Train with pretrained weights question

[wumy27]

I have transformed the custom dataset to coco2017 form and use your pre-trained model d2 to train my model.. But run like this:

My training code: python train.py -c 2 -p coco2017 --batch_size 8 --lr 1e-5 --num_epochs 10 --load_weights /weights/efficientdet-d2.pth

My training process:

When i train with adding the parameter --head_only True , the test result like this:

Can any one give me some help?

[zylo117]

the objects are too small, you should try larger network or smaller anchors to fit you gt boxes

[wumy27]

the objects are too small, you should try larger network or smaller anchors to fit you gt boxes

Thanks for your reply. I will use smaller anchors and see if it can help.

[oukohou]

@wumy27 same problem when using --head_only True, what's the result after you using smaller anchors?

[zylo117]

It has nothing to do with head_only, you should try more suitable anchors or larger network/ larger image size

[oukohou]

@zylo117 after seeing other similar issues, maybe the problem lies on the category id? my categories are like following:

"categories": [
    {
      "id": 2,
      "supercategory": "none",
      "name": "car"
    },
    {
      "id": 1,
      "supercategory": "none",
      "name": "bicycle"
    },
    {
      "id": 5,
      "supercategory": "none",
      "name": "tricycle"
    },
    {
      "id": 3,
      "supercategory": "none",
      "name": "truck"
    },
    {
      "id": 6,
      "supercategory": "none",
      "name": "electrocycle"
    },
    {
      "id": 4,
      "supercategory": "none",
      "name": "crowd"
    },
    {
      "id": 0,
      "supercategory": "none",
      "name": "motorcycle"
    }

which starts from 0, while you said category id starts from 1
I think this caused my model to predict really bad?

For your information: after I set debug option to True, the output images are all whole blue after 17epochs, like following:

[zylo117]

it has to start from 1. And if the pictures are all green, it's probably under fitting. you can visualize the loss graph on tensorboard to get more infomation.

[oukohou]

@zylo117 yeah，thanks, I'll change the category id. And as the regression loss is far smaller than cls loss, I manually add weights to regression loss, which is : loss = cls_loss + reg_loss*4000
do you consider it reasonable?

[zylo117]

You are welcome to try and feel free to share the result. Reg loss is much lower than cls loss becasue regression is much easier to train than classification.

[dthaicuong]

it has to start from 1. And if the pictures are all green, it's probably under fitting. you can visualize the loss graph on tensorboard to get more infomation.

I tried few times without category_id starting from 1. My cat_id starts from 0 and it works pretty good. @oukohou: I met blue images as well. It's from ori_imgs somehow. But I just extract ROIS after invert_affine and plot them onto img = cv2.imread()

[zylo117]

@dthaicuong did you modify the code to adapt to cat_id starting from 0? I've been training well with cat id starting from 1 for months.

[dthaicuong]

did you modify the code to adapt to cat_id starting from 0? I've been training well with cat id starting from 1 for months.

@zylo117 I did not change anything from your repo. For few first tries, I forgot "background" class as usual COCO but results are pretty good (or caused by low diversity in my dataset)