Continuous training on tensorflow breakpoint (principle + code explanation)

One 、 What is breakpoint continuation training

Breakpoint continuation training means that the training is interrupted before the completion of the training for some reasons , The next training can continue on the basis of the previous training . This approach is very friendly for models that require long-term training

Two 、 Model file parsing

checkpoint The file will record the saving information , It allows you to locate the latest saved model ;
.meta The file saves the current NN Network structure ：tf.train.import_meta_graph(‘MODEL_NAME.ckpt-1174.meta’)
.data The file saves the current parameter name and value , Network weight 、 bias 、 Operation etc.
.index The file holds secondary index information , Is an immutable string table
As for the number after the file name 1174 It represents different batches of model training , We usually only need the latest one ;

3、 ... and 、 How to realize breakpoint continuation training

Two prerequisites are met ：

（1）： The snapshot in the model trainer is saved locally （ Breakpoint data saving ）
（2）： You can restore the field environment trained by the model by reading the snapshot （ Breakpoint data recovery ）
among , Both of these operations require tensorflow Medium train.Saver class . Address of official documents

1. establish tensorflow.train.Saver class

saver = tf.train.Saver(max_to_keep=1) # It is allowed to save the latest number of models for training , The default value is 5, If only the latest model is saved, the assignment is 1

2. Use Saver Object's save Method to save the model

saver.save(sess,os.path.join(MODEL_SVAE_PATH , MODEL_NAME ),global_step=epoch)
#sess For the session that needs to be saved ,MODEL_SVAE_PATH Save path for model ,MODEL_NAME Name the model , global_step Model training times

3. Recovery of breakpoint data

3.1 Only load the model, not the graph

saver.restore(sess, ckpt.model_checkpoint_path) # Resume the current session , take ckpt The value in is assigned to w and b
This method is generally used in the breakpoint continuation training , When the network model structure is not very complex , Rebuilding the session graph is also only at the millisecond level . Besides , The graph structure is loaded only once , Because during the whole training process , The network structure will not change .

3.2 The graph structure and parameters are loaded

saver = tf.train.import_meta_graph(ckpt+".meta") # Load graph structure , That is, the structure of neural network
saver.restore(sess, ckpt.model_checkpoint_path) # Resume the current session , take ckpt The value in is assigned to w and b

Four 、 Code details

In this part , I will attach a complete code for the first contact breakpoint continuation , People who want to use it in the program . The code is divided into four parts ： Load what you need package, Load data set , Build a network model , Training models . The first three parts of the code have little relevance to the breakpoint continuation training , Students who already have some basic knowledge can directly look at the last part . Complete code training requires tensorflow>=2.0 and cifar10 Data sets .

reminder ：

If you only save and load model parameters , Each retraining of the model will indeed load the last model parameters , The training effect will continue to be optimized on the basis of the previous one , But the training times of the model start again , under these circumstances , The actual model training times cannot be determined . But in some cases , We need an exact model training parameter to compare with other benchmark models . So we need to keep the last training epoch, Subtract... When starting a new round of training , Only in this way can the training times of the model be consistent .

4.1 Import required package

import tensorflow.compat.v1 as tf  
tf.disable_v2_behavior()  #tf2.0 in placehold The function is obsolete , In order to retain tf1.0 Chinese grammar 
import os                 # It is convenient to create models and save them checkpoint Folder  
import pickle
import numpy as np
import os
import re               # To get the number in the string , Introducing regular expressions 

MODEL_SAVE_PATH = './ckpt/'  #saver.save Will be automatically saved in this folder checkpoint file 
MODEL_NAME = 'vgg model'
batch_size = 20
train_steps = 10000
test_steps = 100

CIFAR_DIR = "D:\Dataset\cifar-10-python\cifar-10-batches-py"  # Change to yourself cifar10 Directory of datasets 
print(os.listdir(CIFAR_DIR))

4.2 Load data （ This part has nothing to do with breakpoint continuation training ）

def load_data(filename):
    """read data from data file."""
    with open(filename, 'rb') as f:
        data = pickle.load(f, encoding='bytes')
        return data[b'data'], data[b'labels']

# tensorflow.Dataset.
class CifarData:
    def __init__(self, filenames, need_shuffle):
        all_data = []
        all_labels = []
        for filename in filenames:
            data, labels = load_data(filename)
            all_data.append(data)
            all_labels.append(labels)
        self._data = np.vstack(all_data)
        self._data = self._data / 127.5 - 1
        self._labels = np.hstack(all_labels)
        print(self._data.shape)
        print(self._labels.shape)
        
        self._num_examples = self._data.shape[0]
        self._need_shuffle = need_shuffle
        self._indicator = 0
        if self._need_shuffle:
            self._shuffle_data()
            
    def _shuffle_data(self):
        # [0,1,2,3,4,5] -> [5,3,2,4,0,1]
        p = np.random.permutation(self._num_examples)
        self._data = self._data[p]
        self._labels = self._labels[p]
    
    def next_batch(self, batch_size):
        """return batch_size examples as a batch."""
        end_indicator = self._indicator + batch_size
        if end_indicator > self._num_examples:
            if self._need_shuffle:
                self._shuffle_data()
                self._indicator = 0
                end_indicator = batch_size
            else:
                raise Exception("have no more examples")
        if end_indicator > self._num_examples:
            raise Exception("batch size is larger than all examples")
        batch_data = self._data[self._indicator: end_indicator]
        batch_labels = self._labels[self._indicator: end_indicator]
        self._indicator = end_indicator
        return batch_data, batch_labels

train_filenames = [os.path.join(CIFAR_DIR, 'data_batch_%d' % i) for i in range(1, 6)]
test_filenames = [os.path.join(CIFAR_DIR, 'test_batch')]

train_data = CifarData(train_filenames, True)
test_data = CifarData(test_filenames, False)

4.3 Build a network model （ It has nothing to do with continuing training ）

x = tf.placeholder(tf.float32, [None, 3072])
y = tf.placeholder(tf.int64, [None])
# [None], eg: [0,5,6,3]
x_image = tf.reshape(x, [-1, 3, 32, 32])
# 32*32
x_image = tf.transpose(x_image, perm=[0, 2, 3, 1])

# conv1:  Neuron diagram , feature_map,  Output image 
conv1_1 = tf.layers.conv2d(x_image,
                           32, # output channel number
                           (3,3), # kernel size
                           padding = 'same',
                           activation = tf.nn.relu,
                           name = 'conv1_1')
conv1_2 = tf.layers.conv2d(conv1_1,
                           32, # output channel number
                           (3,3), # kernel size
                           padding = 'same',
                           activation = tf.nn.relu,
                           name = 'conv1_2')

# 16 * 16
pooling1 = tf.layers.max_pooling2d(conv1_2,
                                   (2, 2), # kernel size
                                   (2, 2), # stride
                                   name = 'pool1')


conv2_1 = tf.layers.conv2d(pooling1,
                           32, # output channel number
                           (3,3), # kernel size
                           padding = 'same',
                           activation = tf.nn.relu,
                           name = 'conv2_1')
conv2_2 = tf.layers.conv2d(conv2_1,
                           32, # output channel number
                           (3,3), # kernel size
                           padding = 'same',
                           activation = tf.nn.relu,
                           name = 'conv2_2')
# 8 * 8
pooling2 = tf.layers.max_pooling2d(conv2_2,
                                   (2, 2), # kernel size
                                   (2, 2), # stride
                                   name = 'pool2')

conv3_1 = tf.layers.conv2d(pooling2,
                           32, # output channel number
                           (3,3), # kernel size
                           padding = 'same',
                           activation = tf.nn.relu,
                           name = 'conv3_1')
conv3_2 = tf.layers.conv2d(conv3_1,
                           32, # output channel number
                           (3,3), # kernel size
                           padding = 'same',
                           activation = tf.nn.relu,
                           name = 'conv3_2')
# 4 * 4 * 32
pooling3 = tf.layers.max_pooling2d(conv3_2,
                                   (2, 2), # kernel size
                                   (2, 2), # stride
                                   name = 'pool3')
# [None, 4 * 4 * 32]
flatten = tf.layers.flatten(pooling3)
y_ = tf.layers.dense(flatten, 10)

loss = tf.losses.sparse_softmax_cross_entropy(labels=y, logits=y_)
# y_ -> sofmax
# y -> one_hot
# loss = ylogy_

# indices
predict = tf.argmax(y_, 1)
# [1,0,1,1,1,0,0,0]
correct_prediction = tf.equal(predict, y)
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float64))

with tf.name_scope('train_op'):
    train_op = tf.train.AdamOptimizer(1e-3).minimize(loss)

4.4 Model training and breakpoint continuation training （ The key ）

init = tf.global_variables_initializer()
saver = tf.train.Saver(max_to_keep=1)   #Saver Class provides methods for saving and restoring models  

# train 10k: 73.4%
with tf.Session() as sess:
    sess.run(init)
    ckpt = tf.train.get_checkpoint_state(MODEL_SAVE_PATH)  
    # Equivalent to the previous sentence ：ckpt = tf.train.latest_checkpoint("./ckpt/")
    print(ckpt)  # The latest saved model name

    if ckpt and ckpt.model_checkpoint_path:
        saver.restore(sess, ckpt.model_checkpoint_path)# Resume the current session , take ckpt The value in is assigned to w and b
        steped = re.findall(r"\d+\.?\d*", str(ckpt))   # Extract numbers from strings 
        print('the step finished last time is ' + steped[0])  # Output the last training times 
        steped = int(steped[0])                # Ensure that the total number of training is certain 
        print('Model restored...')
    else:
        steped = 0
        print('No model')      
    for step in range(train_steps-steped):
        batch_data, batch_labels = train_data.next_batch(batch_size)
        loss_val, acc_val, _ = sess.run(
            [loss, accuracy, train_op],
            feed_dict={
    
                x: batch_data,
                y: batch_labels})
        saver.save(sess, os.path.join(MODEL_SAVE_PATH, MODEL_NAME), global_step=step+steped+1)
        if (step+steped+1) % 100 == 0:      # Calculate the output from the last training times 
            print('[Train] Step: %d, loss: %4.5f, acc: %4.5f' 
                  % (step+steped+1, loss_val, acc_val))
        if (step+steped+1) % 1000 == 0:
            test_data = CifarData(test_filenames, False)
            all_test_acc_val = []
            for j in range(test_steps):
                test_batch_data, test_batch_labels \
                    = test_data.next_batch(batch_size)
                test_acc_val = sess.run(
                    [accuracy],
                    feed_dict = {
    
                        x: test_batch_data, 
                        y: test_batch_labels
                    })
                all_test_acc_val.append(test_acc_val)
            test_acc = np.mean(all_test_acc_val)
            print('[Test ] Step: %d, acc: %4.5f' % (step+steped+1, test_acc))

4.5 The training effect of breakpoint continuation training

It's not easy to code words , If there is any help, please give me a like , thank you ~