search

pangsu0613 / CLOCs

CLOCs: Camera-LiDAR Object Candidates Fusion for 3D Object Detection
MIT License
352 stars 68 forks source link

How to quantize #23

Closed harris-mohamed closed 3 years ago

harris-mohamed commented 3 years ago

How would I go about quantizing the pretrained model for execution on an accelerator? I'm guessing I would quantize the 2D model, then the 3D model, and then finally the fusion layer.

pangsu0613 commented 3 years ago

Could you give more explanation about what is quantization and accelerator? I am not very familiar with these terms.

harris-mohamed commented 3 years ago

My aim with CLOCs is to get it running on a small, indoor robot. The focus of my research is to be able to run CLOCs on a lightweight device, for which I will be using an edge TPU, a very tiny computer which is designed to run inference on models very quickly. PyTorch and Tensorflow support model quantization, which is a method of taking floating point weights and converting them to integer representations to save on model size and inference time without losing too much in accuracy. https://pytorch.org/docs/stable/quantization.html explains it pretty well.

pangsu0613 commented 3 years ago

Ok, I see, then I think you are right. First quantize the 2D and 3D model, then the CLOCs fusion layer.