Un-mix: Rethinking Image Mixtures for Unsupervised Visual Representation Learning
Authors: Zhiqiang Shen, Zechun Liu, Zhuang Liu, Marios Savvides, Trevor Darrell, Eric Xing2216-2224
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments are conducted on CIFAR-10, CIFAR-100, STL-10, Tiny Image Net and standard Image Net-1K with popular unsupervised methods Sim CLR, BYOL, Mo Co V1&V2, Sw AV, etc. |
| Researcher Affiliation | Collaboration | 1 Carnegie Mellon University 2 Reality Labs, Meta Inc. 3 University of California, Berkeley 4 Mohamed bin Zayed University of Artificial Intelligence |
| Pseudocode | No | The paper does not contain any pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | Yes | Code is publicly available at https://github.com/szq0214/Un-Mix. |
| Open Datasets | Yes | Extensive experiments are conducted on CIFAR-10, CIFAR-100, STL-10, Tiny Image Net and standard Image Net-1K. [...] CIFAR-10/100 [Krizhevsky and Hinton 2009] consist of tiny colored natural images [...]. Image Net-1K [Deng et al. 2009], aka ILSVRC 2012 classification dataset consists of 1000 classes, with a number of 1.28 million training images and 50K validation images. |
| Dataset Splits | Yes | Image Net-1K [Deng et al. 2009], aka ILSVRC 2012 classification dataset consists of 1000 classes, with a number of 1.28 million training images and 50K validation images. |
| Hardware Specification | Yes | For example, we use a mini-batch size of 256 with 8 NVIDIA V100 GPUs on Image Net-1K |
| Software Dependencies | No | The paper mentions implementing the method with 'Py Torch codes' but does not specify a version number for PyTorch or any other software dependencies. |
| Experiment Setup | Yes | on CIFAR-10 and CIFAR-100, we train for 1,000 epochs with learning rate 3 10 3; on Tiny Image Net, 1,000 epochs with learning rate 2 10 3; on STL-10, 2,000 epochs with learning rate 2 10 3. We also apply warm-up for the first 500 iterations, and a 0.2 learning rate drop at 50 and 25 epochs before the end. [...] Unless otherwise stated, all the hyperparameter configurations strictly follow the baseline Mo Co V2 on Image Net-1K. For example, we use a mini-batch size of 256 |