Train custom NER component with a base model in spaCy v3 - spacy

Question

I'm having problems training a custom NER component within a base model in spaCy's new version.
So far, I've been training my NER model at CLI with the following command:
python -m spacy train en model training validation --base-model en_core_web_sm --pipeline "ner" -R -n 10
Depending on the use case, I took en_core_web_sm or en_core_web_lg as the base model to make use of the other components like tagger and pos.
In spaCy version 3 a config file is required to handle the command at CLI. I'm using the following configurations for training:
[paths]
train = "training/"
dev = "validation/"
vectors = null
init_tok2vec = null
[system]
gpu_allocator = null
seed = 0
[nlp]
lang = "en"
pipeline = ["ner"]
batch_size = 1000
disabled = []
before_creation = null
after_creation = null
after_pipeline_creation = null
tokenizer = {"#tokenizers":"spacy.Tokenizer.v1"}
[components]
[components.ner]
factory = "ner"
moves = null
update_with_oracle_cut_size = 100
[components.ner.model]
#architectures = "spacy.TransitionBasedParser.v2"
state_type = "ner"
extra_state_tokens = false
hidden_width = 64
maxout_pieces = 2
use_upper = true
nO = null
[corpora]
[corpora.dev]
#readers = "spacy.Corpus.v1"
path = ${paths.dev}
max_length = 0
gold_preproc = false
limit = 0
augmenter = null
[corpora.train]
#readers = "spacy.Corpus.v1"
path = ${paths.train}
max_length = 2000
gold_preproc = false
limit = 0
augmenter = null
[training]
dev_corpus = "corpora.dev"
train_corpus = "corpora.train"
seed = ${system.seed}
gpu_allocator = ${system.gpu_allocator}
dropout = 0.1
accumulate_gradient = 1
patience = 1600
max_epochs = 0
max_steps = 20000
eval_frequency = 200
frozen_components = []
before_to_disk = null
[training.batcher]
#batchers = "spacy.batch_by_words.v1"
discard_oversize = false
tolerance = 0.2
get_length = null
[training.batcher.size]
#schedules = "compounding.v1"
start = 100
stop = 1000
compound = 1.001
t = 0.0
[training.logger]
#loggers = "spacy.ConsoleLogger.v1"
progress_bar = false
[training.optimizer]
#optimizers = "Adam.v1"
beta1 = 0.9
beta2 = 0.999
L2_is_weight_decay = true
L2 = 0.01
grad_clip = 1.0
use_averages = false
eps = 0.00000001
learn_rate = 0.001
[training.score_weights]
ents_per_type = null
ents_f = 1.0
ents_p = 0.0
ents_r = 0.0
[pretraining]
[initialize]
vectors = null
init_tok2vec = null
vocab_data = null
lookups = null
before_init = null
after_init = null
[initialize.components]
Since I'm not familiar to spaCy's new version, these are pretty much the default settings.
Unfortunately, I can only the the model from scratch and I can't find an option anymore, to only train the NER component within an existing language model.
I have also tried to add the parser component in the configuration file with
[components]
[components.parser]
source = "en_core_web_sm"
...
But then the model is not even loadable raising the following error
nn_parser.pyx in spacy.syntax.nn_parser.Parser.from_disk()
nn_parser.pyx in spacy.syntax.nn_parser.Parser.Model()
TypeError: Model() takes exactly 1 positional argument (0 given)

Answer 1

In SpaCy 3.0, what you want to do first is initialize your config file to have components that you need:
python -m spacy init config config.cfg --lang en --pipeline tagger,parser,ner,attribute_ruler,senter,lemmatizer,tok2vec
Then, you want to go to the config.cfg and override settings - for example, you can use vectors from existing model:
[initialize]
vectors = "en_core_veb_lg"
init_tok2vec = null
vocab_data = null
lookups = null
before_init = null
after_init = null
Then you can run the train command:
python -m spacy train config.cfg --paths.train ./path_to_your_train_data.spacy --paths.dev ./path_to_your_validation_data.spacy --output ./your_model_name
I also found that it's possible to just go to the model folder and swap out components manually, as well as load different components from different models in the code into a single pipeline.
If you need to use a component from an existing model, you can use the following setting in your config.cfg:
[components.tagger]
source = "en_core_web_lg"
For more info on using existing models and components go to SpaCy documentation.