DPatch Attack Error in PyTorchObjectDetector

[f4str]

Describe the bug
When running the DPatch attack for the PyTorchObjectDetector and PyTorchFasterRCNN, an error can occur when using any PyTorch-specific preprocessors.

To Reproduce
Run the DPatch patch attack on the PyTorchObjectDetector while using a Faster R-CNN model. When using additional PyTorch-specific preprocessors, the error above will occur.

Expected behavior
The attacks should run without error. It can be easily fixed by adding the following line to the _get_losses method in the PyTorchObjectDetector class:

if x_preprocessed.is_leaf:
    x_preprocessed.requires_grad = True
else:
    x_preprocessed.retain_grad()

This error occurs because the gradient retention property is lost when moving a tensor to a different device.

Screenshots
If applicable, add screenshots to help explain your problem.

System information (please complete the following information):

• OS
• Python version
• ART version or commit number
• TensorFlow / Keras / PyTorch / MXNet version