Early stoping和checkpoint在tensofrflow中的使用

在訓練中使用early stoping終止模型和使用checkpoint儲存模型

當我們訓練模型時，如果epoch設定太長，常常希望可以在loss不再下降或者accuracy不再提高時終止訓練，獲得模型，避免模型浪費時間，這時可以使用tensorflow 中的early stoping終止模型和使用checkpoint儲存模型：

import tensorflow as tf
from tensorflow.keras import datasets, layers, optimizers, Sequential, metrics
import pandas as pd
import numpy as np
from scipy import sparse
from tensorflow.keras.callbacks import EarlyStopping, ModelCheckpoint #導入early stopping，checkpoint
import os

os.environ["CUDA_VISIBLE_DEVICES"] = "0"

batchsz = 256
....
....
.... # 用過pandas讀取資料并建構你自己的x_train,y_train,x_val,y_val,x_test,y_test
db = tf.data.Dataset.from_tensor_slices((x_train, y_train))
db = db.shuffle(6000).batch(batchsz)
ds_val = tf.data.Dataset.from_tensor_slices((x_val, y_val))
ds_val = ds_val.batch(batchsz)
db_test = tf.data.Dataset.from_tensor_slices((x_test, y_test))
db_test = db_test.batch(batchsz)

sample = next(iter(db))
print(sample[0].shape, sample[1].shape)

network = Sequential([layers.Conv1D(512, kernel_size=5, strides=3,padding='same',activation=tf.nn.relu),
                     layers.Dropout(0.4),
                     layers.Conv1D(256, kernel_size=5, strides=3,padding='same',activation=tf.nn.relu),
                     layers.Dropout(0.4),
                     layers.Conv1D(128, kernel_size=5, strides=3,padding='same',activation=tf.nn.relu),
                     layers.Dropout(0.4),
                     layers.Flatten(),
                     layers.Dense(128, activation=tf.nn.relu),
                     layers.Dropout(0.4),
                     layers.Dense(1,activation='sigmoid')]) # 建構你自己network

network.build(input_shape=(None, 2000,4))
network.summary()

# 建構early-stopping，監測目标是測試集的accuracy，8個epoch如果提高不到0。001即終止
early_stopping = EarlyStopping(monitor='val_acc',min_delta=0.001,patience=8) 

#建構checkpoint，儲存model名字為‘conv.h5‘，監測名額為測試集準确率val_acc，model=‘max’儲存val_acc最大的的，save_best_only=true,儲存最好的模型
checkpoint=ModelCheckpoint('conv.h5', monitor='val_acc',model='max',verbose=1,save_best_only=True) 

network.compile(optimizer=optimizers.Adam(lr=0.001),
                loss='binary_crossentropy',
                metrics=['accuracy']
               )

# 在callback中使用early-stopping和checkpoint。
network.fit(db, epochs=100, validation_data=ds_val, validation_steps=2,callbacks=[early_stopping，checkpoint]) 
network.evaluate(db_test)

#測試
sample = next(iter(db_test))
x = sample[0]
y = sample[1] 
pred = network.predict(x)  
           

如果想進一步了解early stoppong，可以檢視jason的文章Use Early Stopping to Halt the Training of Neural Networks At the Right Time