EcoFormer: Energy-Saving Attention with Linear Complexity

Authors: Jing Liu, Zizheng Pan, Haoyu He, Jianfei Cai, Bohan Zhuang

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on both vision and language tasks show that Eco Former consistently achieves comparable performance with standard attentions while consuming much fewer resources.
Researcher Affiliation Academia Department of Data Science & AI, Monash University, Australia
Pseudocode No The paper describes the proposed method in prose and through diagrams, but it does not include structured pseudocode or algorithm blocks.
Open Source Code Yes Code is available at https://github.com/ziplab/EcoFormer.
Open Datasets Yes To investigate the effectiveness of the proposed method, we conduct experiments on Image Net-1K [30], a large-scale image classification dataset that contains 1.2M training images from 1K categories and 50K validation images.
Dataset Splits Yes Image Net-1K [30], a large-scale image classification dataset that contains 1.2M training images from 1K categories and 50K validation images.
Hardware Specification Yes All models in this experiment are trained on 8 V100 GPUs with a total batch size of 256. ... Moreover, we report the on-chip energy consumption according to Table 1 and the throughput with a mini-batch size of 32 on a single NVIDIA RTX 3090 GPU.
Software Dependencies No The paper mentions using the Adam W optimizer and states implementations are based on released code from other papers, but it does not specify software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes All training images are resized to 256 × 256, and 224 × 224 patches are randomly cropped from an image or its horizontal flip, with the per-pixel mean subtracted. ... Next, we finetune each model on Image Net-1K with 100 epochs. ... All models in this experiment are trained on 8 V100 GPUs with a total batch size of 256. We set the initial learning rate to 2.5 × 10−5 for PVTv2 and 1.25 × 10−4 for Twins. We use Adam W optimizer [40] with a cosine decay learning rate scheduler.