nn.conv2d(max_pool_2, l5_w, [1, 1, 1, 1], padding='SAME') + l5_b)
        conv_3 = tf.nn.relu(tf.nn.conv2d(conv_3, l6_w, [1, 1, 1, 1], padding='SAME') + l6_b)
        conv_3 = tf.nn.relu(tf.nn.conv2d(conv_3, l7_w, [1, 1, 1, 1], padding='SAME') + l7_b)
        max_pool_3 = tf.nn.max_pool(conv_3, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')

        conv_4 = tf.nn.relu(tf.nn.conv2d(max_pool_3, l8_w, [1, 1, 1, 1], padding='SAME') + l8_b)
        conv_4 = tf.nn.relu(tf.nn.conv2d(conv_4, l9_w, [1, 1, 1, 1], padding='SAME') + l9_b)
        conv_4 = tf.nn.relu(tf.nn.conv2d(conv_4, l10_w, [1, 1, 1, 1], padding='SAME') + l10_b)
        max_pool_4 = tf.nn.max_pool(conv_4, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')

        conv_5 = tf.nn.relu(tf.nn.conv2d(max_pool_4, l11_w, [1, 1, 1, 1], padding='SAME') + l11_b)
        conv_5 = tf.nn.sigmoid(tf.nn.conv2d(conv_5, l12_w, [1, 1, 1, 1], padding='SAME') + l12_b)
        conv_5 = tf.nn.sigmoid(tf.nn.conv2d(conv_5, l13_w, [1, 1, 1, 1], padding='SAME') + l13_b)
        max_pool_5 = tf.nn.max_pool(conv_5, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')

        shape = max_pool_5.get_shape().as_list()
        reshape = tf.reshape(max_pool_5, [shape[0], shape[1] * shape[2] * shape[3]])
        hidden = tf.nn.sigmoid(tf.matmul(reshape, l14_w) + l14_b)
        return tf.matmul(hidden, l15_w) + l15_b


    logits = model(tf_train_dataset)
    loss = tf.reduce_mean(
        tf.nn.softmax_cross_entropy_with_logits(labels=tf_train_labels, logits=logits))

    # Optimizer.
    optimizer = tf.train.AdamOptimizer(0.001).minimize(loss)

    # Predictions for the training, validation, and test data.
    train_prediction = tf.nn.softmax(logits)
    valid_prediction = tf.nn.softmax(model(tf_valid_dataset))
    test_prediction = tf.nn.softmax(model(tf_test_dataset))

num_steps = 1001

with tf.Session(graph=graph) as session:
    tf.global_variables_initializer().run()
    print('Initialized')
    for step in range(num_steps):
        offset = (step * batch_size) % (train_labels.shape[0] - batch_size)
        batch_data = train_dataset[offset:(offset + batch_size), :, :, :]
        batch_labels = train_labels[offset:(offset + batch_size), :]
        feed_dict = {tf_train_dataset: batch_data, tf_train_labels: batch_labels}
        _, l, predictions = session.run(
            [optimizer, loss, train_prediction], feed_dict=feed_dict)
        if step % 50 == 0:
            print('Minibatch loss at step %d: %f' % (step, l))
            print('Minibatch accuracy: %.1f%%' % accuracy(predictions, batch_labels))
            print('Validation accuracy: %.1f%%' % accuracy(
                valid_prediction.eval(), valid_labels))
print('Test accuracy: %.1f%%' % accuracy(test_prediction.eval(), test_labels))

Answer 1

我同意@cyniikal ，您的网络对于这个数据集来说似乎太复杂了。使用单层模型，我能够在训练数据上实现 93.75% 的准确率，在测试数据上实现 86.7% 的准确率。

在我的模型中，我使用了 GradientDescentOptimizer 来最小化 cross_entropy，就像您所做的那样。我还使用了 16 批量大小。

我认为你的方法和我的方法之间的主要区别在于:

1. OneHot 对标签进行编码
2. 使用单层网络而不是 VGG-16

查看此notebook with my single layer model code sample .

如果您想向神经网络添加层(网络收敛会遇到更多困难)，我强烈建议您阅读 this article on neural nets 。具体来说，由于您添加了 sigmoid 作为最后一个激活函数，我相信您正在遭受 vanishing gradient problem 的困扰。 。请参阅this page解决消失梯度。