keras入门参考网址：
keras分类应用里的人脸预测kaggle：
人脸表情分类与识别：opencv人脸检测+Keras情绪分类（四）
数据量大无法载入时，节约内存model.fit_generator：
1 def generate_arrays_from_file(path):
2 while 1:
3 f = open(path)
4 for line in f:
5 # create Numpy arrays of input data
6 # and labels, from each line in the file
7 x, y = process_line(line)
8 yield (x, y)
9 f.close()
10
11 model.fit_generator(generate_arrays_from_file('/my_file.txt'),
12 samples_per_epoch=10000, nb_epoch=10)
13 ---------------------
14 作者：ShellCollector
15 来源：CSDN
16 原文：https://blog.csdn.net/jacke121/article/details/78877352
1 def generate_batch_data_random(x, y, batch_size): 2 """逐步提取batch数据到显存，降低对显存的占用""" 3 ylen = len(y) 4 loopcount = ylen // batch_size 5 while (True): 6 i = randint(0,loopcount) 7 yield x[i * batch_size:(i + 1) * batch_size], y[i * batch_size:(i + 1) * batch_size]
keras 两种训练模型方式fit和fit_generator(节省内存) 提供了对比，fit 和 fit_generator 两种代码
国外一个大佬Adrian Rosebrock写的博客：实现自定义Keras fit_generator函数（中文） How to use Keras fit and fit_generator (a hands-on tutorial) （英文）
利用fit_generator最小化显存占用比率/数据Batch化
1 #从节省内存的角度，通过生成器的方式来训练 2 def data_generator(data, targets, batch_size): 3 idx = np.arange(len(data)) 4 np.random.shuffle(idx) 5 batches = [idx[range(batch_size*i, min(len(data), batch_size*(i+1)))] for i in range(len(data)/batch_size+1)] 6 while True: 7 for i in batches: 8 xx, yy = np.array(map(gen_matrix, data[i])), np.array(map(gen_target, targets[i])) 9 yield (xx, yy) 10 11 12 batch_size = 1024 13 history = model.fit_generator(data_generator(d['words'], d['label'], batch_size), samples_per_epoch=len(d), nb_epoch=200) 14 model.save_weights('words_seq2seq_final_1.model')
keras数据自动生成器，继承keras.utils.Sequence，结合fit_generator实现节约内存训练
提前shuffle数据：
shuffle数据并且设置callback：存储最佳weight
1 np.random.seed(1024) 2 random.shuffle(index) 3 data = data[index] 4 label = label[index] 5 splitpoint = int(round(num * 0.8)) 6 (X_train, X_val) = (data[0:splitpoint], data[splitpoint:]) 7 (Y_train, Y_val) = (label[0:splitpoint], label[splitpoint:]) 8 X_train=X_train/255 9 X_val=X_val/255
提供了三种方法，非常详细