Trainer a NER from scratch and reusing the parser, tagger, and other components from en_core_web_md model

[oliviercwa]

I have been turning around for quite some time on this one and I guess I am probably missing something.

I want to:
1- Train a NER on new entities starting from the en_core_web_md vectors
2- Use the newly trained NER in a pipeline along with the parser, lemmatizer, tagger and other components from the en_core_web_md mode. In Spacy 2.1.9, we could this but I have not found the right way in spacy 3.1.2

Here is what I do:
1- I have a simple config for the training that essentially defines a pipeline with tok2vec and ner. All configurations are taken from the en_core_web_md model

[paths]
train = "./data_docbins/ner-data-md-all-train.spacy"
dev = "./data_docbins/ner-data-md-all-eval.spacy"
vectors = null
init_tok2vec = null

[system]
gpu_allocator = null
seed = 0

[nlp]
lang = "en"
pipeline = ["tok2vec", "ner"]
batch_size = 1000
disabled = []
before_creation = null
after_creation = null
after_pipeline_creation = null
tokenizer = {"@tokenizers":"spacy.Tokenizer.v1"}

[components]

[components.tok2vec]
factory = "tok2vec"

[components.tok2vec.model]
@architectures = "spacy.Tok2Vec.v2"

[components.tok2vec.model.embed]
@architectures = "spacy.MultiHashEmbed.v2"
width = ${components.tok2vec.model.encode.width}
attrs = ["ORTH","SHAPE"]
rows = [5000,2500]
include_static_vectors = true

[components.tok2vec.model.encode]
@architectures = "spacy.MaxoutWindowEncoder.v2"
width = 96
depth = 4
window_size = 1
maxout_pieces = 3

[components.ner]
factory = "ner"
incorrect_spans_key = null
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

[components.ner.model.tok2vec]
@architectures = "spacy.Tok2VecListener.v1"
width = ${components.tok2vec.model.encode.width}
upstream = "*"

[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 = []
annotating_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.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_f = 1.0
ents_p = 0.0
ents_r = 0.0
ents_per_type = null

[initialize]
vectors = "en_core_web_md"

# These settings must be added manually
[initialize.lookups]
@misc = "spacy.LookupsDataLoader.v1"
lang = ${nlp.lang}
tables = ["lexeme_norm"]

The NER trains correctly

Now I want to use the trained NER in a more complete pipeline with lemmatizer, etc. I defined a config with all elements but of course the tagger, parser, etc. need to reference the components from en_core_web_md while the NER comes from the newly trained model

I have tried many things:
1- I tried to copy the folders for each component (tagger, parser, etc.) into the new trained model ==> The parser splits sentences in weird ways. The toc2vect has changed so it probably does not help
2- I tried to train with all components, freezing all of them except NER and tok2vect ==> I get errors

I have tried many other things.

What is the proper way to do this ?

[oliviercwa]

I have made a bit of progress essentially by using the source attribute in the config

pipeline = ["tok2vec", "tagger","parser", "attribute_ruler", "lemmatizer", "ner"]
batch_size = 1000
disabled = ["parser"]

...

[components.tagger]
source = "en_core_web_md"

[components.parser]
source = "en_core_web_md"

[components.attribute_ruler]
source = "en_core_web_md"

[components.lemmatizer]
source = "en_core_web_md"

[components.ner]
source = "en_core_web_md"

When I load the resulting pipeline, the final model is able to detect the entities properly. However if I enable back the parser, a sentence of 12 words is split into 8 sentences. The same sentence remains a single sentence when using directly the en_core_web_md model.

The cfg, model and moves files in the parser folder of the trained model seems identical to the ones in en_core_web_md folder.

[oliviercwa]

I believe I found the solution. They key was to freeze absolutely everything and to set the tok2vec source to en_core_web_md.

[components]

[components.tok2vec]
source = "en_core_web_md"

[components.tagger]
source = "en_core_web_md"

[components.parser]
source = "en_core_web_md"

[components.attribute_ruler]
source = "en_core_web_md"

[components.lemmatizer]
source = "en_core_web_md"

[components.ner]
source = "en_core_web_md"

[training]
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 = ["tok2vec", "tagger", "parser", "attribute_ruler", "lemmatizer"]
annotating_components = []
before_to_disk = null

[mbrunecky]

Just my $05, as I too am struggling with using the lemmatizer in my spancat pipeline (see #9201)
I do not see replace_listeners in your

[components.tagger]
source = "en_core_web_md"
replace_listeners = ["model.tok2vec"]

I was told I need this to add the "LEMMA" attribute, in order to use it in my (private, embedded in spancat) tok2vec training.