Multi-label Contrastive Predictive Coding

Authors: Jiaming Song, Stefano Ermon

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

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our proposed methods on mutual information estimation, knowledge distillation and unsupervised representation learning.
Researcher Affiliation Academia Jiaming Song Stanford University tsong@cs.stanford.edu Stefano Ermon Stanford University ermon@cs.stanford.edu
Pseudocode Yes We include the pseudo-code and a PyTorch implementation to α-ML-CPC in Appendix B.
Open Source Code Yes Our code is available at https://github.com/jiamings/ml-cpc.
Open Datasets Yes Following the procedure in [47], we evaluate over 13 different student-teacher pairs on CIFAR-100 [30].
Dataset Splits No The paper uses standard datasets like CIFAR-10, CIFAR-100, and ImageNet, which have predefined splits. However, it does not explicitly state the percentages for training, validation, or test splits, nor does it explicitly mention the use of a separate validation set in its experimental setup.
Hardware Specification Yes The time to compute 200 updates on a Nvidia 1080 Ti GPU with the a PyTorch implementation is 1.15 0.06 seconds with CPC and 1.14 0.04 seconds with ML-CPC
Software Dependencies No The paper mentions 'PyTorch implementation' but does not specify version numbers for PyTorch or any other software dependencies, making it difficult to fully reproduce the software environment.
Experiment Setup Yes We use the same values for other hyperparameters as those used in the Mo Co-v2 CPC baseline (more details in Appendix C). For MoCo-v2, we used a batch size of 256, a learning rate of 0.03, cosine decay scheduler for 200 epochs, and a temperature of 0.07.