[Glean] ResNet-50 Architecture and # MACs

Published: April 03, 2021 by SingularityKChen (Last updated: April 03, 2021 )

• Categories:
• Glean 53

• Tags:
• DLA 19
• ResNet-50 1

• ResNet-50 Architecture and # MACs
  • ResNet-50 Architecture
  • Building Block
  • # Weights and # MACs

ResNet-50 Architecture and # MACs

ResNet-50 Architecture¹

From the figure above, ResNet-50 contains 2 separate convolutional layers plus 16 building block where each building block contains three convolutional layers.

Building Block¹

The building block in residual learning contains one residual representations and one shortcut connections which skipping one or more layers.

In ResNet-50, the shortcut connections skipping three layers.

# Weights and # MACs²

We can conclude that the last three building blocks occupy more weights while their computation operations near the same between different building blocks.

1. K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” arXiv:1512.03385 [cs], Dec. 2015, Accessed: Dec. 22, 2019. [Online]. Available: http://arxiv.org/abs/1512.03385. ↩ ↩²

2. 经典神经网络参数的计算【不定期更新】, https://zhuanlan.zhihu.com/p/49842046 ↩