Spancat not being initialized before training

[NixBiks]

I have a config file like this

[paths]
train = null
dev = null
vectors = null
init_tok2vec = null

[system]
gpu_allocator = "pytorch"
seed = 0

[nlp]
lang = "en"
pipeline = ["sentencizer", "domain_entity_ruler", "spancat"]
batch_size = 128
disabled = []
before_creation = null
after_creation = null
after_pipeline_creation = null
tokenizer = {"@tokenizers":"spacy.Tokenizer.v1"}

[components]

[components.sentencizer]
factory = "sentencizer"
overwrite = false
punct_chars = ["!", ".", "\n", "?", "\r", "\r\n"]
scorer = {"@scorers":"spacy.senter_scorer.v1"}

[components.domain_entity_ruler]
factory = "domain_entity_ruler"
include_labels = ["AMOUNT"]

[components.spancat]
factory = "spancat"
max_positive = null
scorer = {"@scorers":"spacy.spancat_scorer.v1"}
spans_key = "sc"
threshold = 0.5

[components.spancat.model]
@architectures = "spacy.SpanCategorizer.v1"

[components.spancat.model.reducer]
@layers = "spacy.mean_max_reducer.v1"
hidden_size = 128

[components.spancat.model.scorer]
@layers = "spacy.LinearLogistic.v1"
nO = null
nI = null

[components.spancat.model.tok2vec]
@architectures = "spacy.Tok2Vec.v1"

[components.spancat.model.tok2vec.embed]
@architectures = "spacy.MultiHashEmbed.v1"
width = 96
rows = [5000,2000,1000,1000]
attrs = ["ORTH","PREFIX","SUFFIX","SHAPE"]
include_static_vectors = false

[components.spancat.model.tok2vec.encode]
@architectures = "spacy.MaxoutWindowEncoder.v1"
width = 96
window_size = 1
maxout_pieces = 3
depth = 4

[components.spancat.suggester]
@misc = "sentence_suggester.v1"

[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 = ["sentencizer"]
annotating_components = ["sentencizer"]
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 = true

[training.optimizer]
@optimizers = "Adam.v1"
beta1 = 0.9
beta2 = 0.999
L2_is_weight_decay = true
L2 = 0.0
grad_clip = 1.0
use_averages = false
eps = 0.00000001
learn_rate = 0.001

[training.score_weights]
sents_f = null
sents_p = null
sents_r = null
spans_sc_f = 1.0
spans_sc_p = 0.0
spans_sc_r = 0.0

[pretraining]

[initialize]
vectors = ${paths.vectors}
init_tok2vec = ${paths.init_tok2vec}
vocab_data = null
lookups = null
after_init = null

[initialize.components]

[initialize.tokenizer]

[initialize.before_init]
@callbacks = "customize_tokenizer"
tokenizers = ["abbreviations_tokenizer", "financial_tokenizer"]

And then I have a script like this to convert to spacy training data

@app.command()
def convert_for_spancat(
    input_file: str = typer.Option(..., "--input"),
    train_file: str = typer.Option(..., "--train"),
    dev_file: str = typer.Option(..., "--dev"),
    config_file: str = typer.Option(..., "--config"),
    split: float = typer.Option(..., "--eval-split"),
):
    """Prepare data for spancat."""
    random.seed(42)
    input_data = models.FilePayload.parse_file(input_file).__root__

    config = load_config(config_file)
    lang_cls = spacy.util.get_lang_class(config["nlp"]["lang"])
    nlp = lang_cls.from_config(config, disable=["spancat"])  # disabling since spancat is not initialized
    train = DocBin()
    dev = DocBin()
    n = len(input_data)
    split_index = int(n * split)
    random.shuffle(input_data)
    span_key = config["components"]["spancat"]["spans_key"]

    for index, example in enumerate(input_data):
        full_text = report_to_text(title=example.title, html=example.html)
        doc = nlp(full_text)
        group = SpanGroup(doc, name=span_key, spans=[])
        for span in doc.sents:
            if example.deal_value is not None and example.deal_value in span.text:
                group.append(span)

        if len(group) == 0 and example.deal_value is not None:
            typer.echo(f"Warning: no spans found for {example.title}")
        doc.spans[span_key] = group
        if index < split_index:
            train.add(doc)
        else:
            dev.add(doc)

    train.to_disk(train_file)
    dev.to_disk(dev_file)

But when I run the following command

spacy train configs/sent_classifier.cfg --output training/ --paths.train corpus/train.spacy --paths.dev corpus/dev.spacy --gpu-id -1

then I get the following error

ValueError: [E143] Labels for component 'spancat' not initialized. This can be fixed by calling add_label, or by providing a representative batch of examples to the component's `initialize` method.

I've been searching around for a solution but to me it looks like my training data and dev data is correctly saved, i.e. doc.spans[span_key] = group for all documents in the DocBins

[NixBiks]

Ahaaa - I had to add a label to the spanthat I add to the SpanGroup, i.e. span.label_ = MY_LABEL. But now I get really strange output from training

=================================== train ===================================
Running command: spacy train configs/sent_classifier.cfg --output training/ --paths.train corpus/train.spacy --paths.dev corpus/dev.spacy --gpu-id -1
ℹ Saving to output directory: training
ℹ Using CPU

=========================== Initializing pipeline ===========================
[2022-10-12 18:06:48,653] [INFO] Set up nlp object from config
[2022-10-12 18:06:48,659] [INFO] Pipeline: ['sentencizer', 'domain_entity_ruler', 'spancat']
[2022-10-12 18:06:48,825] [INFO] Created vocabulary
[2022-10-12 18:06:48,826] [INFO] Finished initializing nlp object
[2022-10-12 18:06:50,673] [INFO] Initialized pipeline components: ['domain_entity_ruler', 'spancat']
✔ Initialized pipeline

============================= Training pipeline =============================
ℹ Pipeline: ['sentencizer', 'domain_entity_ruler', 'spancat']
ℹ Frozen components: ['sentencizer']
ℹ Set annotations on update for: ['sentencizer']
ℹ Initial learn rate: 0.001
E    #       LOSS SPANCAT  SPANS_SC_F  SPANS_SC_P  SPANS_SC_R  SCORE 
---  ------  ------------  ----------  ----------  ----------  ------
  0       0          0.00       29.73       19.30       64.71    0.30
  0     200          0.00       29.73       19.30       64.71    0.30                                                                                
  0     400          0.00       29.73       19.30       64.71    0.30                                                                                
  0     600          0.00       29.73       19.30       64.71    0.30                                                                                
  0     800          0.00       29.73       19.30       64.71    0.30                                                                                
  0    1000          0.00       29.73       19.30       64.71    0.30                                                                                
  1    1200          0.00       29.73       19.30       64.71    0.30                                                                                
  1    1400          0.00       29.73       19.30       64.71    0.30                                                                                
  1    1600          0.00       29.73       19.30       64.71    0.30                                                                                
Epoch 2:   0%|                                                                                                               | 0/200 [00:00<?, ?it/s]✔ Saved pipeline to output directory
training/model-last

FYI my convert script now looks like this

@app.command()
def convert_for_spancat(
    input_file: str = typer.Option(..., "--input"),
    train_file: str = typer.Option(..., "--train"),
    dev_file: str = typer.Option(..., "--dev"),
    config_file: str = typer.Option(..., "--config"),
    split: float = typer.Option(..., "--eval-split"),
):
    """Prepare data for spancat."""
    random.seed(42)
    input_data = models.FilePayload.parse_file(input_file).__root__

    config = load_config(config_file)
    lang_cls = spacy.util.get_lang_class(config["nlp"]["lang"])
    nlp = lang_cls.from_config(config, disable=["spancat"])  # disabling since spancat is not initialized
    nlp.initialize()
    train = DocBin()
    dev = DocBin()
    n = len(input_data)
    split_index = int(n * split)
    random.shuffle(input_data)
    span_key = config["components"]["spancat"]["spans_key"]

    for index, example in enumerate(input_data):
        full_text = report_to_text(title=example.title, html=example.html)
        doc = nlp(full_text)
        group = SpanGroup(doc, name=span_key, spans=[])
        for span in doc.sents:
            if example.deal_value is not None and example.deal_value in span.text:
                span.label_ = "HAS_DEAL_VALUE"
                group.append(span)

        if len(group) == 0 and example.deal_value is not None:
            typer.echo(f"Warning: no spans found for {example.title}")
        doc.spans[span_key] = group
        if index < split_index:
            train.add(doc)
        else:
            dev.add(doc)

    train.to_disk(train_file)
    dev.to_disk(dev_file)

[adrianeboyd]

Hmm, this kind of training output makes it look like the spancat evaluation is reflecting an earlier rule-based component rather than any labels added by spancat itself, which doesn't seem to be learning or predicting anything.

What does domain_entity_ruler do?

Are you sure that the sentence suggester is working as intended? I think you might get this output because there there are no spans being suggested, so there's nothing for the spancat model to work with.