Open GotG opened 4 years ago
Definitely post back if you implement it, or if you find a copy of the .cfg file. I'd also be interested in a tinier/faster config.
https://github.com/jnfrancis/Tinier-YOLO by first author
Training with Tinier-1119-1.cfg and current @AlexeyAB Yolo works like a charm ^^
Is it any good?
Not so bad, about 75 FPS, needs to be trained a bit more
@scamianbas did you make any changes to the config? i used this (low number of batches to test the model)
learning_rate=0.0005 max_batches = 5000 policy=steps steps=4000,4500 scales=.1,.1 burn_in=1000
@GotG I only changed max_batches = 20000 and steps=16000,18000 accordingly And also, accordingly to my dataset, classes=2 and filters=21 ( (classes + 5)x3 )
Hello
I am looking for a tiny yolo v3 that has less parameters (by 1/3 or so), so that it is at least 20% faster on inference. Looking at this paper: https://ieeexplore.ieee.org/document/8941141
Does anyone have experience with this Tinier version, or smth similar that would match what I am looking for?
Thanks