search

pjreddie / darknet

Convolutional Neural Networks
http://pjreddie.com/darknet/
Other
25.75k stars 21.33k forks source link

Why can't compressed models reduce processing time significantly? Let's take a glimpse. #448

Open patrick-ucr opened 6 years ago

patrick-ucr commented 6 years ago
Model Size(MB) Precision(IOU) Recall(Ratio) Processing Time(ms)
Yolo 195 0.74 0.93 2034
Tiny-Yolo 61 0.62 0.77 1160
Compressed Tiny-Yolo 25 0.52 0.58 837
Tiny-Darknet 3.6 0.54 0.63 954

Yolo and Tiny-Yolo are reference models and Compressed Tiny-Yolo (remove 12th and 13th layers) and Tiny-Darknet (based on SqueezeNet) are trained from scratch. One of my observations is (note that processing times are measured on a phone) although Tiny-Darknet can reduce the weight file size, it cannot translate into a significant reduction in processing (inference) time.

I am looking into other techniques such as quantization. Anyone has tried something similar before?

xiaozan2020 commented 4 years ago

Hi @patrick-ucr, I'm research on the same work recently, just wondering that did you figure out the reason?

By the way, did you do all the training work under darknet framework?

Thanks