I have a lot of (domain)-specific text that I want to classify into 100+ categories.
I want to train a wordembedding (FastText) and use that in conjuction with a CNN, thus I'm running into the problem that I have to optimize both the embedder (vector dimension, epochs, window-size, number of negatives) and the network (learning rate, number of neurons etc) at the same time, which takes indeed some time.
I wonder what the best way is to do here; is it to fix the embedder (say to default values), find the optimal network settings and then optimize the embedder, or is it simply trying to optimize both at once (e.g using Optuna)?
Alternative alternate between optimizing the embedder and optimize the network a few (5?) times