rpn_loc_loss is nan, fast_rcnn_loc_loss is 0.0

I have changed the database and converted to tfrecords . when i run it for minutes, fast_rcnn_loc_loss become 0.0 rpn_loc_loss become nan.

``

for t_w = np.log(w/reference_w), I have check picture ,w >0 .So I think reference_w is wrong, I don't know how to modify it .Ask for your help.

Reproduce the training run after the database change and TFRecord conversion, then trace the reported rpn_loc_loss and fast_rcnn_loc_loss values through the training entry points. Done means identifying the input or calculation that produces NaN and zero and confirming corrected loss values in a subsequent run.

Written by the indexing model from the issue text.