Generative Models as a Data Source for Multiview Representation Learning
Authors: Ali Jahanian, Xavier Puig, Yonglong Tian, Phillip Isola
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 4 EXPERIMENTS To study the methods in Sec. 3, we analyze the behaviour and effectiveness of the representations learned from unconditional IGMs here and, in the interest of space, report the results on classconditional IGMs in App. C. We experiment on two unconditional IGMs: the generator from Big Bi GAN (Donahue & Simonyan, 2019) and the Style GAN2 LSUN CAR generator (Karras et al., 2020). |
| Researcher Affiliation | Academia | Ali Jahanian, Xavier Puig, Yonglong Tian, Phillip Isola Massachusetts Institute of Technology Cambridge, MA 02139, USA {jahanian,xpuig,yonglong,phillipi}@mit.edu |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code is available on our project page https://ali-design.github.io/GenRep/. |
| Open Datasets | Yes | Image Net1000 (Deng et al., 2009), Pascal VOC2007 dataset (Everingham et al., 2010), LSUN CAR dataset (recreated via Style GAN2 repository for public access). |
| Dataset Splits | Yes | Stanford Car classification task (Krause et al., 2013) (196 car models with roughly 8K train and 8K val) |
| Hardware Specification | No | The paper does not provide any specific hardware details such as GPU models, CPU types, or cloud computing instances used for running the experiments. |
| Software Dependencies | No | The paper does not specify version numbers for any software dependencies or libraries used in the implementation. |
| Experiment Setup | Yes | In the Image Net1000 setting, the real data encoders are trained on the Image Net1000 dataset, the IGM encoders are trained on 1.3M anchor images (which roughly matches the size of Image Net1000). We use SGD with learning rate of 0.03, batch size of 256, and 20 epochs. In the Imaget Net100 setting, the real data encoders are trained on Image Net100, the unconditional IGM encoders are trained on 130K anchor images sampled unconditionally (note that the unconditional model implicitly is still sampling from all 1000 classes since the generative model itself was fit to Image Net1000). We use SGD with learning rate of 0.03, batch size of 256, for 200 epochs. |