Tensorboard provides access plotting scalars, histograms and images. I am trying to add model predictions as simple vectors to tensorboard. Currently I am doing that through adding matplotlib images, which works great to visual inspection, but it's not possible to infer the prediction traces after model training. There are methods to represent the traces as scalars in tensorboard, as recommended here, but I am curious what other methods exist, since it 1. does not seem that tensorboard provides native support for that, and 2. it is due to scalar logging rather slow.

Are there other methods in tensorboard for vector-based logging? And which other methods could be recommended for that purpose? My best bet would be to simple save them as .npy arrays, but when reading the run computations I would then end up with multiple files in addition to the .tsevents file. So I am just curious what people can recommend.