Bug: Inference on cuda results in fixed bounding box offset! about retinaface HOT 2 CLOSED

I have found the bug! For the final annotation return value in the model.predict_jsons() function the tensor containing the bounding boxes is moved to cpu and converted to a numpy array. But the later resizing operation is done on the original tensor variable rather than on the new numpy one. When running on cpu this bug is not apparent because the tensor is not copied when converted to a numpy array but only a new reference is created.

Original:

boxes_np = boxes.cpu().numpy()
landmarks_np = landmarks.cpu().numpy()
resize_coeff = original_height / transformed_height

boxes *= resize_coeff

Modified:

boxes_np = boxes.cpu().numpy()
landmarks_np = landmarks.cpu().numpy()
resize_coeff = original_height / transformed_height

boxes_np *= resize_coeff

from retinaface.

Thanks. Fixed.

from retinaface.