CPU Mode for This Batch

It has error:
------------CPU Mode for This Batch-------------
gt bboxes: tensor([[ 401.5000, 126.8125, -288.5000, 2.9141]])
bboxes preds: tensor([[ 12.5781, 6.4883, 24.6875, 10.9297],
[ 18.9219, 5.5117, 37.2812, 9.8672],
[ 22.8438, 5.0117, 43.4688, 9.4609],
...,
[235.5000, 255.5000, 485.2500, 708.5000],
[218.6250, 233.1250, 41.9688, 54.6562],
[207.1250, 219.7500, 223.6250, 72.6250]])
gt classes: tensor([0.])
expanded_strides: tensor([[ 8., 8., 8., ..., 32., 32., 32.]])
x_shifts: tensor([[0., 1., 2., ..., 5., 6., 7.]], dtype=torch.float16)
y_shifts: tensor([[0., 0., 0., ..., 7., 7., 7.]], dtype=torch.float16)
gt bbox per image: tensor([[ 401.5000, 126.8125, -288.5000, 2.9141]])
bboxes preds per image: tensor([], size=(0, 4))
cls preds, obj preds: tensor([], size=(0, 3), dtype=torch.float16) tensor([], size=(0, 1), dtype=torch.float16)

I check annotations in coco.py, box has no negative value, why gt_bboxes_per_image has negative values, I turn off mosaic and mixup, Any advices in help!

Start with coco.py and trace the CPU training batch path that produces gt_bboxes_per_image, comparing it with the logged annotations and the disabled mosaic and mixup settings. Reproduce the reported batch if possible; done means the source of the negative coordinates is identified and the issue has a verified correction or clear diagnosis.