用tensorflow迁移学习猫狗分类
2024-10-11 21:46:28
笔者这几天在跟着莫烦学习TensorFlow，正好到迁移学习（至于什么是迁移学习，看这篇），莫烦老师做的是预测猫和老虎尺寸大小的学习。作为一个有为的学生，笔者当然不能再预测猫啊狗啊的大小啦，正好之前正好有做过猫狗大战数据集的图像分类，做好的数据都还在，二话不说，开撸。
既然是VGG16模型，当然首先上模型代码了：
def conv_layers_simple_api(net_in):
with tf.name_scope('preprocess'):
# Notice that we include a preprocessing layer that takes the RGB image
# with pixels values in the range of 0-255 and subtracts the mean image
# values (calculated over the entire ImageNet training set).
mean = tf.constant([123.68, 116.779, 103.939], dtype=tf.float32, shape=[1, 1, 1, 3], name='img_mean')
net_in.outputs = net_in.outputs - mean # conv1
network = Conv2d(net_in, n_filter=64, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv1_1')
network = Conv2d(network, n_filter=64, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv1_2')
network = MaxPool2d(network, filter_size=(2, 2), strides=(2, 2), padding='SAME', name='pool1') # conv2
network = Conv2d(network, n_filter=128, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv2_1')
network = Conv2d(network, n_filter=128, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv2_2')
network = MaxPool2d(network, filter_size=(2, 2), strides=(2, 2), padding='SAME', name='pool2') # conv3
network = Conv2d(network, n_filter=256, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv3_1')
network = Conv2d(network, n_filter=256, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv3_2')
network = Conv2d(network, n_filter=256, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv3_3')
network = MaxPool2d(network, filter_size=(2, 2), strides=(2, 2), padding='SAME', name='pool3') # conv4
network = Conv2d(network, n_filter=512, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv4_1')
network = Conv2d(network, n_filter=512, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv4_2')
network = Conv2d(network, n_filter=512, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv4_3')
network = MaxPool2d(network, filter_size=(2, 2), strides=(2, 2), padding='SAME', name='pool4') # conv5
network = Conv2d(network, n_filter=512, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv5_1')
network = Conv2d(network, n_filter=512, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv5_2')
network = Conv2d(network, n_filter=512, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv5_3')
network = MaxPool2d(network, filter_size=(2, 2), strides=(2, 2), padding='SAME', name='pool5')
return network``
def conv_layers_simple_api(net_in):
with tf.name_scope('preprocess'):
# Notice that we include a preprocessing layer that takes the RGB image
# with pixels values in the range of 0-255 and subtracts the mean image
# values (calculated over the entire ImageNet training set).
mean = tf.constant([123.68, 116.779, 103.939], dtype=tf.float32, shape=[1, 1, 1, 3], name='img_mean')
net_in.outputs = net_in.outputs - mean # conv1
network = Conv2d(net_in, n_filter=64, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv1_1')
network = Conv2d(network, n_filter=64, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv1_2')
network = MaxPool2d(network, filter_size=(2, 2), strides=(2, 2), padding='SAME', name='pool1') # conv2
network = Conv2d(network, n_filter=128, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv2_1')
network = Conv2d(network, n_filter=128, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv2_2')
network = MaxPool2d(network, filter_size=(2, 2), strides=(2, 2), padding='SAME', name='pool2') # conv3
network = Conv2d(network, n_filter=256, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv3_1')
network = Conv2d(network, n_filter=256, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv3_2')
network = Conv2d(network, n_filter=256, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv3_3')
network = MaxPool2d(network, filter_size=(2, 2), strides=(2, 2), padding='SAME', name='pool3') # conv4
network = Conv2d(network, n_filter=512, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv4_1')
network = Conv2d(network, n_filter=512, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv4_2')
network = Conv2d(network, n_filter=512, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv4_3')
network = MaxPool2d(network, filter_size=(2, 2), strides=(2, 2), padding='SAME', name='pool4') # conv5
network = Conv2d(network, n_filter=512, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv5_1')
network = Conv2d(network, n_filter=512, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv5_2')
network = Conv2d(network, n_filter=512, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv5_3')
network = MaxPool2d(network, filter_size=(2, 2), strides=(2, 2), padding='SAME', name='pool5')
return network``
def conv_layers_simple_api(net_in):
with tf.name_scope('preprocess'):
# Notice that we include a preprocessing layer that takes the RGB image
# with pixels values in the range of 0-255 and subtracts the mean image
# values (calculated over the entire ImageNet training set).
mean = tf.constant([123.68, 116.779, 103.939], dtype=tf.float32, shape=[1, 1, 1, 3], name='img_mean')
net_in.outputs = net_in.outputs - mean # conv1
network = Conv2d(net_in, n_filter=64, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv1_1')
network = Conv2d(network, n_filter=64, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv1_2')
network = MaxPool2d(network, filter_size=(2, 2), strides=(2, 2), padding='SAME', name='pool1') # conv2
network = Conv2d(network, n_filter=128, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv2_1')
network = Conv2d(network, n_filter=128, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv2_2')
network = MaxPool2d(network, filter_size=(2, 2), strides=(2, 2), padding='SAME', name='pool2') # conv3
network = Conv2d(network, n_filter=256, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv3_1')
network = Conv2d(network, n_filter=256, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv3_2')
network = Conv2d(network, n_filter=256, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv3_3')
network = MaxPool2d(network, filter_size=(2, 2), strides=(2, 2), padding='SAME', name='pool3') # conv4
network = Conv2d(network, n_filter=512, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv4_1')
network = Conv2d(network, n_filter=512, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv4_2')
network = Conv2d(network, n_filter=512, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv4_3')
network = MaxPool2d(network, filter_size=(2, 2), strides=(2, 2), padding='SAME', name='pool4') # conv5
network = Conv2d(network, n_filter=512, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv5_1')
network = Conv2d(network, n_filter=512, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv5_2')
network = Conv2d(network, n_filter=512, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME',
name='conv5_3')
network = MaxPool2d(network, filter_size=(2, 2), strides=(2, 2), padding='SAME', name='pool5')
return network
笔者偷懒直接用的是TensorLayer库中的Vgg16模型，至于什么是tensorlayer请移步这里
按照莫烦老师的教程，改写最后的全连接层做二分类学习：
def fc_layers(net):
# 全连接层前的预处理
network = FlattenLayer(net, name='flatten')
# tf.layers.dense(self.flatten, 256, tf.nn.relu, name='fc6')
network = DenseLayer(network, n_units=256, act=tf.nn.relu, name='fc1_relu')
# network = DenseLayer(network, n_units=4096, act=tf.nn.relu, name='fc2_relu')
# self.out = tf.layers.dense(self.fc6, 1, name='out')
network = DenseLayer(network, n_units=2, act=tf.identity, name='fc3_relu')
return network
定义输入输出以及损失函数已及学习步骤：
# 输入
x = tf.placeholder(tf.float32, [None, 224, 224, 3])
# 输出
y_ = tf.placeholder(tf.int32, shape=[None, ], name='y_')
net_in = InputLayer(x, name='input')
# net_cnn = conv_layers(net_in) # professional CNN APIs
net_cnn = conv_layers_simple_api(net_in) # simplified CNN APIs
network = fc_layers(net_cnn)
y = network.outputs
# probs = tf.nn.softmax(y)
y_op = tf.argmax(tf.nn.softmax(y), 1)
cost = tl.cost.cross_entropy(y, y_, name='cost')
correct_prediction = tf.equal(tf.cast(tf.argmax(y, 1), tf.float32), tf.cast(y_, tf.float32))
acc = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
# 定义 optimizer
train_params = network.all_params[26:]
# print(train_params)
global_step = tf.Variable(0)
# --------------学习速率的设置（学习速率呈指数下降）--------------------- #将 global_step/decay_steps 强制转换为整数
# learning_rate = tf.train.exponential_decay(1e-2, global_step, decay_steps=1000, decay_rate=0.98, staircase=True)
train_op = tf.train.AdamOptimizer(learning_rate=0.0001, beta1=0.9, beta2=0.999,
epsilon=1e-08, use_locking=False).minimize(cost, var_list=train_params)
读取数据读取训练、验证数据，加载模型参数：
img, label = read_and_decode("F:\\001-python\\train.tfrecords")
img_v, label_v = read_and_decode("F:\\001-python\\val.tfrecords")
# 使用shuffle_batch可以随机打乱输入
X_train, y_train = tf.train.shuffle_batch([img, label],
batch_size=30, capacity=400,
min_after_dequeue=300)
X_Val, y_val = tf.train.shuffle_batch([img_v, label_v],
batch_size=30, capacity=400,
min_after_dequeue=300)
tl.layers.initialize_global_variables(sess)
network.print_params()
network.print_layers()
npz = np.load('vgg16_weights.npz')
params = []
for val in sorted(npz.items())[0:25]:
# print(" Loading %s" % str(val[1].shape))
params.append(val[1])
加载预训练的参数
tl.files.assign_params(sess, params, network)
加载好之后，开始训练，200个epoch：
for epoch in range(n_epoch):
start_time = time.time()
val, l = sess.run([X_train, y_train])
for X_train_a, y_train_a in tl.iterate.minibatches(val, l, batch_size, shuffle=True):
sess.run(train_op, feed_dict={x: X_train_a, y_: y_train_a})
if epoch + 1 == 1 or (epoch + 1) % 5 == 0:
print("Epoch %d of %d took %fs" % (epoch + 1, n_epoch, time.time() - start_time))
train_loss, train_acc, n_batch = 0, 0, 0
for X_train_a, y_train_a in tl.iterate.minibatches(val, l, batch_size, shuffle=True):
err, ac = sess.run([cost, acc], feed_dict={x: X_train_a, y_: y_train_a})
train_loss += err
train_acc += ac
n_batch += 1
print(" train loss: %f" % (train_loss / n_batch))
print(" train acc: %f" % (train_acc / n_batch))
保存训练的参数：
tl.files.save_npz(network.all_params, name='model.npz', sess=sess)
下面就是开始训练啦，笔者很高兴的拿着自己的笔记本显卡呼呼的跑了一遍:
~~~~~~~~~~~~~~~~~~~~~~~~下面是漫长的等待
.......
[TL] Epoch 138 of 150 took 0.999402s
[TL] val loss: 0.687194
[TL] val acc: 0.562500
[TL] Epoch 140 of 150 took 3.782207s
[TL] val loss: 0.619966
[TL] val acc: 0.750000
[TL] Epoch 142 of 150 took 0.983802s
[TL] val loss: 0.685686
[TL] val acc: 0.562500
[TL] Epoch 144 of 150 took 0.986604s
[TL] val loss: 0.661224
[TL] val acc: 0.687500
[TL] Epoch 146 of 150 took 1.022403s
[TL] val loss: 0.675885
[TL] val acc: 0.687500
[TL] Epoch 148 of 150 took 0.991802s
[TL] val loss: 0.682124
[TL] val acc: 0.625000
[TL] Epoch 150 of 150 took 3.487811s
[TL] val loss: 0.674932
[TL] val acc: 0.687500
[TL] Total training time: 319.859640s
[TL] [*] model.npz saved
额~~~~~~~~~~~~~~~~~
0.68的正确率，群里一位朋友看了之后说：跟猜差不多了(一脸黑线)。问题出哪儿呢？难道是笔者训练的次数不够多？莫烦老师可是100次就能出很好的结果啊
不管怎么样，要试试，笔者于是加载刚刚保存的model.npz参数继续跑100个epoch
~~~~~~~~~~~~~~~~~~~~~~~~又是漫长的等待
[TL] Epoch 1 of 100 took 8.477617s
[TL] val loss: 0.685957
[TL] val acc: 0.562500
[TL] Epoch 2 of 100 took 0.999402s
[TL] val loss: 0.661529
[TL] val acc: 0.625000
......
[TL] Epoch 94 of 100 took 0.992208s
[TL] val loss: 0.708815
[TL] val acc: 0.562500
[TL] Epoch 96 of 100 took 0.998406s
[TL] val loss: 0.710636
[TL] val acc: 0.562500
[TL] Epoch 98 of 100 took 0.992807s
[TL] val loss: 0.621505
[TL] val acc: 0.687500
[TL] Epoch 100 of 100 took 0.986405s
[TL] val loss: 0.670647
[TL] val acc: 0.625000
[TL] Total training time: 156.734633s
[TL] [*] model.npz saved
坑爹啊这是，还不如之前的结果。
笔者陷入深深的沉思中，难道是改了全连接层导致的？于是笔者又把之前去掉的全连接层加上：
def fc_layers(net):
# 全连接层前的预处理
network = FlattenLayer(net, name='flatten')
# tf.layers.dense(self.flatten, 256, tf.nn.relu, name='fc6')
network = DenseLayer(network, n_units=256, act=tf.nn.relu, name='fc1_relu')
network = DenseLayer(network, n_units=256, act=tf.nn.relu, name='fc2_relu')
# self.out = tf.layers.dense(self.fc6, 1, name='out')
network = DenseLayer(network, n_units=2, act=tf.identity, name='fc3_relu')
return network
接着训练
~~~~~~~~~~~~~~~~~~~~~~~~下面又是漫长的等待
[TL] Epoch 1 of 100 took 8.477229s
[TL] val loss: 2.370650
[TL] val acc: 0.562500
...
[TL] Epoch 100 of 100 took 1.016002s
[TL] val loss: 0.762171
[TL] val acc: 0.437500
[TL] Total training time: 156.836465s
[TL] [*] model.npz saved
还是一样，笔者已崩溃了，一定是哪儿不对啊啊啊....于是笔者去翻莫烦老师的代码，一点点对下来，每一层参数肯定不会有错，那就是在训练设置的参数有问题。
self.train_op = tf.train.RMSPropOptimizer(0.001).minimize(self.loss) #莫烦的代码
train_op = tf.train.AdamOptimizer(learning_rate=0.0001, beta1=0.9, beta2=0.999,
epsilon=1e-08, use_locking=False).minimize(cost, var_list=train_params)#笔者的
看到train_params难道是这个train_params？笔者只优化了最后的全连接层参数而莫烦老师优化的是全部参数
已经深夜了，笔者表示即使不睡觉也要跑一遍试试，于是改成
# 定义 optimizer
train_params = network.all_params
~~~~~~~~~~~~~~~~~~~~~~~~于是又是是漫长的等待 [TL] Epoch 1 of 100 took 20.286640s
[TL] val loss: 11.938850
[TL] val acc: 0.312500
[TL] Epoch 2 of 100 took 3.091806s
[TL] val loss: 2.890055
[TL] val acc: 0.625000
[TL] Epoch 4 of 100 took 3.074205s
[TL] val loss: 24.055895
[TL] val acc: 0.687500
[TL] ....
[TL] val loss: 0.699907
[TL] val acc: 0.500000
[TL] Epoch 98 of 100 took 3.089206s
[TL] val loss: 0.683627
[TL] val acc: 0.562500
[TL] Epoch 100 of 100 took 3.091806s
[TL] val loss: 0.708496
[TL] val acc: 0.562500
[TL] Total training time: 375.727307s
[TL] [*] model.npz saved
效果变得更差了....
排除参数的问题，已经深夜1点了，明天还要上班，不得不睡啦。
继续崩溃第三天~~~
第四天~~~
第五天，今天供应商过来公司调试机器，正好是一个学图像处理的小伙子，我提到这个说：我为啥训练了这么多代为啥还是像猜一样的概率....？小伙儿说：莫不是过拟合了吧？我说:不可能啊现成的数据现成的模型和参数，不应该的啊！
不过我还是得检查一下数据处理的代码
# 生成是数据文件
def create_record(filelist):
random.shuffle(filelist)
i = 0
writer = tf.python_io.TFRecordWriter(recordpath)
for file in filelist:
name = file.split(sep='.')
lable_val = 0
if name[0] == 'cat':
lable_val = 0
else:
lable_val = 1
img_path = file_dir + file
img = Image.open(img_path)
img = img.resize((240, 240))
img_raw = img.tobytes() # 将图片转化为原生bytes
example = tf.train.Example(features=tf.train.Features(feature={
"label": tf.train.Feature(int64_list=tf.train.Int64List(value=[lable_val])),
'img_raw': tf.train.Feature(bytes_list=tf.train.BytesList(value=[img_raw]))
})) #example对象对label和image进行封装
writer.write(example.SerializeToString())
i=i+1
print(name[1])
print(lable_val)
print(i)
writer.close()
# 用队列形式读取文件
def read_and_decode(filename):
# 根据文件名生成一个队列
filename_queue = tf.train.string_input_producer([filename])
reader = tf.TFRecordReader()
_, serialized_example = reader.read(filename_queue) # 返回文件名和文件
features = tf.parse_single_example(serialized_example,
features={
'label': tf.FixedLenFeature([], tf.int64),
'img_raw': tf.FixedLenFeature([], tf.string),
})
img = tf.decode_raw(features['img_raw'], tf.uint8)
img = tf.reshape(img, [224, 224, 3])
img = tf.cast(img, tf.float32) * (1. / 255) - 0.5
label = tf.cast(features['label'], tf.int32)
return img, label
img = tf.cast(img, tf.float32) * (1. / 255) - 0.5 难道是这一步处理多余？注销之后，训练模型
Epoch 85 of 200 took 1.234071s
train loss: 14.689816
train acc: 0.900000
[TL] [*] model3.npz saved
Epoch 90 of 200 took 1.241071s
train loss: 17.104382
train acc: 0.800000
[TL] [*] model3.npz saved
Epoch 95 of 200 took 1.236071s
train loss: 11.190630
train acc: 0.850000
[TL] [*] model3.npz saved
Epoch 100 of 200 took 1.238071s
train loss: 0.000000
train acc: 1.000000
[TL] [*] model3.npz saved
Epoch 105 of 200 took 1.236071s
train loss: 7.622324
train acc: 0.900000
[TL] [*] model3.npz saved
Epoch 110 of 200 took 1.234071s
train loss: 2.164670
train acc: 0.950000
[TL] [*] model3.npz saved
Epoch 115 of 200 took 1.237071s
train loss: 0.000000
train acc: 1.000000
[TL] [*] model3.npz saved
准确度1，停停停...不用跑完了，Perfect!
原来如此，必须要真实的像素值.......心好累......，笔者已经不记得哪儿抄来的这一行了。
嗯，VGG16模型的迁移学习到此结束，代码见github
最新文章
- 固态硬盘与机械硬盘 SQL Server 单表插入性能对比测试
- CCF 201604-4 游戏
- 来聊聊apply和call
- [转载]寻找两个有序数组中的第K个数或者中位数
- 指向函数的指针与iOS-Block相关知识
- MySQL Workbench类型之MySQL客户端工具的下载、安装和使用
- 设置控件全局显示样式 appearance
- 自制stm32板子无法烧录程序的问题
- ajaxFileUpload上传带参数,返回值改成json格式
- JS案例四：表格的隔行换色以及高亮显示
- Unity添加多个可视镜头Preview功能（一）
- Java高并发秒杀API之Service层
- C和C指针小记(十五)-结构和联合
- 黄聪：通过 itms:services://? 在线安装ipa ，跨过appstore
- python3学习笔记二（注释、缩进）
- MongoDB阅读精要
- e770. 确定按钮租中已选的单选按钮
- node：爬虫爬取网页图片
- Mac 下,Redis(集群)的安装和配置
- OVN实战---《OVN and Containers》翻译