I'm studying Python on my own.
With Keras code,
model.fit(x_train,y_train,verbose=1,validation_split=0.2,shuffle=True,epochs=20000)
At that time, we understand that learning occurs over 20,000 epochs, with 80% of the data being used for learning and 20% for validation.
shuffle=True
Therefore, the selection of validation data becomes random, and I think this code will be cross-validation, or k=5 divided cross-validation. Is that correct?
When I looked into Keras' code for k-division cross verification, I was worried about whether it was correct or not, so I asked you a question.
I apologize for the rudimentary question, but I would appreciate it if you could let me know.Thank you for your cooperation.
python deep-learning keras
Not cross-validation.
validation_split is an argument to truncate the specified percentage of data for validation before shuffling the data set.Also, validation data takes a specified percentage from the end of the data set.Therefore, all epochs use the same data for validation, and the extraction method is not random.
shuffle is an argument that shuffles a data set for learning.These arguments are provided because 3,4 are known to reduce loss by shuffling learning data per epoch.
Keras documentation describes the arguments for model.fit as follows:
validation_split: floating point number between 0 and 1. percentage of training data used as validation data. [omitted] This validation data is sampled from the back of the given x and y data before shuffle.shuffle—Truth value (whether to shuffle training data before each attempt [epoc]) or string (for 'batch').[abbreviated]Keras documentation (https://keras.io/ja/models/model/), last accessed date: 2022-01-16.Square brackets and highlights are appended by respondents.
Keras does not provide K-split cross-validationAs you pointed out, you need to use another library, such as scikit-learn.
1The relevant implementation is the definition of model.fit (GitHub).validation_split as the argument is performed before applying shuffle.
2Implementation is similar to validation_split (GitHub).
3Examples of experiments with or without shuffles: https://qiita.com/hikobotch/items/d8ff5bebcf70083de089
4Description of why the loss is reduced: https://stats.stackexchange.com/a/311318
model.fit (GitHub). You can see that validation_split was used as an argument before applying shuffle.2Implementation is similar to validation_split (GitHub).3Examples of experiments with or without shuffles: https://qiita.com/hikobotch/items/d8ff5bebcf70083de0894Description of why the loss is reduced: https://stats.stackexchange.com/a/311318
© 2024 OneMinuteCode. All rights reserved.