Detectron: How can i train model from scratch
Hello.
I want to train mask rcnn from scratch(not using the pre-trained weight)
I hope weight parameter starts from random initialization.
How can i do this?
Hwang-dae-won
All 3 comments
Training from scratch is possible in terms of coding and can be done without much modification in this code. However, there may be convergence problems, e.g., caused by not using BN, or using BN but with a small mini-batch size. We encourage more research to be done on this.
KaimingHe
on 27 Jan 2018
One caveat to add: we noticed just before release that there is currently a bug that will cause a crash when trying to train from scratch (the scale and bias parameters of the AffineChannel ops will not be initialized). We have a patch for this that will hopefully roll out this week. Once that is fixed, leaving TRAIN.WEIGHTS as the empty string will trigger training from scratch. As @KaimingHe says, more research needs to be done before one should expect to get good results.
rbgirshick
on 27 Jan 2018
Since e59c30bb1a6ced1a310b72d563bd9a60aba84999 was committed, it is now possible to train from scratch by setting TRAIN.WEIGHTS to the empty string (equiv. delete from your yaml file). But I want to reinforce Kaiming's point that significant experimentation will be needed to get reasonable results from doing so.
rbgirshick
on 31 Jan 2018
Related issues
twmht
·
3Comments
lilichu
·
3Comments
kampelmuehler
·
4Comments
Adhders
·
3Comments
olgaliak
·
4Comments
Most helpful comment
Training from scratch is possible in terms of coding and can be done without much modification in this code. However, there may be convergence problems, e.g., caused by not using BN, or using BN but with a small mini-batch size. We encourage more research to be done on this.