Exponentional/Proportional Derivative (E/PD) and Double Event-Based E/PD learning rate algorithms

Code for paper Feedback Control for Online Training of Neural Networks and Event-Based Control for Online Training of Neural Networks which published in 2019 IEEE Conference on Control Technology and Applications (CCTA) and The IEEE Control Systems Letters.

The basic code structure is designed for google cloud ML-engine platform, the code is based on Python 3.5, Tenserflow and Keras.

Feedback Control for Online Training of Neural Networks (E/PD)

This paper introduces a Exponentional/Proportional Derivative (E/PD) learning rate algorithm, which dynamically adopt learning rate based on the training/testing loss. The algorithm is defined in trainer/EPD.py. One demo for testing E/PD algorith is provided in trainer/task-EPD.py, if user does not want to run it on google cloud, one can run direclty:

python task-EPD-local.py

This code will run E/PD on CIFAR10, the CIFAR10 dataset will be chunked into 5 subsets, each subset will be trained 60 epochs.

Event-Based Control for Online Training of Neural Networks (Double Event-Based E/PD)

This paper introduces two algorithms:

Event-Based Learning Rate (EBLR)
Event-Based Learning Epoch (EBLE)

Code for EBLR is defined in trainer/EPD.py, there is a parameter eventbasedLR, it will turn on/off this event-based control by setting the parameter True/False. Code for EBLE is defined in trainer/EventBasedLE.py. One demo for testing these two event-based control is provided in trainer/task-D-EB-EPD.py, if user does not want to run it on google cloud, one can run direclty:

python task-D-EB-EPD-local.py

This code will run E/PD and (EBLR & EBLE) on CIFAR10, the CIFAR10 dataset will be chunked into 5 subsets, maximum successive training epochs for one subset is 60, maximum total training epoch sets to 300.

Usage

Import the customized callbacks:

from .EPD import LossLearningRateScheduler
from .EventBasedLE import EventBasedLearningEpochStopper

As our algorithms are all based on SGD, so the optimizer must be defined as (value of parameter lr here does not matter, because we will adjust its value before the training, setting by our algorithm):

optimizer = SGD(lr=0.01, decay=0, momentum=0)

To use the callbacks, we can call it as that:

h  =   model.fit_generator(datagen.flow(X_clean_iteration, y_clean_iteration, batch_size=batch_size),
                    steps_per_epoch=X_clean_iteration.shape[0]//batch_size, epochs=epochs_training,
                    validation_data=(X_test, y_test), 
                    callbacks = [LossLearningRateScheduler(base_lr = 0.01, kd_coef = 5, eventbasedLR = True), 
                    EventBasedLearningEpochStopper(alpha_threshold = -0.001, loss_zero = 0, lookbackward_epoch = 5, accumulated_epoch = 0, limit_epoch = total_epoch_training, ending_epoch = 10)]
                    )

One thing to notice that for parameter loss_zero in EventBasedLearningEpochStopper, if this is the first to run the training process, as we do not have loss_zero (the loss value of first training epoch), we must set it to 0. Once we have the loss_zero value, we will pass this value when we train the model.

Citing

If you use E/PD in your research, please cite:

@INPROCEEDINGS{zhao2019epd, 
author={Z. {Zhao} and S. {Cerf} and B. {Robu} and N. {Marchand}}, 
booktitle={2019 IEEE Conference on Control Technology and Applications (CCTA)}, 
title={Feedback Control for Online Training of Neural Networks}, 
year={2019}, 
pages={136-141}, 
doi={10.1109/CCTA.2019.8920662}, 
ISSN={null}, 
month={Aug},}

If you use EBLR or EBLE in your research, please cite:

@ARTICLE{zhao2020lcss, 
author={Z. {Zhao} and S. {Cerf} and B. {Robu} and N. {Marchand}}, 
journal={IEEE Control Systems Letters}, 
title={Event-Based Control for Online Training of Neural Networks}, 
year={2020}, 
doi={10.1109/LCSYS.2020.2981984}, 
ISSN={2475-1456}, 
}

License