Model Architecture textcat.train-batch

, ,
1

I was wondering what the exact architecture of the CNN underlying the textcat.train-batch is? I have seen it described as a 'simple cnn' and I assume its the same one being used by spacy's textcat component. Thanks

2

Yes, the architectures are the same ones used in spaCy. As of spaCy v2.1 you can control these by passing the "architecture" argument to the constructor (or via nlp.create_pipe()).

The simple_cnn architecture uses the same token-to-vector strategy as spaCy's NER, dependency parser, tagger etc components. You can find the source code in spacy._ml.Tok2Vec.

3

thanks much!