Some questions about cls_weight in rein_dinov2_mask2former.py

[mingsiliu557]

Dear author,
Thanks for ur share with ur code. But when I trained the head with the dataset REFUGE2, I find some problem here:
class_weight=[1.0] * num_classes + [0.1]
I noticed u set num_classes = 19, but I dont know why need to set a list at the length of 19+1 for class_weight

and I set reduce_zero_label = True for REFUGE2 dataset, and the classes=('background', ' Optic Cup', 'Optic Disc'), I set num_class = 2 for model, but I only get 'background' and 'Optic Cup' in my segmentation mask.
when I set num_class = 3 and class_weight = [0.1 ,1, 2], the index is out of boundary , I check my label:
class_weight: tensor([0.1000, 1.0000, 2.0000], device='cuda:0')
label:tensor([3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3,
1, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3,
3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 2, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3,
3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 0, 3, 3, 3, 3, 3, 3,
3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 0, 3,
3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3,
3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 1,
3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 2, 3, 3, 3, 3, 3,
3, 3, 3, 3, 3, 3, 3, 3], device='cuda:0')
I dont know why my label got a pixel at 3, I have transfered all the pixes to 0~2 before training

[w1oves]

It appears to be due to the fact that `reduce_zero_label=True`. Although I am not aware of the precise reasoning, I can affirm that for N classes, `num_classes` should be set to N, where the "+1" denotes the 'void' class.

…

---- Replied Message ---- | From | ***@***.***> | | Date | 04/02/2024 22:04 | | To | ***@***.***> | | Cc | ***@***.***> | | Subject | [w1oves/Rein] Some questions about cls_weight in rein_dinov2_mask2former.py (Issue #17) | Dear author, Thanks for ur share with ur code. But when I trained the head with the dataset REFUGE2, I find some problem here: class_weight=[1.0] * num_classes + [0.1] I noticed u set num_classes = 19, but I dont know why need to set a list at the length of 19+1 for class_weight and I set reduce_zero_label = True for REFUGE2 dataset, and the classes=('background', ' Optic Cup', 'Optic Disc'), I set num_class = 2 for model, but I only get 'background' and 'Optic Cup' in my segmentation mask. when I set num_class = 3 and class_weight = [0.1 ,1, 2], the index is out of boundary , I check my label: class_weight: tensor([0.1000, 1.0000, 2.0000], device='cuda:0') label:tensor([3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 1, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 2, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 0, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 0, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 1, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 2, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3], device='cuda:0') I dont know why my label got a pixel at 3, I have transfered all the pixes to 0~2 before training — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[w1oves]

If you set reduce_zero_label=True, the code will perform the following actions:

1. Replace all occurrences of 0 with 255.
2. Replace 1 with 0, 2 with 1, and so on.
3. When num_classes=2, the model will only predict classes 0 and 1.

My recommended configuration is as follows:

• If 0 represents the background and not unlabeled regions, you should set reduce_zero_label=False and num_classes=3.
• Otherwise, if 0 represents unlabeled regions, you should set reduce_zero_label=True and num_classes=2. Also, define CLASSES=['Optic Cup', 'Optic Disc'].
• You should maintain class_weight=[1.0] * num_classes + [0.1].

[mingsiliu557]

Thanks for ur kind help, my problem solved WEI, ZHIXIANG ***@***.***>于2024年4月5日 周五17:10写道：

…

If you set reduce_zero_label=True, the code will perform the following actions: 1. Replace all occurrences of 0 with 255. 2. Replace 1 with 0, 2 with 1, and so on. 3. When num_classes=2, the model will only predict classes 0 and 1. My recommended configuration is as follows: - If 0 represents the background and not unlabeled regions, you should set reduce_zero_label=False and num_classes=3. - Otherwise, if 0 represents unlabeled regions, you should set reduce_zero_label=True and num_classes=2. Also, define CLASSES=['Optic Cup', 'Optic Disc']. - You should maintain class_weight=[1.0] * num_classes + [0.1]. — Reply to this email directly, view it on GitHub <#17 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BC4KLOI3ZRGVQM2K5MTZITLY3ZTARAVCNFSM6AAAAABFTOYU36VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDAMZZGI4DOMZTGY> . You are receiving this because you authored the thread.Message ID: ***@***.***>