最基础的tensorflow代码学习方法———多敲多理解
我最近在学习tensorflow实战深度学习框架，发现书上的部分代码看起来很简单，理论部分也比较基础。可是当我自己想要敲点代码的时候，发现怎么也敲不出来，于是我反思，到底哪里出了问题。原来我为了求快，迅速的把tensorflow的教程和深度学习的书本看了一遍，只是基本的知道了里面的理论和原理，具体的代码细节并不理解。这犯了初学者的大忌：只看不动手实践。
程序员的代码学习方法——多敲多理解，每学一件新的东西（算法、程序），必须要把它敲出来，只有敲出来，理解了才是你的东西。不然哪怕你copy过来实现了基本功能，你也不理解细节。
像以下最为基础的猫狗分类代码，如果不自己亲自写一遍，对于其中的细节也只能停留在表面的理解上，永远得不到具体的理解。
# -*- coding: utf-8 -*-
"""
Created on Fri Mar 19 19:27:26 2021
cat_VS_dag 猫狗分类识别
@author: chenfeng
"""
import tensorflow as tf
from tensorflow import keras
import numpy as np
import os
import random
num_epochs = 100 #遍历数据集的次数
num_class = 2 #类别数
num_train = 23000 #训练样本总数
num_test = 2000 #测试样本总数
input_rows,input_cols = 100,100 #输入图片尺寸
batch_size = 32 #批次大小
learning_rate = 0.01 #初始学习率
loss_value = [] #记录损失值
num_parallel_calls = 4
#================================================================#
#=注意！！！num_parallel_calls是数据增强使使用的CPU核心数，根据实际调整=#
#================================================================#
DATA_DIR = 'E:/UserData/Desktop/catVSdog_dataSet/' #数据根目录
LABLE = {0:'cats',1:'dogs'} #分类标签
#'F:/data/cats-and-dogs/train/cats' #猫的训练样本路径，其余类似
#============================ model ============================#
#=================================================================#
def CNN():
model = keras.Sequential()
#block1
model.add(keras.layers.Conv2D(32,[3,3],padding = 'same',kernel_initializer='he_normal',activation = tf.nn.elu,input_shape = (input_rows,input_cols,3)))
model.add(keras.layers.BatchNormalization())
model.add(keras.layers.Conv2D(32,[3,3],padding = 'same',kernel_initializer='he_normal',activation = tf.nn.elu))
model.add(keras.layers.BatchNormalization())
model.add(keras.layers.MaxPool2D(pool_size = [2,2]))
model.add(keras.layers.Dropout(0.2))
#block2
model.add(keras.layers.Conv2D(64,[3,3],padding = 'same',kernel_initializer='he_normal',activation = tf.nn.elu))
model.add(keras.layers.BatchNormalization())
model.add(keras.layers.Conv2D(64,[3,3],padding = 'same',kernel_initializer='he_normal',activation = tf.nn.elu))
model.add(keras.layers.BatchNormalization())
model.add(keras.layers.MaxPool2D(pool_size = [2,2]))
model.add(keras.layers.Dropout(0.2))
#block3
model.add(keras.layers.Conv2D(128,[3,3],padding = 'same',kernel_initializer='he_normal',activation = tf.nn.elu))
model.add(keras.layers.BatchNormalization())
model.add(keras.layers.Conv2D(128,[3,3],padding = 'same',kernel_initializer='he_normal',activation = tf.nn.elu))
model.add(keras.layers.BatchNormalization())
model.add(keras.layers.MaxPool2D())
model.add(keras.layers.Dropout(0.2))
#block4
model.add(keras.layers.Conv2D(256,[3,3],padding = 'same',kernel_initializer='he_normal',activation = tf.nn.elu))
model.add(keras.layers.BatchNormalization())
model.add(keras.layers.Conv2D(256,[3,3],padding = 'same',kernel_initializer='he_normal',activation = tf.nn.elu))
model.add(keras.layers.BatchNormalization())
model.add(keras.layers.MaxPool2D())
model.add(keras.layers.Dropout(0.5))
#block5
model.add(keras.layers.Flatten())
model.add(keras.layers.Dense(64,activation = tf.nn.elu,kernel_initializer='he_normal'))
model.add(keras.layers.BatchNormalization())
model.add(keras.layers.Dropout(0.5))
#block6
model.add(keras.layers.Dense(32,activation = tf.nn.elu,kernel_initializer='he_normal'))
model.add(keras.layers.BatchNormalization())
model.add(keras.layers.Dropout(0.5))
#block7
model.add(keras.layers.Dense(num_class,activation = tf.nn.softmax,kernel_initializer='he_normal'))
print(model.summary(),'\n')
return model
#============================= data ============================#
#=================================================================#
def data_set(DATA_DIR,option): #根据option是train还是test加载对应文件夹的数据
if option == 'train':
k = num_train
else:
k = num_test
Filenames = []
Lable = np.zeros(shape = (k,1))
i = 0
for lable in LABLE: #lable是0即进入cats文件夹，是1则进入dogs文件夹
for filename in os.listdir(DATA_DIR + option + '/' + LABLE[lable]):
Filenames.append(DATA_DIR + option + '/' + LABLE[lable] + '/' + filename)
Lable[i] = (lable)
i += 1
return Filenames,Lable #返回图片地址list和对应标签np数组
def map_fun(filename,lable): #数据增强函数，包含旋转，镜像，翻转
image_string = tf.io.read_file(filename)
image = tf.image.decode_jpeg(image_string,channels =3)
#必须要进行上面两步，将图像地址转换为图像数据
rand = random.random()
if rand > 0.5:
image = tf.image.rot90(image)
image = tf.image.random_flip_left_right(image)
image = tf.image.random_flip_up_down(image)
image = image / 255 #数据处理为[0,1]区间的值，据说有利于计算和收敛
return image,lable
#下面是main中的内容，给搬过来了
#======================== funs and main ========================#
#=================================================================#
TRAN_Filenames,TRAN_Lable = data_set(DATA_DIR,'train')
tran_dataset = tf.data.Dataset.from_tensor_slices((TRAN_Filenames,TRAN_Lable))
#上面这步需要注意注意再注意！！！一定是(TRAN_Filenames,TRAN_Lable)，带上括号作为参数，不然，嘿嘿，报错都找不见在哪
tran_dataset = tran_dataset.map(map_func = map_fun,num_parallel_calls = num_parallel_calls)
tran_dataset = tran_dataset.shuffle(23000)
tran_dataset = tran_dataset.batch(batch_size)
tran_dataset = tran_dataset.prefetch(tf.data.experimental.AUTOTUNE)
TEST_Filenames,TEST_Lable = data_set(DATA_DIR,'test')
test_dataset = tf.data.Dataset.from_tensor_slices((TEST_Filenames,TEST_Lable))
test_dataset = test_dataset.map(map_func = map_fun,num_parallel_calls=num_parallel_calls)
test_dataset = test_dataset.batch(batch_size)
#======================== funs and main ========================#
#=================================================================#
class LossHistory(keras.callbacks.Callback):
def on_train_begin(self,logs = {}):
self.losses = []
def on_batch_end(self,batch,logs = {}):
self.losses.append(logs.get('loss'))
if __name__ == '__main__':
TRAN_Filenames,TRAN_Lable = data_set(DATA_DIR,'train')
tran_dataset = tf.data.Dataset.from_tensor_slices((TRAN_Filenames,TRAN_Lable))
tran_dataset = tran_dataset.map(map_func = map_fun,num_parallel_calls=num_parallel_calls)
tran_dataset = tran_dataset.shuffle(23000)
tran_dataset = tran_dataset.batch(batch_size)
tran_dataset = tran_dataset.prefetch(tf.data.experimental.AUTOTUNE)
TEST_Filenames,TEST_Lable = data_set(DATA_DIR,'test')
test_dataset = tf.data.Dataset.from_tensor_slices((TEST_Filenames,TEST_Lable))
test_dataset = test_dataset.map(map_func = map_fun,num_parallel_calls=num_parallel_calls)
test_dataset = test_dataset.batch(batch_size)
#================上面是data相关处理================#
model = CNN() #赋予模型
model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate),
loss=tf.keras.losses.sparse_categorical_crossentropy,
metrics=[tf.keras.metrics.sparse_categorical_accuracy])
#使用compile函数定义优化器optimizer，详见资料1
#以下四个参数与callbacks的使用有关
reduce_lr = keras.callbacks.ReduceLROnPlateau(monitor = 'val_loss',
factor = 0.2,
patience = 4,
verbose = 1,
min_delta = 0.00001)
#动态调整学习率
earlystop = keras.callbacks.EarlyStopping(monitor = 'val_loss',
min_delta = 0,
patience = 10,
verbose = 1,
restore_best_weights = True)
#当模型不在进步时提前结束训练
history = LossHistory()
#通过LossHistory将训练loss值保存history中
checkpoint = keras.callbacks.ModelCheckpoint('best_model.h5',
monitor = 'val_loss',
mode = 'min',
save_best_only = True,
verbose = 1)
#查看当前模型是否最优，只保存训练中最优的模型
callbacks = [history,reduce_lr,earlystop,checkpoint]
#将需要参数打包成list
model.fit(tran_dataset,batch_size = batch_size,epochs = num_epochs, callbacks = callbacks,validation_data = test_dataset)
#训练模型，设置callbacks，validation_data = test_dataset可以在一个epoch后用测试集进行测试，
#上面reduce_lr等中的'val_loss'就是测试集的loss，以此作为保存最优模型，和调整学习率的指标
#tf.saved_model.save(model, "F:/data/cats-and-dogs")