How to train and test a model on custom multi class dataset ?

[Usmani255]

We have applied this DeTR method, following exactly the same in this. It works perfectly fine for single-class Face detection detection & it also worked well for our own single-class object detection using Transformer. But when we extend this DeTR code for muti class custom dataset using this tutorial, it didn't work well, I sought the tutorial helps for defining num_class (as num_classes+1) in detr.py file. But even then it didn't work well. If someone could guide us how to apply DeTR for multi-class object detection or if could please refer some DeTR tutorial for multi-class object detection. Thanx!

[Usmani255]

@colesbury @Celebio @pdollar @likethesky Hi sir how are you,Please help me regarding this I run this code on google colab for single class it works well but i failed to get results on multiple class 1.My datasets contains the 15 classes and my last class is is 15 so i change in detr.py to num_classes=16

2.in every Epoch class rate is decreasing in training but it is remained to 100 in after every Epoch during test

3.I am also sharing my built checkpoint.pth file with you

[TalentedMUSE]

I have the same problem! the train class error is always 99

[TalentedMUSE]

We have applied this DeTR method, following exactly the same in this. It works perfectly fine for single-class Face detection detection & it also worked well for our own single-class object detection using Transformer. But when we extend this DeTR code for muti class custom dataset using this tutorial, it didn't work well, I sought the tutorial helps for defining num_class (as num_classes+1) in detr.py file. But even then it didn't work well. If someone could guide us how to apply DeTR for multi-class object detection or if could please refer some DeTR tutorial for multi-class object detection. Thanx!

I solve the problem. First of all, if your dataset is coco and has 15 classes, you should writenum_classes = xxx if args.dataset_file != 'coco' else 15; Then, you should load pretrained model given by the author. In main.py, you should do some modifications: if args.resume: if args.resume.startswith('https'): checkpoint = torch.hub.load_state_dict_from_url( args.resume, map_location='cpu', check_hash=True) else: checkpoint = torch.load(args.resume, map_location='cpu') del checkpoint['model']['class_embed.weight'] del checkpoint['model']['class_embed.bias'] model_without_ddp.load_state_dict(checkpoint['model'], strict=False)

[TalentedMUSE]

Then when you call python main.py, you should add parameter --resume 'xxx.pth'

[aa200647963]

@TalentedMUSE I have a problem of the num_classes，in the detr.py loss_labels function， I find the background class is set to be 91, and the model works well, but when I set the back groundclass to 0, the model works bad, can you tell me why? thx