Training on COCO-Person dataset

[QCHighing]

Hi, Thanks for the advance work you have done! This model helped me a lot. I would also like to retrain the model for one class(coco-person), and I have browsed many related issues, but I encountered some doubts.

1. Why the score is getting lower and lower？Looks like it finally converged to around 0.08.
   【0】 10 || B: 1.702 | C: 3.531 | M: 2.854 | S: 0.607 | T: 8.694 || ETA: 7:15:17 || timer: 0.633 【0】100 || B: 1.460 | C: 2.346 | M: 2.624 | S: 0.353 | T: 6.784 || ETA: 7:07:40 || timer: 0.606 【0】300 || B: 1.438 | C: 1.732 | M: 2.513 | S: 0.169 | T: 5.852 || ETA: 7:07:18 || timer: 0.638 【0】800 || B: 1.522 | C: 1.586 | M: 2.468 | S: 0.106 | T: 5.682 || ETA: 7:02:46 || timer: 0.679 【0】2550 || B: 1.486 | C: 1.518 | M: 2.430 | S: 0.087 | T: 5.520 || ETA: 6:45:35 || timer: 0.672

2. How to set lr in this case? I just need to retrain on Person class.

Looking forward to your reply，Thanks a lot!

---

My config is:

yolact_coco_person_config = yolact_base_config.copy({
    'name': 'yolact_coco_person',
    'dataset': coco2017_dataset_person,
    'num_classes': len(coco2017_dataset_person.class_names) + 1,
    # Training params
    'max_iter': 40000,
    'lr': 1e-4,
    'momentum': 0.9,
    'decay': 5e-4,
    'gamma': 0.1,
    'lr_steps': (.35 * 40000, .75 * 40000, .88 * 40000, .93 * 40000),
})

I have removed other annotations and categories in the COCO dataset.
This is my dataset:

coco2017_dataset_person = dataset_base.copy({
    'name': 'COCO 2017 Person',
    'train_images': '../datasets/COCO2017/train2017',
    'train_info': '../datasets/COCO2017/annotations/instances_train2017_person.json',
    'valid_images': '../datasets/COCO2017/val2017',
    'valid_info': '../datasets/COCO2017/annotations/instances_val2017_person.json',
    'has_gt': True,
    'class_names': ('person',),
    'label_map': None,
})

Train command is: python train.py --config=yolact_coco_person_config --resume=weights/yolact_base_54_800000.pth --start_iter=0

[abhigoku10]

@QCHighing are you training only person data? what is the intention of doing this ? is it to obtain high accuracy for person dataset since the total mAP for yolact 29.8 and yolact_plus 34.6 will be get high accruacy for only person

[QCHighing]

@abhigoku10 Yeah, thanks for your Timely reply. I only need to detect the person, as same as Issue #65 . It would be better if I could get higher accuracy. If my environment can be configured, I will also try yolact_plus, but not. Can you give me some suggestions on model training for this issue? Why the score is getting lower and lower？What went wrong with my configuration? Thanks again!

[dbolya]

S is not score but an extra semantic segmentation loss. That going down is perfectly normal (and is what should be happening).

As for lr, I think what you have is fine? Make sure you're loading the pretrained weights with --resume and have applied the fine-tuning change in #36.

[QCHighing]

Thanks a lot.

[Maxinho96]

@QCHighing did you figure out how to improve person mAP? I tried retraining with your settings but person mAP gets worst than the pretrained model on the whole 80 classes.

[abhigoku10]

@Maxinho96 i increased the person data and then used yolact++ architecture i was able to get good improvement

[Maxinho96]

@Maxinho96 i increased the person data and then used yolact++ architecture i was able to get good improvement

Did you use a custom dataset or another online person dataset? I have found Open Images and Cityscapes but I am not sure if they can give an improvement.

[QCHighing]

@Maxinho96 I have stopped this job. I can reopen the issue.

[abhigoku10]

@Maxinho96 i have used pedestrian dataset since my applications required pedestrains scenes

[Maxinho96]

@Maxinho96 i have used pedestrian dataset since my applications required pedestrains scenes

You used Cityscapes maybe? Or you built your dataset? Building an instance segmentation dataset is expensive, so Cityscapes would save a lot of time.

[JiaCheng-Lin]

@QCHighing did you figure out how to improve person mAP? I tried retraining with your settings but person mAP gets worst than the pretrained model on the whole 80 classes.

I have the same problem... @dbolya Do you have any good solutions instead of adding new dataset? Thank you very much!!!

[sam1006]

@QCHighing Hi, did you managed to retain using COCO-person?

[sam1006]

@QCHighing May I request for pre-trained model only for COCO-person class?