I'm training tok2vec, parser, tagger, and morphologizer. When making predictions, my tagger is constantly making 'XX' predictions and my parser is predicting all relations as 'dep'. My config file was built as such. I did not make any modifications to the config after running the below command.

 python -m spacy init config data/config.cfg --lang en --pipeline tok2vec,tagger,parser,morphologizer --optimize accuracy --gpu 

Im relatively certain that the problem is not my data. The data I originally trained on was an augmented set of OntoNotes 5.0. I tried training my same config on a completely unchanged OntoNotes 5.0. I still get many 'XX' predictions and 'dep' relations. Am I missing something from the pipeline?

This augmented OntoNotes 5.0 does NOT contain any new part of speech tags or dependency relations, it simply copies ~40% of sentences and changes the word form, the POS, and the tag based on some rules.

My metrics are:

    "tag_acc":0.9548473684,
    "dep_uas":0.9040159128,
    "dep_las":0.8812835371,
    ...
    "sents_p":0.9303902061,
    "sents_r":0.926391081,
    "sents_f":0.9283863369,
    "pos_acc":0.9589583168,
    "morph_acc":1.0,
    "morph_per_feat":0.0,
    "tok2vec_loss":111378.3423347773,
    "tagger_loss":9228.9996576309,
    "parser_loss":35001.1950423486,
    "morphologizer_loss":8242.14793396

I only need the parser, tagger, and morphologizer trained. Below is an image that shows a sample output when loading my model as such nlp = spacy.load('data/models/train_5/model-last')