Help me implement "Likelihood Ratio" for detecting OOD-Images in Tensorflow and Jupyter Notebook.

I want to implement the "Likelihood Ratio" from the attached paper. The paper covers both OOD (out-of-distribution) detection of images and of genomic sequences, but the authors only have a github for the latter. I need help to implement a similar code for OOD detection of images. As image data I want to use the MNIST database of handwritten digits. Preferably in Tensorflow and in the Jupyter Notebook Interface.

Here's the github for the detection of OOD genomic sequences: https://github.com/google-research/google-research/tree/master/genomics_ood