How many rounds of deep learning training? How many iterations?

Deep learning is one of the hottest research directions nowadays . With its excellent learning ability , Realized AI The key function of . Generally speaking , One deep study is certainly not enough , So how many rounds of deep learning training ？ How many iterations ？ Issues of concern , Here comes the answer .

How many iterations of deep learning training ？

I have turned over your research achievements , Trained 10 Iterations , Yes 30 individual , There is even 1 More than 10000 . This makes Xiaobai confused , How many training sessions is appropriate ? What is the test of training ?

Some big bulls just say how many times they have trained , Nor explain why , Let's say it's based on experience , We can't learn this . Other experts have given some reasons , Let's summarize .

01 When loss The iteration ends when the value converges

A key principle of deep learning is to compare learning results with sample labels . In theory, the smaller the gap , It shows that the better the effect of learning . The gap is loss value .

Loss The value cannot change to 0, It can only be approached infinitely 0. So you can think of it through your toes , When loss When the value cannot be reduced , This is called convergence , Is the end of learning . Such as the following figure ：

02 Use the verification set to test the training results

Do you think that one convergence can solve all problems ? It's not that simple .

Deep learning often encounters a problem —— Over fitting . When training, the learning effect is very good , But the test results in other places are not good .

That is to say , It doesn't have to be the best time to stop learning . So how to judge the timing of stopping ?

Some scholars put forward the verification set .

That is to say , Divide the training set into 2 part , such as 70% For training ,30% Used to verify . Like the following code .

X_train, X_test, y_train, y_test = train_test_split(X, Y, test_size=0.33, random_state=seed)

Then add the validated parameters during the training , Just like this. ：

autoencoder.fit(train_data, train_data,

epochs=50,

batch_size=128,

shuffle=True,

validation_data=(noisy_imgs, data_test)

)

Then it is to observe the validation curve , When was it verified loss Minimum value , Select the training model for test and application .

For example, below , Above is the training curve , Here is the validation curve . Study 15 The effect is already better .

03 Use the test set to verify the training results

With the extensive development of research and the ultimate pursuit of learning effect , We gradually discovered a problem —— Sometimes the optimal value seen from the validation curve is not the best result of the real prediction process . For example, in this case ： Deep learning denoising for large-scale data .

Say first conclusion ： The best method is to determine the number of iterations through the test results .

How did you do it ? For example, a case like this ： A large two-dimensional data , such as 300*300, Like this, ：

Share some of my artificial intelligence learning materials for free , Including some AI Common framework actual combat video 、 Image recognition 、OpenCV、NLQ、 machine learning 、pytorch、 Computer vision 、 Videos such as deep learning and neural network 、 Courseware source code 、 Famous essence resources at home and abroad 、AI Hot papers 、 Industry reports, etc .

For better systematic learning AI, I recommend that you collect one .

Here are some screenshots , Free download method is attached at the end of the article .

One 、 Artificial intelligence courses and projects

Two 、 Famous essence resources at home and abroad

3、 ... and 、 Collection of papers on artificial intelligence

Four 、 AI Industry Report

Learn Artificial Intelligence well , Read more , Do more , practice , If you want to improve your level , We must learn to settle down and learn slowly and systematically , Only in the end can we gain something .

Click on the business card below , Scan the code and download the information for free .