When you’re working with Python or PyTorch and you encounter the CUDA error: device-side assert triggered, you’re probably having trouble with the number of labels/classes in your code. Another common problem is an incorrect loss function input. In order to fix the problem, you can use the CUDA_LAUNCH_BLOCKING=1 flag.
If you get the device-side assert error while using the GPU, try resetting the kernel. The problem may be related to the unet_learner. For example, if the target number is too high, it causes the assertion to fail. If you don’t change it, you’ll need to restart the kernel session.