fionser
commented
10 months ago Hi
You can check out the flag #if USE_CHEETAH to see how we mannual modify the auto-generated code.
Two main modifications.
- For the first layer, we can skip some computation since both input are not secret share. To do so, in CHEETAH, we set a global flag
kIsSharedInput = false. After the first layer, we set this flag to true;
For example, https://github.com/Alibaba-Gemini-Lab/OpenCheetah/blob/main/networks/main_resnet50.cpp#L4989
- The fused normalization which take quite a lot man power to replace all the normalization layer with the fused one. For example https://github.com/Alibaba-Gemini-Lab/OpenCheetah/blob/main/networks/main_resnet50.cpp#L5095
tanjina12
Hi,
I wanted to express my interest in your work and seek your guidance. Recently, I have been experimenting with Cheetah using different sized neural networks. I successfully compiled the model and generated the ‘.cpp file’ using the script provided by Athos. I placed the ‘.cpp’ file in this directory: OpenCheetah/networks. However, I am aware that you manually modified the auto generated ‘.cpp file’, and I’m curious about the specific modifications required to ensure compatibility with Cheetah.
For instance, if I want to compile and run models like ResNet50 using Cheetah as the backend, what are the necessary adjustments I should make? I greatly appreciate your kind assistance in this matter.
Thank you in advance for your time and help.