Designing Concise ConvNets with Columnar Stages

Authors: Ashish Kumar, Jaesik Park

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our comprehensive evaluations show that Co SNet rivals many renowned Conv Nets and Transformer designs under resource-constrained scenarios. Code: https://github.com/ashishkumar822/Co SNet.We evaluate Co SNet on Image Net (Deng et al., 2009) dataset consisting of 1.28M train and 50k validation images of 1000 categories.
Researcher Affiliation Collaboration Ashish Kumar Score Labs AI Atlanta, USA EMAIL Jaesik Park Seoul National University Seoul, South Korea EMAIL
Pseudocode Yes F PYTORCH CODE All codes shall be open-sourced in Py Torch Paszke et al. (2019) post the review process. Here, we provide a code snippet of a Co SNet-Unit. Please see until the end of this document.
Open Source Code Yes Code: https://github.com/ashishkumar822/Co SNet.
Open Datasets Yes We evaluate Co SNet on Image Net (Deng et al., 2009) dataset consisting of 1.28M train and 50k validation images of 1000 categories. Our training methodology is consistent with recent Vanilla Net (Chen et al., 2023). We apply Co SNet to state-of-the-art object Detection Transformer, DN-DETR (Li et al., 2022) to demonstrate the effectiveness of Co SNet in the downstream task. We experiment on MS-COCO (Lin et al., 2014) benchmark and utilized DN-DETR s default training settings.
Dataset Splits Yes We evaluate Co SNet on Image Net (Deng et al., 2009) dataset consisting of 1.28M train and 50k validation images of 1000 categories.
Hardware Specification Yes We train models in Py Torch Paszke et al. (2019) using eight NVIDIA A40 GPUs.
Software Dependencies No We train models in Py Torch Paszke et al. (2019) using eight NVIDIA A40 GPUs. All codes shall be open-sourced in Py Torch Paszke et al. (2019) post the review process. Here, we provide a code snippet of a Co SNet-Unit. Please see until the end of this document. The paper mentions PyTorch but does not provide a specific version number.
Experiment Setup Yes Our training methodology is consistent with recent Vanilla Net (Chen et al., 2023). We use data augmentation techniques in (Chen et al., 2023; Liu et al., 2022). See the appendix at the end of this paper for more details. Latency is measured with batch size 1. Table A2: Ablations are conducted at 120 epochs. Table A3: The effect of batch sizes of the baseline approaches, listing batch sizes such as 256, 2048, 4096.