Question

我编辑了MNIST data set的TensorFlow示例，在我的电脑上准确率达到了90％，并尝试在CIFAR-10 dataset上使用类似的代码。然而，准确度仅为0-15％，从未达到20％。

import six.moves.cPickle as cPickle
from pprint import pprint

def unpickle():
  dict=[]
  fo = open(r'C:\train\cifar-10-batches-py\data_batch_1', 'rb')
  dict.append(cPickle.load(fo, encoding='latin1'))
  fo.close()
  return dict

def testpickle():
  afo = open(r'C:\train\cifar-10-batches-py\test_batch', 'rb')
  adict = cPickle.load(afo, encoding='latin1')
  afo.close()
  return adict

dt=unpickle()
import tensorflow as tf
import numpy as np
datadt=np.empty([5,10000,1024])

####to arrange input data properly####
for p in range(len(dt)):
  print(p)
  for i in range(len(dt[p]["labels"])):
    a=dt[p]["labels"][i]
    dt[p]["labels"][i]=[0,0,0,0,0,0,0,0,0,0]
    dt[p]["labels"][i][a]=1
    datadt[p][i]=(dt[p]["data"][i].tolist()[:1024])

tdt=testpickle()

###arrange test data properly###
testdt=np.empty([10000,1024])
for i in range(len(tdt["labels"])):
  a=tdt["labels"][i]
  tdt["labels"][i]=[0,0,0,0,0,0,0,0,0,0]
  tdt["labels"][i][a]=1
  testdt[i]=(tdt["data"][i].tolist()[:1024])

sess = tf.InteractiveSession()

x = tf.placeholder(tf.float32, shape=[None, 1024])
y_ = tf.placeholder(tf.float32, shape=[None, 10])

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

W_conv1=weight_variable([5,5,1,8])
b_conv1=bias_variable([8])
x_image=tf.reshape(x,[-1,32,32,1])

h_conv1=tf.nn.relu(conv2d(x_image,W_conv1)+b_conv1)
h_pool1=max_pool_2x2(h_conv1)

W_conv2 = weight_variable([5, 5, 8, 16])
b_conv2 = bias_variable([16])
h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)
h_pool2 = max_pool_2x2(h_conv2)

W_fc1 = weight_variable([8 * 8 * 16, 32])
b_fc1 = bias_variable([32])
h_pool2_flat = tf.reshape(h_pool2, [-1, 8*8*16])
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)

keep_prob = tf.placeholder(tf.float32)
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)

W_fc2 = weight_variable([32, 10])
b_fc2 = bias_variable([10])
y_conv = tf.matmul(h_fc1_drop, W_fc2) + b_fc2

cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y_, logits=y_conv))

train_step = tf.train.AdamOptimizer(0.5).minimize(cross_entropy)

sess.run(tf.global_variables_initializer())
tshaped_x=testdt
tshaped_y=tdt["labels"]
correct_prediction = tf.equal(tf.argmax(y_conv,1), tf.argmax(y_,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

k=100
import random
for i in range(len(dt)):
  for u in range(99):
    shaped_x=datadt[i][(u*k):(u*k+k)]#np.reshape(dt["data"][i], (-1,3072))
    shaped_y=dt[i]["labels"][(u*k):(u*k+k)]#np.reshape(dt["labels"][i], (-1,10))
    train_step.run(feed_dict={x: shaped_x, y_:shaped_y,keep_prob:0.5})
    r=random.randint(0,9000)
    print(accuracy.eval(feed_dict={x:tshaped_x[r:r+50], y_:tshaped_y[r:r+50],keep_prob:1.0}))

代码的神经网络部分与样本非常相似，但结果如下：

0.08
0.06
0.12
0.2
0.14
0.14
0.1
0.12
0.1
0.1
0.04
0.14
0.14

（为方便起见，我只使用每个图片数据RGB的红色数据作为输入 - 原始3072 int表示R，G，B和i使用前1024个整数，如dt[p]["data"][i].tolist()[:1024]所示）

我一直在寻找不同网站的答案，但遗憾地失败了。作为Tensorflow的初学者，抱歉天真。感谢您提前慷慨解囊！

P.S。无论我如何将AdamOptimizer的学习率从0.0001改为999，结果都是一样的（非常相似）

Answer 1

在权重的初始化中，降低标准偏差，比如大约0.01，或者更多地调整它。您的网络将开始学习！

请参阅：https://stats.stackexchange.com/questions/198840/cnn-xavier-weight-initialization

请记住，这些是给出的差异，我们需要提供标准偏差，以便平均根。

Tensorflow - CIFAR-10数据集上的振荡学习率

1 个答案: