search

aim-uofa / AdelaiDet

AdelaiDet is an open source toolbox for multiple instance-level detection and recognition tasks.
https://git.io/AdelaiDet
Other
3.34k stars 643 forks source link

Weird output of ABCNet #371

Open NguyenHongSon1103 opened 3 years ago

NguyenHongSon1103 commented 3 years ago

Hi ! I trained BAText model with my custom dataset, but rec_loss is very high (around 0.7) even though after 300k iteration. image

And when I used a checkpoint for visualize result on some training image, it all output wrong with same format of text 0006333 0032673

I've checked my json but nothing wrong with it. Im using batchsize = 1, and lr=0.0005, maybe it relate ? Thank for your help !

xcharxlie commented 3 years ago

@NguyenHongSon1103 Did you solve it yet? I'm now encountering a problem similar to yours. May I ask if you pretrained the model as the instruction says or just build one on your own datasets?

NguyenHongSon1103 commented 3 years ago

@xcharxlie I think my problem is my dataset not enough for model's converging. I used it for Vietnamese language so cannot use Syntexth150k or pretrained model on it. Then, I created my own data, trained on it and fine-tune with practical data, now it work

AmjadAlsulami commented 2 years ago

Hi @NguyenHongSon1103 Can you please share your steps in in how you trained and fine-tune Abcnet with your own data, I'm having issues now while trying to do this

NguyenHongSon1103 commented 2 years ago

@AmjadAlsulami I created synthetic data with Vietnamese modification version of this https://github.com/ankush-me/SynthText, then followed the instruction of Abcnet