Tensorflow实现迁移学习

当训练样本比较少，可以使用预先训练好的模型，也可以将预训练好的模型用来初始化。本文介绍如何使用Tensorflow进行迁移学习。

预处理

import glob
import os.path
import numpy as np
import tensorflow as tf
from tensorflow.python.platform import gfile
import tensorflow.contrib.slim as slim

# 输入数据目录，共有5个子目录，分别对应5种类别的花
INPUT_DATA = '/home/seisinv/data/flower_photos/'
OUTPUT_DATA = '/home/seisinv/data/flower_photos_process.npy'

# 测试和验证数据的比例
VALIDATION_PERCENTAGE = 10
TEST_PERCENTAGE = 10

# 读取数据，并分割成训练数据、验证数据和测试数据
def create_image_list(sess, testing_percentage, validation_percentage):
    sub_dirs = [x[0] for x in os.walk(INPUT_DATA)]
    is_root_dir = True
    
    # 初始化各个数据集
    training_images = []
    training_labels = []
    testing_images = []
    testing_labels = []
    validation_images = []
    validation_labels = []
    current_label = 0
    
    # 读取子目录
    for sub_dir in sub_dirs:
        if is_root_dir:
            is_root_dir = False
            continue
        # 获取一个子目录中所有的图片文件
        extensions = ['jpg', 'jpeg', 'JPG', 'JPEG']
        file_list = []
        dir_name = os.path.basename(sub_dir)
        for extension in extensions:
            file_glob = os.path.join(INPUT_DATA, dir_name, '*.', extension)
            file_list.extend(glob.glob(file_glob))
            if not file_list: continue
            
            # 处理图片数据
            for file_name in file_list:
                # 读取图片
                image_raw_data = gfile.FastGFile(file_name,'rb').read()
                # 解析图片
                image = tf.image.decode_jpeg(image_raw_data)
                # 图片转为float32型
                if image.dtype != tf.float32:
                    image = tf.image_convert_image_dtype(
                        image, dtype=tf.float32)
                # 图片大小改为299*299，以满足inception-v3模型的需求
                image = tf.image_resize_images(image, [299, 299])
                image_value = sess.run(image)
                
                # 随机划分数据集
                chance = np.random.randint(100)
                if chance < validation_percentage:
                    validation_images.append(image_value)
                    validation_labels.append(current_label)
                elif chance < (testing_percentage + validation_percentage):
                    testing_images.append(image_value)
                    validation_labels.append(current_label)
                else:
                    training_images.append(image_value)
                    training_labels.append(current_label)
            current_label += 1
            
    # 将训练速度打乱
    state = np.random.get_state()
    np.random.shuffle(training_images)
    np.random.set_state(state)
    np.random.shuffle(training_labels)
    
    return np.asarray([training_images, training_labels,
                      validation_images, validation_labels,
                      testing_images, testing_labels])

tf.reset_default_graph()

with tf.Session() as sess:
    processed_data = create_image_list(
    sess, TEST_PERCENTAGE, VALIDATION_PERCENTAGE)
    np.save(OUTPUT_DATA, processed_data)

---------------------------------------------------------------------------

FileNotFoundError                         Traceback (most recent call last)

<ipython-input-3-89177c393fa3> in <module>()
      4     processed_data = create_image_list(
      5     sess, TEST_PERCENTAGE, VALIDATION_PERCENTAGE)
----> 6     np.save(OUTPUT_DATA, processed_data)


~/anaconda3/envs/fwi_ai/lib/python3.5/site-packages/numpy/lib/npyio.py in save(file, arr, allow_pickle, fix_imports)
    488         if not file.endswith('.npy'):
    489             file = file + '.npy'
--> 490         fid = open(file, "wb")
    491         own_fid = True
    492     elif is_pathlib_path(file):


FileNotFoundError: [Errno 2] No such file or directory: '/home/seisinv/data/flower_photos_process.npy'

迁移学习

# 加载Tensorflow-Slim定义好的inception_v3模型
import tensorflow.contrib.slim.python.slim.nets.inception_v3 as inception_v3

# 处理后的数据
INPUT_DATA = '/home/seisinv/data/flower_photos_process.npy'
# 保存训练后的模型
TRAIN_FILE = '/home/seisinv/data/model'
# 谷歌提供的训练好的模型
CKPT_FILE = '/home/seisinv/data/model/inception_v3.ckpt'

# 定义训练过程所使用的参数
LEARNING_RATE = 0.0001
STEPS = 300
BATCH = 23
N_CLASSES = 5

# 不需要从预训练模型中加载的参数名称，这里指最后的全连接层，给出参数的前缀
CHECKPOINT_EXCLUDE_SCOPES = 'InceptionV3/Logits,InceptionV3/AuxLogits'
# 需要训练的参数名称，同样指的是最后的全连接层，给出参数前缀
TRAINABLE_SCOPES = 'InceptionV3/Logits,InceptionV3/AuxLogits'

# 获取所有需要从预训练模型中加载的参数
def get_tuned_variables():
    exclusions = [scope.strip() for scope in\
                 CHECKPOINT_EXCLUDE_SCOPES.split(',')]
    variables_to_restore = []
    for var in slim.get_model_variables():
        excluded = False
        for exclusion in exclusions:
            if var.op.name.startswith(exclusion):
                excluded = True
                break
        if not excluded:
            variables_to_restore.appen(var)
    return variables_to_restore

# 获取所有需要训练的变量列表
def get_trainable_variable():
    scopes = [scope.strip() for scope in\
             TRAINABLE_SCOPES.split(',')]
    variables_to_train = []
    for scope in scopes:
        variables = tf.get_collection(
            tf.GraphKeys.TRAINABLE_VARIABLES, scope)
        variables_to_train.extend(variables)
    return variables_to_train

def main():
    processed_data = np.load(INPUT_DATA)
    training_images = processed_data[0]
    n_training_examples = len(training_images)
    training_labels = processed_data[1]
    validation_images = processed_data[2]
    validation_labels = processed_data[3]
    testing_images = processed_data[4]
    testing_labels = processed_data[5]
    print("%d training examples, %d validation examples and %d "
         "testing examples." %(n_training_examples, len(validation_images), len(testing_images)))
    
    # 定义inception-v3的输入
    images = tf.placeholder(
        tf.float32, [None, 299, 299, 3],
        name='input_image')
    labels = tf.placeholder(
        tf.int64, [None], name='labels')

print("a""b")

ab

tf.reset_default_graph()

更多谷歌训练好的模型可以参考这里

参考资料

• 郑泽宇、梁博文和顾思宇，Tensorflow: 实战Google深度学习框架（第二版）