python - tensorflow 线性分类器未训练

Question

我正在尝试为 MNIST 数据创建一个简单的线性分类器，但我无法降低损失。可能是什么问题呢？ 这是我的代码:

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

class LinearClassifier(object):
    def __init__(self):
        print("LinearClassifier loading MNIST")
        self._mnist = input_data.read_data_sets("mnist_data/", one_hot = True)
        self._buildGraph()

    def _buildGraph(self):
        self._tf_TrainX = tf.placeholder(tf.float32, [None, self._mnist.train.images.shape[1]])    
        self._tf_TrainY = tf.placeholder(tf.float32, [None, self._mnist.train.labels.shape[1]]) 

        self._tf_Weights = tf.Variable(tf.random_normal([784,10]), tf.float32)
        self._tf_Bias = tf.Variable(tf.zeros([10]), tf.float32)
        self._tf_Y = tf.nn.softmax(tf.matmul(self._tf_TrainX, self._tf_Weights) + self._tf_Bias)

        self._tf_Loss = tf.reduce_mean(-tf.reduce_sum(self._tf_TrainY * tf.log(self._tf_Y), reduction_indices=[1]))
        self._tf_TrainStep = tf.train.GradientDescentOptimizer(learning_rate=0.001).minimize(self._tf_Loss)

        self._tf_CorrectGuess = tf.equal(tf.argmax(self._tf_Y, 1), tf.arg_max(self._tf_TrainY, 1))
        self._tf_Accuracy = tf.reduce_mean(tf.cast(self._tf_CorrectGuess, tf.float32))

        self._tf_Initializers = tf.global_variables_initializer()

    def train(self, epochs, batch_size):
        self._sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))
        self._sess.run(self._tf_Initializers)

        for i in range(epochs):
            batchX, batchY = self._mnist.train.next_batch(batch_size)
            self._loss, _, self._accurracy = self._sess.run([self._tf_Loss, self._tf_TrainStep, self._tf_Accuracy], feed_dict ={self._tf_TrainX: batchX, self._tf_TrainY: batchY})
            print("Epoch: {0}, Loss: {1}, Accuracy: {2}".format(i, self._loss, self._accurracy))

当我通过以下方式运行此命令时:

lc = LinearClassifier()
lc.train(1000, 100)

...我得到这样的东西:

Epoch: 969, Loss: 8.19491195678711, Accuracy: 0.17999999225139618
Epoch: 970, Loss: 9.09421157836914, Accuracy: 0.1899999976158142
....
Epoch: 998, Loss: 7.865959167480469, Accuracy: 0.17000000178813934
Epoch: 999, Loss: 9.281349182128906, Accuracy: 0.10999999940395355

tf.train.GradientDescentOptimizer 未正确训练我的权重和偏差的原因可能是什么？

Answer 1

最主要的是你的学习率(0.001)太低了。我在将其更改为 0.5 后运行了此命令，就像他们在 mnist tensorflow tutorial 中所做的那样我得到的准确性和损失更像是:

Epoch: 997, Loss: 0.6437355875968933, Accuracy: 0.8999999761581421
Epoch: 998, Loss: 0.6129786968231201, Accuracy: 0.8899999856948853
Epoch: 999, Loss: 0.6442205905914307, Accuracy: 0.8999999761581421

另一件有点不寻常的事情是在你的原始代码中你有这个

self._tf_Y = tf.nn.softmax(tf.matmul(self._tf_TrainX, self._tf_Weights) + self._tf_Bias)
self._tf_Loss = tf.reduce_mean(-tf.reduce_sum(self._tf_TrainY * tf.log(self._tf_Y), reduction_indices=[1]))

在这种情况下，您将执行两次 softmax。我在更改之前确实运行过它，火车准确率约为 85%，所以它确实有所不同。另外，进行两次 softmax 理论上的解释性较差。

最后，他们在教程中提到，使用上面的softmax形式，-reduce_sum(label * log(y))，在数值上不稳定，所以最好使用内置的softmax层它计算解析上等效但数值更稳定的 softmax。应用这两项更改后，受影响的行如下所示:

self._tf_Y = tf.matmul(self._tf_TrainX, self._tf_Weights) + self._tf_Bias
self._tf_Loss = tf.reduce_mean(
  tf.nn.softmax_cross_entropy_with_logits(labels=self._tf_TrainY, logits=self._tf_Y))