Tensorflow - CIFAR-10数据集的振荡学习率

Question

我编辑了TensorFlow样本MNIST data set，它在我的电脑上达到了〜90％的准确度，并尝试使用CIFAR-10 dataset上的类似代码。然而，准确度只有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})) 

代码的神经网络的部分是非常相似的样品，然而结果：

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 

（为了方便起见我只是用各画面数据RGB的红色的数据作为输入 - 原来3072 INT表示R，G，B，并且我使用了前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，但是很失败。作为Tensorflow的初学者，抱歉太天真了。感谢您的慷慨帮助！

P.S.无论我如何将AdamOptimizer的学习率从0.0001改为999，结果都是相同的（非常相似）

Answer 1

在初始化权重时，降低标准偏差，比如0.01左右，或者更多地调整它。你的网络将开始学习！

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

记住那些给定的差异，我们需要养活的标准偏差，从而平方根的。