Please try below solution which is working on my side.
Change the training image from jpg to png.
$ for i in *.jpg ; do convert "$i" "${i%.*}.png" ; done
The fire can be detected during inference. And there is no Nan issue in evaluation.
More, after changing training .jpg files to .png files, you can also use below loss parameter.
- loss: “cross_entropy”
- weight: 2e-06
- crop_and_resize_prob : 0.01