On the Convergence of Local Stochastic Compositional Gradient Descent with Momentum

Authors: Hongchang Gao, Junyi Li, Heng Huang

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Additionally, extensive experimental results demonstrate the superior empirical performance over existing methods, confirming the efficacy of our method. 5. Numerical Experiments
Researcher Affiliation Academia 1Department of Computer and Information Sciences, Temple University, PA, USA. 2Department of Electrical and Computer Engineering, University of Pittsburgh, PA, USA.
Pseudocode Yes Algorithm 1 Local-SCGDM
Open Source Code No The paper does not provide any statement about making its source code publicly available or a link to a code repository.
Open Datasets Yes We evaluate our proposed algorithm Local SCGDM over a 1-D sinusoid regression problem... and the Few-Shot Classification task over the Omniglot dataset.
Dataset Splits Yes we construct 25 different training tasks by choosing A = {1, 2, 3, 4, 5} and b = {1, 2, 3, 4, 5} and randomly and evenly distribute them over 5 clients. Then during training, we randomly sample 3 tasks for every client per meta-iteration. For each task we choose K = 10 samples of x [ 5, 5] randomly. We follow the experimental protocols of Vinyals et al. (2016) to divide the alphabets to train/validation/test with 33/5/12, respectively. for each task, we sample K samples for training and 15 samples for validation.
Hardware Specification Yes All experiments are run over a machine with Intel Xeon Gold 6248 CPU and 4 Nvidia Tesla V100 GPUs.
Software Dependencies No The paper mentions "Pytorch" and "Pytorch.distributed package" but does not specify version numbers for these software components.
Experiment Setup Yes The inner learning rate is 0.001 for all methods. For other hyper-parameters, we perform grid search for all methods and choose the setting with the best results. More precisely, for Local-BSGD (Local-MAML), we choose meta learning rate 0.01; for Local-SCGD, we choose meta learning rate 0.01 and the inner state momentum coefficient 0.9 (this algorithm diverges with smaller values); for Local-MOML, we choose meta learning rate 0.01, inner state momentum coefficient 0.7; for our Local-SCGDM, we choose η as 1, meta learning rate coefficient β as 0.01, meta momentum coefficient α as 0.8 and inner state momentum coefficient γ as 0.7. We set the number of local epochs as 5 in all comparison experiments.