數據也就用了作者上傳的60000張Discuz驗證碼。作者是創建了一個 類 封裝了所有的變量和函數，我看了他的代碼之后自己嘗試着不用類去實現該網絡。
作者說自己可以訓練到90%以上的精度。然而我看了他的代碼后發現，作者是用訓練過的數據來進行測試，即訓練集和測試集是一樣的。
我想着，測試集應該是不能參與訓練過程中的，比如說我們在做mnist手寫數字識別的時候，訓練集與測試集就一定是不一樣的。
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("/tmp/data/", one_hot=False)
Extracting /tmp/data/train-images-idx3-ubyte.gz
Extracting /tmp/data/train-labels-idx1-ubyte.gz
Extracting /tmp/data/t10k-images-idx3-ubyte.gz
Extracting /tmp/data/t10k-labels-idx1-ubyte.gz
於是我在自己實現的過程中，將數據集打亂后取10000個作為測試集，不參與訓練，剩余的50000張驗證碼作為訓練集。
訓練過程中發現只有將學習率設置為0.001時，loss才會降下去，太高，loss會卡在0.07；其次，我的訓練精度最多只能到50%左右，但是我用訓練數據來測試保存的模型，精度確實達到了90%，即作者看到的精度。不過這個模型不具有泛化能力，它在沒見過的測試集上只有50%的精確度。
同時這個代碼還有問題：測算精確度時，一張圖中4個驗證碼有兩個錯誤的話，正確率是50%而不是0.當一張圖中4個驗證碼識別有一個錯誤時，該驗證碼識別就應該是失敗的。因此這個精確度實在是有相當大的水分。
於是要考慮解決辦法。首先我嘗試着下調學習率，發現永遠還是到50%就上不去了。
接下來我在原來的3層卷積層上，又加了一層卷積層。然而這並沒有提升多少精度。
隨后我又加入了一層全連接層，期望可以擬合得更好一些，但是這樣讓我陷入了麻煩。
我的loss值卡在了0.07，無論我的學習率是0.1還是0.00001.哪怕迭代一百萬次也是如此。這時候的測試精度只有······3%。
我不知道是什么問題更不知道如何改進。
這更讓我覺得沒有人帶我，多么地難受；同時也更深刻地體驗到理論知識是多么地重要（當然我一直知道）。
我自己的代碼附上，大家可以相互交流。數據可以在文章頂部的鏈接里下載，作者壓縮好的。
以下是訓練腳本：（理論上python3和python2應該都能跑。我是用2寫的）
訓練中我使用了學習率衰減，本來還想用dropout結果發現這個訓練基本不給我過擬合的機會所以訓練加了沒有意義。
from __future__ import print_function, division, absolute_import
import tensorflow as tf
import os
import cv2
import matplotlib.pyplot as plt
import random
import numpy as np
from optparse import OptionParser
path = 'Discuz/' #存放數據的路徑
imgs = os.listdir(path) #以列表形式讀取所有圖片名稱
random.shuffle(imgs) #打亂
max_steps = 1000000 #最大迭代步數
save_path = 'model4cnn-1fcn' #保存模型的路徑，會自動生成
dropout = 1 #沒用到
trainnum = 50000 #定義訓練集和測試集的大小
testnum = 10000
traindatas = imgs[:trainnum] #取出訓練集和測試集及其標簽
trainlabels = list(map(lambda x: x.split('.')[0],traindatas))
testdatas = imgs[trainnum:]
testlabels = list(map(lambda x: x.split('.')[0],testdatas))
#定義取數據集的指針
train_ptr = 0
test_ptr = 0
def next_batch(batch=100, train_flag=True):
global train_ptr
global test_ptr
batch_x = np.zeros([batch,30*100])
batch_y = np.zeros([batch, 4*63])
if train_flag == True:
if batch + train_ptr < trainnum:
trains = traindatas[train_ptr:(train_ptr+batch)]
labels = trainlabels[train_ptr:(train_ptr+batch)]
train_ptr += batch
else:
new_ptr = (train_ptr + batch) % trainnum
trains = traindatas[train_ptr:] + traindatas[:new_ptr]
labels = trainlabels[train_ptr:] + traindatas[:new_ptr]
train_ptr = new_ptr
for index, train in enumerate(trains):
img = np.mean(cv2.imread(path + train), -1)
batch_x[index,:] = img.flatten() /255
for index, label in enumerate(labels):
batch_y[index,:] = text2vec(label)
else:
if batch + test_ptr < testnum:
tests = testdatas[test_ptr:(test_ptr+batch)]
labels = testlabels[test_ptr:(test_ptr+batch)]
test_ptr += batch
else:
new_ptr = (test_ptr + batch) % testnum
tests = testdatas[test_ptr:] + testdatas[:new_ptr]
labels = testlabels[test_ptr:] + testlabels[:new_ptr]
test_ptr = new_ptr
for index, test in enumerate(tests):
img = np.mean(cv2.imread(path + test), -1)
batch_x[index, :] = img.flatten() /255
for index, label in enumerate(labels):
batch_y[index,:] = text2vec(label)
return batch_x, batch_y
def text2vec(text):
if len(text) > 4:
raise ValueError('too long captcha')
vector = np.zeros(4*63)
def char2pos(c):
if c == '_':
k = 62
return k
k = ord(c)-48
if k > 9:
k = ord(c)-55
if k > 35:
k = ord(c) - 61
if k > 61:
raise ValueError('No Map')
return k
for i, c in enumerate(text):
idx = i*63 + char2pos(c)
vector[idx] = 1
return vector
X = tf.placeholder(tf.float32, [None, 30*100])
Y = tf.placeholder(tf.float32, [None,4*63])
_lr = tf.placeholder(tf.float32)
keep_prob = tf.placeholder(tf.float32)
def conv2d(x, W, b, strides=1):
x = tf.nn.conv2d(x, W, strides=[1, strides, strides, 1], padding='SAME')
x = tf.nn.bias_add(x, b)
return tf.nn.relu(x)
def max_pool2d(x, k=2):
x = tf.nn.max_pool(
x, ksize=[
1, k, k, 1], strides=[
1, k, k, 1], padding='SAME')
return x
weights = {
'wc1': tf.Variable(0.01*tf.random_normal([3, 3, 1, 32])),
'wc2': tf.Variable(0.01*tf.random_normal([3, 3, 32, 64])),
'wc3': tf.Variable(0.01*tf.random_normal([3, 3, 64, 64])),
'wc4': tf.Variable(0.01*tf.random_normal([3, 3, 64, 64])),
'wf1': tf.Variable(0.01*tf.random_normal([2 * 7 * 64, 1024])),
'wf2': tf.Variable(0.01*tf.random_normal([1024, 1024])),
'wout': tf.Variable(0.01*tf.random_normal([1024, 4*63]))
}
biases = {
'bc1': tf.Variable(0.1*tf.random_normal([32])),
'bc2': tf.Variable(0.1*tf.random_normal([64])),
'bc3': tf.Variable(0.1*tf.random_normal([64])),
'bc4': tf.Variable(0.1*tf.random_normal([64])),
'bf1': tf.Variable(0.1*tf.random_normal([1024])),
'bf2': tf.Variable(0.1*tf.random_normal([1024])),
'bout': tf.Variable(0.1*tf.random_normal([4*63]))
}
def conv_net(x, weights, biases, dropout):
x = tf.reshape(x, [-1,100,30,1])
conv1 = conv2d(x, weights['wc1'], biases['bc1'], 1)
conv1 = max_pool2d(conv1, 2)
conv2 = conv2d(conv1, weights['wc2'], biases['bc2'], 1)
conv2 = max_pool2d(conv2, 2)
conv3 = conv2d(conv2, weights['wc3'], biases['bc3'], 1)
conv3 = max_pool2d(conv3, 2)
conv4 = conv2d(conv3, weights['wc4'], biases['bc4'], 1)
conv4 = max_pool2d(conv4, 2)
fc1 = tf.reshape(
conv4, shape=[-1, weights['wf1'].get_shape().as_list()[0]])
fc1 = tf.matmul(fc1, weights['wf1'])
fc1 = tf.add(fc1, biases['bf1'])
fc1 = tf.nn.relu(fc1)
out = tf.add(tf.matmul(fc1, weights['wout']), biases['bout'])
return out
output = conv_net(X, weights, biases, keep_prob)
loss_op = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(
logits=output, labels=Y))
optimizer = tf.train.AdamOptimizer(learning_rate=_lr).minimize(loss_op)
y = tf.reshape(output, [-1,4,63])
y_ = tf.reshape(Y, [-1,4,63])
correct_pred = tf.equal(tf.argmax(y, 2), tf.argmax(y_,2))
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
init = tf.global_variables_initializer()
lr = 0.001
saver = tf.train.Saver()
with tf.Session() as sess:
sess.run(init)
for step in range(1,1+max_steps):
batch_x, batch_y = next_batch(100,True)
loss_value,_ = sess.run([loss_op, optimizer],
feed_dict = {X:batch_x, Y:batch_y, keep_prob:dropout,_lr:lr})
if step % 10 == 0:
batch_x_test, batch_y_test = next_batch(100, False)
acc = sess.run(accuracy,
feed_dict={X:batch_x_test, Y:batch_y_test,keep_prob:1})
print('step{}, loss={}, accuracy={}'.format(step,loss_value, acc))
if step % 500 == 0:
random.shuffle(traindatas)
trainlabels = list(map(lambda x: x.split('.')[0],traindatas))
if step % 3000 == 0:
lr *= 0.9
if step % 10000 == 0:
saver.save(sess, save_path + "/model.ckpt-%d" % step)
print('model saved!')
接下來是我寫的一個直觀觀察訓練效果的，新建一個腳本，添加如下代碼，然后運行該腳本，將會隨機展示4張驗證碼和你的預測結果，終端還會顯示本次預測的精確度。
from __future__ import print_function, division, absolute_import
import tensorflow as tf
import os
import cv2
import matplotlib.pyplot as plt
import random
import numpy as np
from datasplit import use
#from optparse import OptionParser
testnumber = 4 #要更改的話需要改畫圖部分的代碼否則會出錯
path = 'Discuz/'
imgs = os.listdir(path)
model_path = 'model4cnn-1fcn/model.ckpt-500000' #讀取你訓練好的模型
testdatas = random.sample(imgs,testnumber)
testlabels = list(map(lambda x: x.split('.')[0],testdatas))
#testnum = len(testdatas)
#test_ptr = 0
X = tf.placeholder(tf.float32, [None, 30*100])
Y = tf.placeholder(tf.float32, [None,4*63])
keep_prob = tf.placeholder(tf.float32)
def text2vec(text):
if len(text) > 4:
raise ValueError('too long captcha')
vector = np.zeros(4*63)
def char2pos(c):
if c == '_':
k = 62
return k
k = ord(c)-48
if k > 9:
k = ord(c)-55
if k > 35:
k = ord(c) - 61
if k > 61:
raise ValueError('No Map')
return k
for i, c in enumerate(text):
idx = i*63 + char2pos(c)
vector[idx] = 1
return vector
def vec2text(vec):
char_pos = vec.nonzero()[0]
text = []
for i, c in enumerate(char_pos):
char_at_pos = i #c/63
char_idx = c % 63
if char_idx < 10:
char_code = char_idx + ord('0')
elif char_idx < 36:
char_code = char_idx - 10 + ord('A')
elif char_idx < 62:
char_code = char_idx - 36 + ord('a')
elif char_idx == 62:
char_code = ord('_')
else:
raise ValueError('error')
text.append(chr(char_code))
return "".join(text)
batch_x = np.zeros([testnumber,30*100])
batch_y = np.zeros([testnumber, 4*63])
for index, test in enumerate(testdatas):
img = np.mean(cv2.imread(path + test), -1)
batch_x[index, :] = img.flatten() /255
for index, label in enumerate(testlabels):
batch_y[index, :] = text2vec(label)
def conv2d(x, W, b, strides=1):
x = tf.nn.conv2d(x, W, strides=[1, strides, strides, 1], padding='SAME')
x = tf.nn.bias_add(x, b)
return tf.nn.relu(x)
def max_pool2d(x, k=2):
x = tf.nn.max_pool(
x, ksize=[
1, k, k, 1], strides=[
1, k, k, 1], padding='SAME')
return x
weights = {
'wc1': tf.Variable(0.01*tf.random_normal([3, 3, 1, 32])),
'wc2': tf.Variable(0.01*tf.random_normal([3, 3, 32, 64])),
'wc3': tf.Variable(0.01*tf.random_normal([3, 3, 64, 64])),
'wc4': tf.Variable(0.01*tf.random_normal([3, 3, 64, 64])),
'wf1': tf.Variable(0.01*tf.random_normal([2 * 7 * 64, 1024])),
'wf2': tf.Variable(0.01*tf.random_normal([1024, 1024])),
'wout': tf.Variable(0.01*tf.random_normal([1024, 4*63]))
}
biases = {
'bc1': tf.Variable(0.1*tf.random_normal([32])),
'bc2': tf.Variable(0.1*tf.random_normal([64])),
'bc3': tf.Variable(0.1*tf.random_normal([64])),
'bc4': tf.Variable(0.1*tf.random_normal([64])),
'bf1': tf.Variable(0.1*tf.random_normal([1024])),
'bf2': tf.Variable(0.1*tf.random_normal([1024])),
'bout': tf.Variable(0.1*tf.random_normal([4*63]))
}
def conv_net(x, weights, biases, dropout):
x = tf.reshape(x, [-1,100,30,1])
conv1 = conv2d(x, weights['wc1'], biases['bc1'], 1)
conv1 = max_pool2d(conv1, 2)
conv2 = conv2d(conv1, weights['wc2'], biases['bc2'], 1)
conv2 = max_pool2d(conv2, 2)
conv3 = conv2d(conv2, weights['wc3'], biases['bc3'], 1)
conv3 = max_pool2d(conv3, 2)
conv4 = conv2d(conv3, weights['wc4'], biases['bc4'], 1)
conv4 = max_pool2d(conv4, 2)
fc1 = tf.reshape(
conv4, shape=[-1, weights['wf1'].get_shape().as_list()[0]])
fc1 = tf.matmul(fc1, weights['wf1'])
fc1 = tf.add(fc1, biases['bf1'])
fc1 = tf.nn.relu(fc1)
out = tf.add(tf.matmul(fc1, weights['wout']), biases['bout'])
return out
output = conv_net(X, weights, biases, keep_prob)
y = tf.reshape(output, [-1,4,63])
y_ = tf.reshape(Y, [-1,4,63])
predict = tf.argmax(y,2)
correct_pred = tf.equal(predict, tf.argmax(y_,2))
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
saver = tf.train.Saver()
with tf.Session() as sess:
saver.restore(sess, model_path)
pred, acc = sess.run([predict,accuracy], feed_dict ={ X:batch_x, Y:batch_y,keep_prob:1})
print('accuracy={}'.format(acc))
for i in range(1,testnumber+1):
plt.subplot(2,2,i)
img = cv2.imread(path+testdatas[i-1])
plt.imshow(img)
plt.title('number%d' %i)
plt.xticks([])
plt.yticks([])
vect = np.zeros([4*63])
#print(pred[i-1])
for ind,j in enumerate(pred[i-1]):
vect[ind*63+j] = 1
xlabel = 'True label:{};Pred label:{}'.format(testlabels[i-1], vec2text(vect))
plt.xlabel(xlabel)
plt.show()
有任何問題歡迎討論。