search

Media-Smart / vedadet

A single stage object detection toolbox based on PyTorch
Apache License 2.0
498 stars 128 forks source link

Training tinaface on mmdetection #40

Closed yossibiton closed 3 years ago

yossibiton commented 3 years ago

I'm trying to train a compatible model to tinaface on mmdetection repo. The only difference is that I don't use the IOU-aware head, but normal RetinaHead.

On the first batches, I can see that on mmdetection the bbox loss is twice higher than vedadet. Maybe you can explain that? Do you use different weights initialization than mmdetection? I was really surprised by that because I see your code is based on mmdetection (it actually a fork, although you didn't use github fork)

mileistone commented 3 years ago

you'd better use vedadet.

yossibiton commented 3 years ago

Why should I prefer it ? It looks like you copied the code from mmdetection, and implemented the IoUAware head. Can you explain why didn't you use git fork ?

On Fri, Mar 19, 2021, 07:52 mileistone @.***> wrote:

you'd better use vedadet.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/Media-Smart/vedadet/issues/40#issuecomment-802576781, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABRXKAA7UJWGISRTL5Y4BGLTELRANANCNFSM4ZNAVQGA .

mileistone commented 3 years ago

We re-design MMDetection based on our taste and needs, it's not the same as MMDetection, you may spend some time to compare vedadet and MMDetection if you have interest. If you like MMDetection, you can also use MMDetection, it's OK.