Lookahead: A Far-sighted Alternative of Magnitude-based Pruning

Authors: Sejun Park*, Jaeho Lee*, Sangwoo Mo, Jinwoo Shin

ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experimental results demonstrate that the proposed method consistently outperforms magnitude-based pruning on various networks, including VGG and Res Net, particularly in the high-sparsity regime.
Researcher Affiliation Academia KAIST EE KAIST AI {sejun.park,jaeho-lee,swmo,jinwoos}@kaist.ac.kr
Pseudocode Yes Algorithm 1 Lookahead Pruning (LAP)
Open Source Code Yes See https://github.com/alinlab/lookahead_pruning for codes.
Open Datasets Yes FCN is trained on MNIST dataset (Lecun et al., 1998), Conv-6, VGG, and Res Net are trained on CIFAR-10 dataset (Krizhevsky & Hinton, 2009), and VGG, Res Net, and WRN are trained on Tiny-Image Net.
Dataset Splits No The paper mentions 'test accuracy' and 'retraining phase' but does not explicitly provide information on training/test/validation splits with percentages, sample counts, or citations to predefined validation splits.
Hardware Specification Yes The computations have taken place on 40 CPUs of Intel Xeon E5-2630v4 @ 2.20GHz. ...additionally used a single NVidia Ge Force GTX1080
Software Dependencies No The paper mentions 'Py Torch' and 'Back PACK' but does not provide specific version numbers for these software dependencies. It mentions 'Adam optimizer' but this is an algorithm, not a software dependency with a version.
Experiment Setup Yes We use Adam optimizer (Kingma & Ba, 2015) with batch size 60. We use a learning rate of 1.2 10 3 for FCN and 3 10 4 for all other models. For FCN, we use [50k, 50k] for the initial training phase and retraining phase. For Conv-6, we use [30k, 20k] steps.