Hello,
If I have a bunch of images that I can perform classification on, and I also have some metadata on which I can train as well, how do I combine the 2 models? Do I take the average of their predictions (maybe weighted) or is there some other way I can achieve this? One example of such a dataset would be RSNA intracranial hemorrhage dataset on Kaggle.
3 Likes
I’ve done an average of the two (or what’s called vote ensembling IIRC) but I haven’t messed around with weighted a whole bunch. I think you’d need an understanding of how each model performs on another hold out set to really justify and play with just how much weight each model should have. Interested to see other’s opinions on the topic 
1 Like
You could consider doing this all in one model, rather than using two separate models. Somethinge like this: https://www.pyimagesearch.com/2019/02/04/keras-multiple-inputs-and-mixed-data/
2 Likes