Official Object Detection Finetuning tutorial code returns NaN loss after a few batches

I’m following this official tutorial on how to fine tune Fast R-CNN object detection, following the code as it is with the same hyperparameter choices, after training for only a few batches the returned loss is NaN. The only difference between my code and the tutorial’s code is that I’m using a subset of the COCO dataset objects to fine tune the model but I wouldn’t imagine that to be the problem since the model is already pre-trained on COCO. Has anyone run into this?