Instance-Conditioned GAN
Authors: Arantxa Casanova, Marlene Careil, Jakob Verbeek, Michal Drozdzal, Adriana Romero Soriano
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on Image Net and COCO-Stuff show that IC-GAN significantly improves over unconditional models and unsupervised data partitioning baselines. |
| Researcher Affiliation | Collaboration | Arantxa Casanova Facebook AI Research École Polytechnique de Montréal Mila, Quebec AI Institute Marlène Careil Facebook AI Research Télécom Paris Jakob Verbeek Facebook AI Research Michał Dro zd zal Facebook AI Research Adriana Romero-Soriano Facebook AI Research Mc Gill University |
| Pseudocode | No | The paper describes the Instance-conditioned GAN method in Section 2 but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code and trained models to reproduce the reported results are available at https://github.com/facebookresearch/ic_gan. |
| Open Datasets | Yes | We evaluate our model in the unlabeled scenario on Image Net [45] and COCO-Stuff [6]. The Image Net dataset contains 1.2M and 50k images for training and evaluation, respectively. COCO-Stuff is a very diverse and complex dataset which contains multi-object images and has been widely used for complex scene generation. We use the train and evaluation splits of [8], and the (un)seen subsets of the evaluation images with only class combinations that have (not) been seen during training. These splits contain 76k, 2k, 675 and 1.3k images, respectively. For the class-conditional image generation, we use Image Net as well as Image Net-LT [34]. The latter is a long-tail variant of Image Net that contains a subset of 115k samples, where the 1,000 classes have between 5 and 1,280 samples each. Moreover, we use some samples of four additional datasets to highlight the transfer abilities of IC-GAN: Cityscapes [10], Met Faces [28], PACS [31] and Sketches [15]. |
| Dataset Splits | Yes | The Image Net dataset contains 1.2M and 50k images for training and evaluation, respectively. |
| Hardware Specification | No | The paper mentions various network architectures like ResNet50, BigGAN, and StyleGAN2, but it does not specify any hardware details such as GPU or CPU models, or memory configurations used for running the experiments. |
| Software Dependencies | No | The paper mentions using specific models like 'Res Net50' and 'Big GAN' as backbone architectures but does not provide version numbers for any software dependencies. |
| Experiment Setup | Yes | Unless stated otherwise, we set the size of the neighborhoods to k =50 for Image Net and k =5 for both COCO-Stuff and Image Net-LT. See the supplementary material for details on the architecture and optimization hyperparameters. |