Assertion `input_val >= zero && input_val <= one` failed

[llongbif]

I'm training a YOLOX model for object detection using a dataset that includes:

• Annotated images (positive samples)
• Unannotated images (negative samples, i.e., no bounding boxes)

When I include negative images in the training dataset (images with no annotations), I consistently encounter the following CUDA error:

Assertion input_val >= zero && input_val <= one failed

From this trace:

../aten/src/ATen/native/cuda/Loss.cu:94: operator(): block: [0,0,0], thread: [0,0,0]
Assertion input_val >= zero && input_val <= one failed.
...
terminate called after throwing an instance of 'c10::DistBackendError'
what(): Process group watchdog thread terminated with exception: CUDA error: device-side assert triggered

The unannotated images in the .json file are only in the images section; they do not appear in the annotations section.

Is happening in the epoch 3 or later.

How can I safely include negative samples (images without annotations) in YOLOX training without triggering the CUDA assert?