Equi-normalization of Neural Networks

Authors: Pierre Stock, Benjamin Graham, Rémi Gribonval, Hervé Jégou

ICLR 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We analyze our approach by carrying out experiments on the MNIST and CIFAR-10 datasets and on the more challenging Image Net dataset.
Researcher Affiliation Collaboration Pierre Stock1,2, Benjamin Graham1, R emi Gribonval2 and Herv e J egou1 1Facebook AI Research 2Univ Rennes, Inria, CNRS, IRISA
Pseudocode Yes Algorithm 1: Pseudo-code of Equi-normalization
Open Source Code Yes Our code is available at https://github.com/facebookresearch/enorm.
Open Datasets Yes We analyze our approach by carrying out experiments on the MNIST and CIFAR-10 datasets and on the more challenging Image Net dataset.
Dataset Splits Yes We split the train set into one training set (40,000 images) and one validation set (10,000 images).
Hardware Specification No The paper mentions 'We train on 8 GPUs' but does not specify the model or type of the GPUs (e.g., NVIDIA A100, Tesla V100), nor does it provide details about CPUs or memory.
Software Dependencies Yes Although the additional overhead of implicit ENorm is theoretically negligible, we observed an increase of the training time of a Res Net-18 by roughly 30% using Py Torch 4.0 (Paszke et al., 2017).
Experiment Setup Yes We select the learning rate in {0.001, 0.01, 0.1} and decay it linearly to zero. We use a batch size of 256 and SGD with no momentum and a weight decay of 0.001.