Delay-agnostic Asynchronous Coordinate Update Algorithm

Authors: Xuyang Wu, Changxin Liu, Sindri Magnússon, Mikael Johansson

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The effectiveness of DEGAS is demonstrated by numerical experiments on classification problems. We evaluate the practical performance of DEGAS on Lasso and regularized logistic regression problems on the CIFAR-100 dataset (Krizhevsky et al., 2009). We plot the convergence in terms of the number of computed Ti in Figures 2 3
Researcher Affiliation Academia 1Division of Decision and Control Systems, EECS, KTH Royal Institute of Technology, Stockholm, Sweden 2Department of Computer and System Science, Stockholm University, Stockholm, Sweden.
Pseudocode Yes Algorithm 1 DEGAS 1: Setup: initial iterate x(0). 2: Initialization: the master sets x = x(0) and broadcasts x to all workers. 3: while not interrupted by master: each worker w [n] asynchronously and continuously do 4: receive x from the master and assign xw = x. 5: sample i [m] uniformly at random. 6: compute Ti(xw). 7: send (Ti(xw), i) to the master. 8: end while 9: while not converged: the master do 10: receive (Ti(xw), i) from a worker w. 11: update xi Ti(xw). 12: send x to the worker w. 13: end while
Open Source Code No The paper does not provide a direct link to a source code repository or an explicit statement of code release for the methodology described.
Open Datasets Yes We evaluate the practical performance of DEGAS on Lasso and regularized logistic regression problems on the CIFAR-100 dataset (Krizhevsky et al., 2009).
Dataset Splits No The paper mentions using the CIFAR-100 dataset but does not explicitly provide details about specific train/validation/test splits, such as percentages or sample counts.
Hardware Specification Yes We set m = 20 and implement all the methods on a 10-core machine (1 master and 9 workers)
Software Dependencies No The paper mentions 'MPI4py (Dalcın et al., 2008)' but does not provide specific version numbers for MPI4py or any other software dependencies, which are required for reproducibility.
Experiment Setup Yes In these methods, we choose the operator T as (20) with γ = 1/L in BCD and (29) in the extended ADMM. We set m = 20... We consider both theoretical and hand-tuned parameters. In the former setting, we fine-tune the step-size of ARock within its theoretical range... while the other two methods have no parameters to tune. In the hand-tune step-size setting, we run all the methods for finding the fixed point of Id +λ(T Id), λ > 0 and tune λ.