This is the implementation of AdaptiveLRF [1]. This code was written based on keras/tensorflow and tested on tensorflow 1.15.0.
The code contains just a class named AdaptiveLRF. This class extends Callback (a class in keras library).
Then, what just you need to do is add a instance of this class in your callback list in fit function.
The constructor parameter is in the following:
- A trainable model. The acceptable type is
keras.models.Model. - The number of the classes. The acceptable type is
int. - A part of the input of the train samples. The acceptable type is
numpy.ndarray. - A part of the output of the train samples. The acceptable type is
numpy.ndarray. - The batch size which the conditional analysis is done on this size of the batch. The acceptable type is
int. - The value of
k. This value was described in the paper. - The number of patients after overfitting.
model.compile(...)
reg = AdaptiveLRF(model, 10, ...)
model.fit(..., callbacks=[reg])
Notice, you have to give the validation or test data to the fit function!
[1] Mohammad Mahdi Bejani and Mehdi Ghatee, Adaptive Low-Randk Factorization to Regularize Shallow and Deep Neural Networks, ArXiv, 2020