Summary of Faster Inference Of Integer Swin Transformer by Removing the Gelu Activation, By Mohammadreza Tayaranian et al.
Faster Inference of Integer SWIN Transformer by Removing the GELU Activation
by Mohammadreza Tayaranian, Seyyed Hasan Mozafari, James J. Clark, Brett Meyer, Warren Gross
First submitted to arxiv on: 2 Feb 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Artificial Intelligence (cs.AI)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The paper presents a method to improve the inference latency of the state-of-the-art Swin Transformer model, which is known for its high accuracy in image classification tasks but has slower inference compared to other deep neural networks. The authors achieve this by removing floating-point operations associated with the GELU activation function and replacing it with ReLU activation, which has lower memory and computation complexity. They also use iterative knowledge distillation to compensate for any loss of accuracy due to the replacement. The proposed method is shown to improve the inference latency of the quantized Swin Transformer by at least 11% on an RTX 4090 NVIDIA GPU while maintaining an accuracy drop of under 0.5% on the ImageNet evaluation dataset. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper helps make a popular image classification model, called Swin Transformer, run faster without losing too much accuracy. To do this, they replace a special activation function called GELU with ReLU, which uses less computer power and memory. They also use a technique called knowledge distillation to help the new model remember what it learned from the old one. The result is that the model can process images 11% faster while still being almost as accurate on the ImageNet dataset. |
Keywords
» Artificial intelligence » Image classification » Inference » Knowledge distillation » Relu » Transformer
