Problem with empty bounding box

🐛 Bug

Sometimes during the training procedure I get: ValueError: Expected target boxes to be a tensorof shape [N, 4], got torch.Size([0]). It is caused by the transformation: train_tfms = tfms.A.Adapter([*tfms.A.aug_tfms(size=size, presize=presize), tfms.A.Normalize()]). Sometimes tfms is for “aggressive” that it moves the image in such a way that no bbox is left. It crashes the training procedure

To Reproduce Steps to reproduce the behavior:

1. Create the dataset, Wheat is fine for this
2. Apply tfms such as: train_tfms = tfms.A.Adapter([*tfms.A.aug_tfms(size=size, presize=presize), tfms.A.Normalize()])
3. Set faster_rcnn as model
4. Run pytorch lightning training
5. At some random (since aug_tfms is random) point the training will crash.

Expected behavior The training shouldn’t crash, such step should be handled by the training_step/validation_step in the lightning_module or at the model level

Desktop (please complete the following information): Happens both at kaggle notebooks and at colab

[tugot17]

In Wheat dataset it is a case for img with imageid = 2507 (See attached screen). It is rare for this dataset but happens and crashes training.

You can find such situation in your dataset by runnign simple loop

idx_list = []

for idx, x in enumerate(tqdm(train_ds)):
    if not len(x["bboxes"]):
        idx_list.append(idx)
        print(idx)

If not at first time, then try a few times more (since crop is random)

You can chceck out my notebook, hope it helps you reproduce the error

[tugot17]

Yeah solved 😃