AADL is a Python package that implements the Anderson acceleration to speed-up the training of deep learning (DL) models using the PyTorch library.\ AA is an extrapolation technique that can accelerate fixed-point iterations such those arising from the iterative training of DL models. However, large volume of data are typically processed in sequential random batches which introduces stochastic oscillations in the fixed-point iteration that hinders AA acceleration. AADL implements a moving average that reduces the oscillations and results in a smoother sequence of gradient descent updates which enables the use of AA. AADL uses a criterion to automatically decide if the moving average is needed by monitoring if the relative standard deviation between consecutive stochastic gradient updates exceeds a tolerance defined by the user.
Python 3.5 or greater\ PyTorch (any version works)
AADL comes with a setuptools install script:
python3 setup.py installimport torch
import torch.nn
import torch.optim
import AADL
# Creation of the DL model (neural network)
class model(torch.nn.Module):
...
# Definition of the stochastic optimizer used to train the model
optimizer = torch.optim.SGD(model.parameters(), lr=1e-3, momentum=0.9, nesterov = True)
# Parameters for Anderson acceleration
relaxation = 0.5
wait_iterations = 0
history_depth = 10
store_each_nth = 10
frequency = store_each_nth
reg_acc = 0.0
safeguard = True
average = True
# Over-writing of the torch.optim.step() method
AADL.accelerate(optimizer_anderson, "anderson", relaxation, wait_iterations, history_depth, store_each_nth, frequency, reg_acc, average)
relaxation: Float. Linear mixing parameter between a standard gradient descent update and the Anderson update\
wait_iterations: Integer. Number of initial gradient descent updates to wait before starting the Anderson scheme\
history_depth: Integer. Number of gradient updates used to compute the Anderson mixing. The history is updated with a first-in-first-out policy\
store_each_nth: Integer. Number of gradient updates to skip between two vector updates consecutively stored in the history window\
frequency: Integer. Number of gradient updates to skip between two consecutive Anderson steps\
reg_acc: Float. Tikhonov regularization factor used to stabilize the least-squares problem solved to compute the Anderson mixing vector\
safeguard: Boolean. If set to True, the Anderson step overwrites the gradient step only if the training loss function is further reduced\
average: Boolean. If set to True, a movign average is applied to the history window before computing the Anderson step\
Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.
M. Lupo Pasini, V. Reshniak, and M. K. Stoyanov. AADL: Anderson Accelerated Deep Learning. Computer Software. https://github.com/ORNL/AADL.git. 06 Sep. 2021. Web. doi:10.11578/dc.20210723.1. Copyright ID#: 81927550
M. Lupo Pasini, J. Yin, V. Reshniak and M. K. Stoyanov, "Anderson Acceleration for Distributed Training of Deep Learning Models," SoutheastCon 2022, 2022, pp. 289-295, doi: 10.1109/SoutheastCon48659.2022.9763953.