Issues Encountered When Reproducing DOVER++ Effects on DEVIDE17k

[wenhao7841]

I encountered some issues while attempting to reproduce the effects of DOVER++ on the DEVIDE17k dataset. I have detailed the problems below, along with the corresponding code snippets and error messages.

Issue 1: AttributeError: 'list' object has no attribute 'float'

Error message:

Linear Epoch 0:
Training in epoch 0: 0%| | 0/228 [00:19<?, ?it/s]
y = data["gt_label"].float().detach().to(device).unsqueeze(-1)
AttributeError: 'list' object has no attribute 'float'

The error occurs at the following line of code:

DOVER/training_with_divide.py

Line 114 in 411c62b

y = data["gt_label"].float().detach().to(device).unsqueeze(-1)

Issue 2: KeyError: 'gt_label_a'

Error message:

y_a = data["gt_label_a"].float().detach().to(device).unsqueeze(-1)
KeyError: 'gt_label_a'

The error occurs at the following line of code:

DOVER/training_with_divide.py

Line 146 in 411c62b

y_a = data["gt_label_a"].float().detach().to(device).unsqueeze(-1)

It appears that there might be an issue with the dataset class. Could this be due to an undisclosed dataset class causing these errors? Alternatively, what steps should I follow to train the model and achieve the effects of the public weights? Any guidance would be greatly appreciated.

[BubbleYu-ZJU]

Setting fully_supervised and plus_plus to False can resolve the issue.