Aux-NAS: Exploiting Auxiliary Labels with Negligibly Extra Inference Cost
Authors: Yuan Gao, WEIZHONG ZHANG, Wenhan Luo, Lin Ma, Jin-Gang Yu, Gui-Song Xia, Jiayi Ma
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments with six tasks on NYU v2, City Scapes, and Taskonomy datasets using VGG, Res Net, and Vi T backbones validate the promising performance. |
| Researcher Affiliation | Collaboration | Yuan Gao1, Weizhong Zhang2, Wenhan Luo3, Lin Ma4, Jin-Gang Yu5, Gui-Song Xia1 , Jiayi Ma1 1Wuhan University, 2Fudan University, 3HKUST, 4Meituan, 5South China University of Technology |
| Pseudocode | No | The paper does not contain any explicitly labeled 'Pseudocode' or 'Algorithm' blocks, nor does it present structured steps formatted like code. |
| Open Source Code | Yes | The codes are available at https://github.com/ethanygao/Aux-NAS. |
| Open Datasets | Yes | We perform our experiments on the NYU v2 Silberman et al. (2012), the City Scapes Cordts et al. (2016), and the Taskonomy Zamir et al. (2018) datasets. |
| Dataset Splits | Yes | We use the official train/val split for the NYU v2 Eigen & Fergus (2015) and the City Scapes Cordts et al. (2016) datasets. |
| Hardware Specification | No | The paper mentions using VGG-16, Res Net-50, and Vi TBase backbones, but it does not specify any details about the hardware (e.g., GPU models, CPU types) used for running the experiments. |
| Software Dependencies | No | The paper mentions using the 'huggingface timm package' for Vi TBase, but it does not provide specific version numbers for this or any other software dependencies, such as the deep learning framework used (e.g., PyTorch, TensorFlow). |
| Experiment Setup | Yes | We use 321 321 image samples for the CNN backbones (i.e., Res Net-50 and VGG-16), and 224 224 image samples for the transformer backbone (i.e., Vi TBase Dosovitskiy et al. (2021))... We initialize the single task branches with the pretrained single task model weights. For the fusion operations, we initialize the 1x1 convolution with all 0. We gradually increase λ of Eq. 10 from 0 to 100 during the training, and initialize all α s of Eqs. 13 and 14 to 0.5. |