keras 两种训练模型方式详解fit和fit_generator(节省内存)

import keras
from keras.models import Sequential
from keras.layers import Dense
import numpy as np
from sklearn.preprocessing import LabelEncoder
from sklearn.preprocessing import OneHotEncoder
from sklearn.model_selection import train_test_split
#读取数据
x_train = np.load("D:\\machineTest\\testmulPE_win7\\data_sprase.npy")[()]
y_train = np.load("D:\\machineTest\\testmulPE_win7\\lable_sprase.npy")
# 获取分类类别总数
classes = len(np.unique(y_train))
#对label进行one-hot编码，必须的
label_encoder = LabelEncoder()
integer_encoded = label_encoder.fit_transform(y_train)
onehot_encoder = OneHotEncoder(sparse=False)
integer_encoded = integer_encoded.reshape(len(integer_encoded), 1)
y_train = onehot_encoder.fit_transform(integer_encoded)
#shuffle
X_train, X_test, y_train, y_test = train_test_split(x_train, y_train, test_size=0.3, random_state=0)
model = Sequential()
model.add(Dense(units=1000, activation='relu', input_dim=784))
model.add(Dense(units=classes, activation='softmax'))
model.compile(loss='categorical_crossentropy',
optimizer='sgd',
metrics=['accuracy'])
model.fit(X_train, y_train, epochs=50, batch_size=128)
score = model.evaluate(X_test, y_test, batch_size=128)
# #fit参数详情
# keras.models.fit(
# self,
# x=None, #训练数据
# y=None, #训练数据label标签
# batch_size=None, #每经过多少个sample更新一次权重，defult 32
# epochs=1, #训练的轮数epochs
# verbose=1, #0为不在标准输出流输出日志信息，1为输出进度条记录，2为每个epoch输出一行记录
# callbacks=None,#list，list中的元素为keras.callbacks.Callback对象，在训练过程中会调用list中的回调函数
# validation_split=0., #浮点数0-1，将训练集中的一部分比例作为验证集，然后下面的验证集validation_data将不会起到作用
# validation_data=None, #验证集
# shuffle=True, #布尔值和字符串，如果为布尔值，表示是否在每一次epoch训练前随机打乱输入样本的顺序，如果为"batch"，为处理HDF5数据
# class_weight=None, #dict,分类问题的时候，有的类别可能需要额外关注，分错的时候给的惩罚会比较大，所以权重会调高，体现在损失函数上面
# sample_weight=None, #array,和输入样本对等长度,对输入的每个特征+个权值，如果是时序的数据，则采用(samples，sequence_length)的矩阵
# initial_epoch=0, #如果之前做了训练，则可以从指定的epoch开始训练
# steps_per_epoch=None, #将一个epoch分为多少个steps，也就是划分一个batch_size多大，比如steps_per_epoch=10，则就是将训练集分为10份，不能和batch_size共同使用
# validation_steps=None, #当steps_per_epoch被启用的时候才有用，验证集的batch_size
# **kwargs #用于和后端交互
# )
# 
# 返回的是一个History对象，可以通过History.history来查看训练过程，loss值等等

# 第二种,可以节省内存
'''
Created on 2018-4-11
fit_generate.txt，后面两列为lable,已经one-hot编码
1 2 0 1
2 3 1 0
1 3 0 1
1 4 0 1
2 4 1 0
2 5 1 0
'''
import keras
from keras.models import Sequential
from keras.layers import Dense
import numpy as np
from sklearn.model_selection import train_test_split
count =1 
def generate_arrays_from_file(path):
global count
while 1:
datas = np.loadtxt(path,delimiter=' ',dtype="int")
x = datas[:,:2]
y = datas[:,2:]
print("count:"+str(count))
count = count+1
yield (x,y)
x_valid = np.array([[1,2],[2,3]])
y_valid = np.array([[0,1],[1,0]])
model = Sequential()
model.add(Dense(units=1000, activation='relu', input_dim=2))
model.add(Dense(units=2, activation='softmax'))
model.compile(loss='categorical_crossentropy',
optimizer='sgd',
metrics=['accuracy'])
model.fit_generator(generate_arrays_from_file("D:\\fit_generate.txt"),steps_per_epoch=10, epochs=2,max_queue_size=1,validation_data=(x_valid, y_valid),workers=1)
# steps_per_epoch 每执行一次steps,就去执行一次生产函数generate_arrays_from_file
# max_queue_size 从生产函数中出来的数据时可以缓存在queue队列中
# 输出如下:
# Epoch 1/2
# count:1
# count:2
# 
# 1/10 [== ...........................] - ETA: 2s - loss: 0.7145 - acc: 0.3333count:3
# count:4
# count:5
# count:6
# count:7
# 
# 7/10 [==================== .........] - ETA: 0s - loss: 0.7001 - acc: 0.4286count:8
# count:9
# count:10
# count:11
# 
# 10/10 [==============================] - 0s 36ms/step - loss: 0.6960 - acc: 0.4500 - val_loss: 0.6794 - val_acc: 0.5000
# Epoch 2/2
# 
# 1/10 [== ...........................] - ETA: 0s - loss: 0.6829 - acc: 0.5000count:12
# count:13
# count:14
# count:15
# 
# 5/10 [============== ...............] - ETA: 0s - loss: 0.6800 - acc: 0.5000count:16
# count:17
# count:18
# count:19
# count:20
# 
# 10/10 [==============================] - 0s 11ms/step - loss: 0.6766 - acc: 0.5000 - val_loss: 0.6662 - val_acc: 0.5000

#coding=utf-8
'''
Created on 2018-7-10
'''
import keras
import math
import os
import cv2
import numpy as np
from keras.models import Sequential
from keras.layers import Dense
class DataGenerator(keras.utils.Sequence):
def __init__(self, datas, batch_size=1, shuffle=True):
self.batch_size = batch_size
self.datas = datas
self.indexes = np.arange(len(self.datas))
self.shuffle = shuffle
def __len__(self):
#计算每一个epoch的迭代次数
return math.ceil(len(self.datas) / float(self.batch_size))
def __getitem__(self, index):
#生成每个batch数据，这里就根据自己对数据的读取方式进行发挥了
# 生成batch_size个索引
batch_indexs = self.indexes[index*self.batch_size:(index+1)*self.batch_size]
# 根据索引获取datas集合中的数据
batch_datas = [self.datas[k] for k in batch_indexs]
# 生成数据
X, y = self.data_generation(batch_datas)
return X, y
def on_epoch_end(self):
#在每一次epoch结束是否需要进行一次随机，重新随机一下index
if self.shuffle == True:
np.random.shuffle(self.indexes)
def data_generation(self, batch_datas):
images = []
labels = []
# 生成数据
for i, data in enumerate(batch_datas):
#x_train数据
image = cv2.imread(data)
image = list(image)
images.append(image)
#y_train数据 
right = data.rfind("\\",0)
left = data.rfind("\\",0,right)+1
class_name = data[left:right]
if class_name=="dog":
labels.append([0,1])
else: 
labels.append([1,0])
#如果为多输出模型，Y的格式要变一下，外层list格式包裹numpy格式是list[numpy_out1,numpy_out2,numpy_out3]
return np.array(images), np.array(labels)
# 读取样本名称，然后根据样本名称去读取数据
class_num = 0
train_datas = [] 
for file in os.listdir("D:/xxx"):
file_path = os.path.join("D:/xxx", file)
if os.path.isdir(file_path):
class_num = class_num + 1
for sub_file in os.listdir(file_path):
train_datas.append(os.path.join(file_path, sub_file))
# 数据生成器
training_generator = DataGenerator(train_datas)
#构建网络
model = Sequential()
model.add(Dense(units=64, activation='relu', input_dim=784))
model.add(Dense(units=2, activation='softmax'))
model.compile(loss='categorical_crossentropy',
optimizer='sgd',
metrics=['accuracy'])
model.compile(optimizer='sgd', loss='categorical_crossentropy', metrics=['accuracy'])
model.fit_generator(training_generator, epochs=50,max_queue_size=10,workers=1)