Channel Equilibrium Networks for Learning Deep Representation
Authors: Wenqi Shao, Shitao Tang, Xingang Pan, Ping Tan, Xiaogang Wang, Ping Luo
ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We extensively evaluate the proposed CE on two basic vision tasks, image classification on Image Net (Russakovsky et al., 2015) and object detection/segmentation on COCO (Lin et al., 2014). We first evaluate CE on the Image Net benchmark. The models are trained on the 1.28M training images and evaluate on the 50,000 validation images. The top-1 and top5 accuracies are reported. We are particularly interested in whether the proposed CE has better generalization to testing samples in various modern CNNs such as Res Nets (He et al., 2016), Mobile Netv2 (Sandler et al., 2018), Shuffle Netv2 (Ma et al., 2018) compared with the SE block (Hu et al., 2018). |
| Researcher Affiliation | Collaboration | 1Department of Electronic Engineering, The Chinese University of Hong Kong 2Department of Computer Science, Simon Fraser University 3Department of Information Engineering, The Chinese University of Hong Kong 4Department of Computer Science,The University of Hong Kong. Correspondence to: Wenqi Shao <weqish@link.cuhk.edu.hk>, Ping Luo <pluo@cs.hku.hk>. |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Models and code are available at https://github.com/Tangshitao/CENet. |
| Open Datasets | Yes | We extensively evaluate the proposed CE on two basic vision tasks, image classification on Image Net (Russakovsky et al., 2015) and object detection/segmentation on COCO (Lin et al., 2014). |
| Dataset Splits | Yes | We extensively evaluate the proposed CE on two basic vision tasks, image classification on Image Net (Russakovsky et al., 2015) and object detection/segmentation on COCO (Lin et al., 2014). We train our model on the union of 80k training images and 35k validation images and report the performance on the mini-val 5k images. |
| Hardware Specification | Yes | Additionally, the CPU and GPU inference time of CENet is nearly the same with SENet. 1The CPU type is Intel Xeon CPU E5-2682 v4, and the GPU is NVIDIA GTX1080TI. The implementation is based on Pytorch |
| Software Dependencies | Yes | 1The CPU type is Intel Xeon CPU E5-2682 v4, and the GPU is NVIDIA GTX1080TI. The implementation is based on Pytorch |
| Experiment Setup | Yes | The training details are illustrated in Sec.G of the Appendix. For fair comparisons, we use publicly available code and reimplement baseline models and SE modules with their respective best settings in a unified Pytorch framework. To save computation, the CE blocks are selectively inserted into the last normalization layer of each residual block. Specifically, for Res Net18, we plug the CE block into each residual block. For Res Net50, CE is inserted into all residual blocks except for those layers with 2048 channels. For Res Net101, the CE blocks are employed in the first seven residual blocks. |