zpwiki/pages/students/2016/jakub_maruniak/dp2021/annotation/custom_train.py

152 lines
5.1 KiB
Python
Raw Permalink Normal View History

"""Scripts used for training and evaluation of NER models
Usage example:
$ python custom_train.py train ./model ./train.jsonl ./eval.jsonl -o ./output_dir -n 15
Requirements:
spacy>=2.2.3
https://github.com/explosion/projects/tree/master/ner-drugs
"""
import spacy
from spacy.cli.train import _load_pretrained_tok2vec
from timeit import default_timer as timer
from pathlib import Path
import srsly
from wasabi import msg
import random
import plac
import sys
import tqdm
def format_data(data):
result = []
labels = set()
for eg in data:
if eg["answer"] != "accept":
continue
ents = [(s["start"], s["end"], s["label"]) for s in eg.get("spans", [])]
labels.update([ent[2] for ent in ents])
result.append((eg["text"], {"entities": ents}))
return result, labels
@plac.annotations(
model=("The base model to load or blank:lang", "positional", None, str),
train_path=("The training data (Prodigy JSONL)", "positional", None, str),
eval_path=("The evaluation data (Prodigy JSONL)", "positional", None, str),
n_iter=("Number of iterations", "option", "n", int),
output=("Optional output directory", "option", "o", str),
tok2vec=("Pretrained tok2vec weights to initialize model", "option", "t2v", str),
)
def train_model(
model, train_path, eval_path, n_iter=10, output=None, tok2vec=None,
):
"""
Train a model from Prodigy annotations and optionally save out the best
model to disk.
"""
spacy.util.fix_random_seed(0)
with msg.loading(f"Loading '{model}'..."):
if model.startswith("blank:"):
nlp = spacy.blank(model.replace("blank:", ""))
else:
nlp = spacy.load(model)
msg.good(f"Loaded model '{model}'")
train_data, labels = format_data(srsly.read_jsonl(train_path))
eval_data, _ = format_data(srsly.read_jsonl(eval_path))
ner = nlp.create_pipe("ner")
for label in labels:
ner.add_label(label)
nlp.add_pipe(ner)
t2v_cfg = {
"embed_rows": 10000,
"token_vector_width": 128,
"conv_depth": 8,
"nr_feature_tokens": 3,
}
optimizer = nlp.begin_training(component_cfg={"ner": t2v_cfg} if tok2vec else {})
if tok2vec:
_load_pretrained_tok2vec(nlp, Path(tok2vec))
batch_size = spacy.util.compounding(1.0, 32.0, 1.001)
best_acc = 0
best_model = None
row_widths = (2, 8, 8, 8, 8)
msg.row(("#", "L", "P", "R", "F"), widths=row_widths)
for i in range(n_iter):
random.shuffle(train_data)
losses = {}
data = tqdm.tqdm(train_data, leave=False)
for batch in spacy.util.minibatch(data, size=batch_size):
texts, annots = zip(*batch)
nlp.update(texts, annots, drop=0.2, losses=losses)
with nlp.use_params(optimizer.averages):
sc = nlp.evaluate(eval_data)
if sc.ents_f > best_acc:
best_acc = sc.ents_f
if output:
best_model = nlp.to_bytes()
acc = (f"{sc.ents_p:.3f}", f"{sc.ents_r:.3f}", f"{sc.ents_f:.3f}")
msg.row((i + 1, f"{losses['ner']:.2f}", *acc), widths=row_widths)
msg.text(f"Best F-Score: {best_acc:.3f}")
if output and best_model:
with msg.loading("Saving model..."):
nlp.from_bytes(best_model).to_disk(output)
msg.good("Saved model", output)
@plac.annotations(
model=("The model to evaluate", "positional", None, str),
eval_path=("The evaluation data (Prodigy JSONL)", "positional", None, str),
)
def evaluate_model(model, eval_path):
"""
Evaluate a trained model on Prodigy annotations and print the accuracy.
"""
with msg.loading(f"Loading model '{model}'..."):
nlp = spacy.load(model)
data, _ = format_data(srsly.read_jsonl(eval_path))
sc = nlp.evaluate(data)
result = [
("Precision", f"{sc.ents_p:.3f}"),
("Recall", f"{sc.ents_r:.3f}"),
("F-Score", f"{sc.ents_f:.3f}"),
]
msg.table(result)
@plac.annotations(
model=("The model to evaluate", "positional", None, str),
data=("Raw data as JSONL", "positional", None, str),
)
def wps(model, data):
"""
Measure the processing speed in words per second. It's recommended to
use a larger corpus of raw text here (e.g. a few million words).
"""
with msg.loading(f"Loading model '{model}'..."):
nlp = spacy.load(model)
texts = (eg["text"] for eg in srsly.read_jsonl(data))
n_docs = 0
n_words = 0
start_time = timer()
for doc in nlp.pipe(texts):
n_docs += 1
n_words += len(doc)
end_time = timer()
wps = int(n_words / (end_time - start_time))
result = [
("Docs", f"{n_docs:,}"),
("Words", f"{n_words:,}"),
("Words/s", f"{wps:,}"),
]
msg.table(result, widths=(7, 12), aligns=("l", "r"))
if __name__ == "__main__":
opts = {"train": train_model, "evaluate": evaluate_model, "wps": wps}
cmd = sys.argv.pop(1)
if cmd not in opts:
msg.fail(f"Unknown command: {cmd}", f"Available: {', '.join(opts)}", exits=1)
try:
plac.call(opts[cmd])
except KeyboardInterrupt:
msg.warn("Stopped.", exits=1)