Improved Feature Distillation via Projector Ensemble

Authors: Yudong Chen, Sen Wang, Jiajun Liu, Xuwei Xu, Frank de Hoog, Zi Huang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on different datasets with a series of teacher-student pairs illustrate the effectiveness of the proposed method.
Researcher Affiliation Academia 1The University of Queensland 2CSIRO Data61 {yudong.chen,sen.wang,xuwei.xu}@uq.edu.au {jiajun.liu,frank.dehoog}@csiro.au huang@itee.uq.edu.au
Pseudocode Yes Algorithm 1 Improved Feature Distillation via Projector Ensemble.
Open Source Code Yes Code is available at https://github.com/chenyd7/PEFD.
Open Datasets Yes Datasets. Two benchmark datasets are used for evaluation in our experiments. Image Net [25] contains approximately 1.28 million training images and 50,000 validation images from 1,000 classes. ... CIFAR-100 [18] dataset includes 50,000 training images and 10,000 testing images from 100 classes.
Dataset Splits Yes Image Net [25] contains approximately 1.28 million training images and 50,000 validation images from 1,000 classes. The validation images are used for testing. CIFAR-100 [18] dataset includes 50,000 training images and 10,000 testing images from 100 classes.
Hardware Specification Yes All the experiments are performed on an NVIDIA V100 GPU.
Software Dependencies No The paper mentions software like 'PyTorch' implicitly through a GitHub link for settings, but does not provide specific version numbers for any software dependencies or libraries.
Experiment Setup Yes Following the settings of previous methods2, the batch size, epochs, learning rate decay rate and weight decay rate are 256/64, 100/240, 0.1/0.1, and 0.0001/0.0005, respectively on Image Net/CIFAR-100. The initial learning rate is 0.1 on Image Net, and 0.01 for Mobile Net V2, 0.05 for the other students on CIFAR-100. Besides, the learning rate drops at every 30 epochs on Image Net and drops at 150, 180, 210 epochs on CIFAR-100. The optimizer is Stochastic Gradient Descent (SGD) with momentum 0.9.