【TensorFlow】微调（fine tuning）

ゝ一世哀愁。 2024-04-18 14:58 70阅读 0赞

> ## 一、使用slim及nets库，调用经典模型结构 ##
> 
> （调用模型，自己训练）
> 
>     import tensorflow as tf
>     import tensorflow.contrib.slim as slim
>     import tensorflow.contrib.slim.nets as nets
>     
>     images = tf.placeholder(tf.float32, [None, 224, 224, 3])
>     net = nets.vgg.vgg_16(images)
> 
> 下图是nets.py文件，在上面那段代码里ctrl+右键 点一下nets就能进去。
> 
> 我们可以看到这个文件里有很多经典网络模型，通过nets.就能直接调用。
> 
> ![vici\_\_][vici]
> 
> 我们打开vgg.py文件，里面会有很多函数，比如vgg\_16()、vgg\_19()，函数里面写了这个模型的结构。
> 
> ![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L1ZpY2lfXw_size_16_color_FFFFFF_t_70][]
> 
> 也就是说，这里只是把一些模型封装起来，当我们想用这个网络模型结构时，不再需要一层层地写，直接调用就行了。
> 
> 如果我们自己去把整个神经网络从头到尾重新训练一遍的话，可能会耗费大量的时间，而且训练结果并不一定很好。
> 
> 那么我们可以用别人训练好的参数，直接把别人的参数导入到模型中。

> ### 二、将参数导入到网络模型中 ###
> 
> **1、从.ckpt文件导出参数：**
> 
> **详情可以看下这篇博客[【TensorFlow】使用slim从ckpt里导出指定层的参数][TensorFlow_slim_ckpt]**
> 
>     import tensorflow as tf
>     import tensorflow.contrib.slim as slim
>     import tensorflow.contrib.slim.nets as nets
>      
>     s = tf.Session()
>      
>     images = tf.placeholder(tf.float32, [None, 224, 224, 3])
>     net = nets.vgg.vgg_16(images, 200)
>     
>     # ！！！
>     variables_to_restore = slim.get_variables_to_restore()
>     init_assign_op, init_feed_dict = slim.assign_from_checkpoint('./vgg16.ckpt', variables_to_restore)
>     s.run(init_assign_op, init_feed_dict)
> 
> **2、从.pb文件导出网络模型和参数：**
> 
>     # .pb 文件中既存储了inception的网络结构，也存储了对应数据
>     
>     from tensorflow.python.platform import gfile
>      
>     sess = tf.Session()
>     
>     with gfile.FastGFile(pb_file_path+'model.pb', 'rb') as f:
>         graph_def = tf.GraphDef()
>         graph_def.ParseFromString(f.read())
>         sess.graph.as_default()
>         tf.import_graph_def(graph_def, name='') # 导入计算图
>      
>     # 需要有一个初始化的过程    
>     sess.run(tf.global_variables_initializer())
>      
>     # 需要先复原变量
>     print(sess.run('b:0'))
>     # 1
>      
>     # 输入
>     input_x = sess.graph.get_tensor_by_name('x:0')
>     input_y = sess.graph.get_tensor_by_name('y:0')
>      
>     op = sess.graph.get_tensor_by_name('op_to_store:0')
>      
>     ret = sess.run(op,  feed_dict={input_x: 5, input_y: 5})
>     print(ret)
>     # 输出 26
> 
> **附参考博客：[tensorflow保存数据为.pb格式和加载pb文件][tensorflow_.pb_pb]**

> ### 三、fine tuning对模型做微调 ###
> 
> 那么如果我想对网络结构做一下微调，把最后一层的输出改一下呢？  
> 比如VGG网络，输出是个1000分类的，当我们想识别猫狗两类的时候，就需要稍作改动了。
> 
>     def vgg_16(inputs):
>         with slim.arg_scope([slim.conv2d, slim.fully_connected],  
>                             activation_fn=tf.nn.relu,  
>                             weights_initializer=tf.truncated_normal_initializer(0.0, 0.01),  
>                             weights_regularizer=slim.l2_regularizer(0.0005),
>                             scope='vgg_16'):
>             net = slim.conv2d(inputs, 64, [3, 3], scope='vgg_16/conv1/conv1_1')
>             net = slim.conv2d(net, 64, [3, 3], scope='vgg_16/conv1/conv1_2')
>             net = slim.max_pool2d(net, [2, 2], scope='pool1')
>     
>             net = slim.conv2d(net, 128, [3, 3], scope='vgg_16/conv2/conv2_1')
>             net = slim.conv2d(net, 128, [3, 3], scope='vgg_16/conv2/conv2_2')
>             net = slim.max_pool2d(net, [2, 2], scope='pool2')  
>     
>             net = slim.conv2d(net, 256, [3, 3], scope='vgg_16/conv3/conv3_1')
>             net = slim.conv2d(net, 256, [3, 3], scope='vgg_16/conv3/conv3_2')
>             net = slim.conv2d(net, 256, [3, 3], scope='vgg_16/conv3/conv3_3')
>             net = slim.max_pool2d(net, [2, 2], scope='pool3')  
>     
>             net = slim.conv2d(net, 512, [3, 3], scope='vgg_16/conv4/conv4_1')
>             net = slim.conv2d(net, 512, [3, 3], scope='vgg_16/conv4/conv4_2')
>             net = slim.conv2d(net, 512, [3, 3], scope='vgg_16/conv4/conv4_3')
>             net = slim.max_pool2d(net, [2, 2], scope='pool4')  
>     
>             net = slim.conv2d(net, 512, [3, 3], scope='vgg_16/conv5/conv5_1')
>             net = slim.conv2d(net, 512, [3, 3], scope='vgg_16/conv5/conv5_2')
>             net = slim.conv2d(net, 512, [3, 3], scope='vgg_16/conv5/conv5_3')
>             net = slim.max_pool2d(net, [2, 2], scope='pool5')  
>     
>             net = slim.flatten(net,scope='flat5')
>             net = slim.fully_connected(net, 4096, scope='vgg_16/fc6')  
>             net = slim.dropout(net, 0.5, scope='dropout6')  
>             net = slim.fully_connected(net, 4096, scope='vgg_16/fc7')  
>             net = slim.dropout(net, 0.5, scope='dropout7')  
>             # net = slim.fully_connected(net, 1000, activation_fn=None, scope='vgg_16/fc8') 
>             net = slim.fully_connected(net, 2, activation_fn=None, scope='vgg_16/fc8') 
>         return net
>     
>     with tf.Session() as sess:
>         net = vgg_16(inputs=x)
>         variables_to_restore = slim.get_variables_to_restore(exclude=['vgg_16/fc6', 'vgg_16/fc7', 'vgg_16/fc8'])
>         init_assign_op, init_feed_dict = slim.assign_from_checkpoint('D:/Python_code/007_fine_tuning/VGG16/vgg_16_2016_08_28/vgg_16.ckpt', variables_to_restore)
>         sess.run(init_assign_op, init_feed_dict)
> 
> 首先，我们可以自定义网络结构，只把最后一层的输出维数1000改为2，其它层不变化；
> 
> 然后我们需要自己训练一下输出层，我们在从ckpt文件中导出参数时，用**exclude=\["vgg\_16/fc8"\]**把fc8这层除外。
> 
> 其余代码都一样。

[vici]: https://image.dandelioncloud.cn/pgy_files/images/2024/04/18/e56c386fb5be4579a6a11b24fee77d84.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L1ZpY2lfXw_size_16_color_FFFFFF_t_70]: https://image.dandelioncloud.cn/pgy_files/images/2024/04/18/be9e54de9f654be6aaa2c6c7bf225ff6.png
[TensorFlow_slim_ckpt]: https://blog.csdn.net/Vici__/article/details/100626467
[tensorflow_.pb_pb]: https://blog.csdn.net/fu6543210/article/details/80343345