使用keras建立checkpoint

1.简介

checkpoint是一种系统状态的快照方法，可以直接使用。checkpoint是模型的权重，可以用来预测，也可以用来继续训练。
keras中的回调函数callbacks提供了checkpoint功能。
Tensorboard是一种训练可视化的操作。在keras的回调函数中也有相应的功能。
下面这个示例，将两种的情况都包涵在内了。

2.示例

#!/usr/bin/env python
# encoding: utf-8

import pandas as pd
import numpy as np
import matplotlib.pylab as plt
from keras.models import Sequential
from keras.layers import Dense
from keras.layers import LSTM
from sklearn.preprocessing import MinMaxScaler
from sklearn.metrics import mean_squared_error
from keras.callbacks import ModelCheckpoint
from keras.callbacks import TensorBoard
#读取数据
dataf = pd.read_csv("./data/data_complete.csv")[['sump']].values[0:2*288]
#print(dataf)
def create_dataset(dataset, timesteps):
    datax=[]
    datay=[]
    for each in range(len(dataset)-timesteps):
        x = dataset[each:each+timesteps,0]
        y = dataset[each+timesteps,0]
        datax.append(x)
        datay.append(y)
    return np.array(datax),np.array(datay)

#构造train and test
scaler = MinMaxScaler(feature_range=(0,1))
dataf = scaler.fit_transform(dataf)
trainsize = int(len(dataf)*0.7)
train = dataf[0:trainsize]
test = dataf[trainsize:len(dataf)]
timesteps = 288
trainx, trainy = create_dataset(train, timesteps)
testx, testy = create_dataset(test,timesteps)
#print(trainx)
#print(trainx.shape)
#变换
trainx = np.reshape(trainx,(trainx.shape[0],timesteps,1))
testx = np.reshape(testx, (testx.shape[0],timesteps,1))
#print(trainx)
#print(trainx.shape)
#lstm
model = Sequential()
model.add(LSTM(4,input_shape=(timesteps,1)))
model.add(Dense(1))
#model.load_weights("")#可以在这里加载checkpoint权重模型，继续训练。
model.compile(loss="mean_squared_error",optimizer="adam",metrics=["accuracy"])#metrics要设置
#checkpoint
filepath = "weights-impovement-{epoch:02d}--{val_acc:.2f}.hdf5"
checkpoint = ModelCheckpoint(filepath,monitor="val_acc", verbose=1,
                             save_best_only=True,mode="max")#checkpoint的示例
callbacks_list = [checkpoint]
tensorboard = TensorBoard(log_dir="log")#tensorboard的示例
callbacks_tensor = [tensorboard]
model.fit(trainx,trainy, epochs=3, batch_size=5,
          validation_split=0.25,callbacks=callbacks_tensor)#callbacks=callbacks_list
#
# #predict
# train_predict = model.predict(trainx)
# test_predict = model.predict(testx)
# #invert
# train_predict = scaler.inverse_transform(train_predict)
# trainy = scaler.inverse_transform([trainy])
# print(train_predict[0],trainy[0])
#
# test_predict = scaler.inverse_transform(test_predict)
# testy = scaler.inverse_transform([testy])
# #error
# train_score = np.sqrt(mean_squared_error(trainy[0],train_predict[:,0]))
# test_score = np.sqrt(mean_squared_error(testy[0],test_predict[:,0]))
# print("train score RMSE: %.2f"% train_score)
# print("test score RMSE: %.2f"% test_score)
#
# #plot
# # shift train predictions for plotting
# trainPredictPlot = np.empty_like(dataf)
# trainPredictPlot[:, :] = np.nan
# trainPredictPlot[timesteps:len(train_predict)+timesteps, :] = train_predict
#
# # shift test predictions for plotting
# testPredictPlot = np.empty_like(dataf)
# testPredictPlot[:,:] = np.nan
# testPredictPlot[len(train_predict)+(timesteps*2):len(dataf), :] = test_predict
#
# # plot baseline and predictions
# plt.plot(scaler.inverse_transform(dataf))
# plt.plot(trainPredictPlot)
# plt.plot(testPredictPlot)
# plt.show()

checkpoint:
Train on 86 samples, validate on 29 samples
Epoch 1/3

5/86 [>…] - ETA: 22s - loss: 9.5544e-04 - acc: 0.0000e+00
10/86 [>…] - ETA: 12s - loss: 9.2816e-04 - acc: 0.1000
15/86 [==>…] - ETA: 8s - loss: 7.5216e-04 - acc: 0.0667

Epoch 00001: val_acc improved from -inf to 0.00000, saving model to weights-impovement-01–0.00.hdf5
Epoch 2/3
Epoch 00003: val_acc did not improve from 0.00000

tensorboard:
生成一个log文件夹，理由有个tensorboard文件。
在cmd下.py脚本目录下执行：tensorboard --logdir=log
然后在浏览器localhost:6006，即可看到可视化过程。

3.参考

如何为Keras中的深度学习模型建立Checkpoint