Supervised Contrastive Learning

Authors: Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, Dilip Krishnan

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

Reproducibility Variable Result LLM Response
Research Type Experimental On Res Net-200, we achieve top-1 accuracy of 81.4% on the Image Net dataset, which is 0.8% above the best number reported for this architecture. We show consistent outperformance over cross-entropy on other datasets and two Res Net variants.
Researcher Affiliation Collaboration Prannay Khosla Google Research Piotr Teterwak Boston University Chen Wang Snap Inc. Aaron Sarna Google Research Yonglong Tian MIT Phillip Isola MIT Aaron Maschinot Google Research Ce Liu Google Research Dilip Krishnan Google Research
Pseudocode No The paper describes the contrastive loss functions with mathematical equations and describes the representation learning framework components, but it does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes Our loss function is simple to implement and reference Tensor Flow code is released at https://t.ly/supcon 1. Py Torch implementation: https://github.com/Hobbit Long/Sup Contrast
Open Datasets Yes We evaluate our Sup Con loss (Lsup out, Eq. 2) by measuring classification accuracy on a number of common image classification benchmarks including CIFAR-10 and CIFAR-100 [27] and Image Net [7].
Dataset Splits Yes We evaluate our Sup Con loss (Lsup out, Eq. 2) by measuring classification accuracy on a number of common image classification benchmarks including CIFAR-10 and CIFAR-100 [27] and Image Net [7].
Hardware Specification Yes On Image Net, with a memory size of 8192 (requiring only the storage of 128-dimensional vectors), a batch size of 256, and SGD optimizer, running on 8 Nvidia V100 GPUs, Sup Con is able to achieve 79.1% top-1 accuracy on Res Net-50.
Software Dependencies No The paper mentions 'reference Tensor Flow code' and a 'Py Torch implementation' but does not specify version numbers for these software dependencies or any other libraries.
Experiment Setup Yes The Sup Con loss was trained for 700 epochs during pretraining for Res Net-200 and 350 epochs for smaller models. We trained our models with batch sizes of up to 6144... All our results used a temperature of τ = 0.1. We experimented with standard optimizers such as LARS [58], RMSProp [20] and SGD with momentum [39]...