BMRS: Bayesian Model Reduction for Structured Pruning

Authors: Dustin Wright, Christian Igel, Raghavendra Selvan

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on multiple datasets and neural networks of varying complexity showed that the two BMRS methods offer a competitive performance-efficiency trade-off compared to other pruning methods.
Researcher Affiliation Academia Dustin Wright, Christian Igel, and Raghavendra Selvan Department of Computer Science, University of Copenhagen {dw,igel,raghav}@di.ku.dk
Pseudocode Yes Algorithm 1: Training and pruning with BMRS
Open Source Code Yes Source code: https://github.com/saintslab/bmrs-structured-pruning/
Open Datasets Yes We use the following datasets (full details in Appendix B): MNIST [22], Fashion-MNIST [37], CIFAR10 [21], and Tiny Imagenet.
Dataset Splits Yes We use the original 10,000 image test set for testing and split the 60,000 image train set into 80% training and 20% validation images.
Hardware Specification Yes All experiments were run on a shared cluster. Requested jobs consisted of 16GB of RAM and 4 Intel Xeon Silver 4110 CPUs. We used a single NVIDIA Titan X GPU with 24GB of RAM for all experiments
Software Dependencies No The paper mentions the use of the Adam optimizer and implicit use of PyTorch for neural network training but does not provide specific version numbers for these or other software libraries.
Experiment Setup Yes For each model and dataset we use the Adam optimizer with no weight decay. We train for 50 epochs for each experiment with an MLP and Lenet5, and for 100 epochs for each experiment with Resnet50 and Vi T. Further details about each model are given as follows: MNIST: Number of layers: 7; Hidden dimension: 100; Batch size: 128; Learning rate: 8.5 10 4.