spaCy/spacy/train.py

90 lines
3.2 KiB
Python
Raw Normal View History

# coding: utf8
from __future__ import absolute_import, unicode_literals
import random
2017-03-11 17:12:48 +00:00
import tqdm
from thinc.neural.optimizers import Adam
from thinc.neural.ops import NumpyOps, CupyOps
from .gold import GoldParse, merge_sents
from .scorer import Scorer
class Trainer(object):
"""
Manage training of an NLP pipeline.
"""
def __init__(self, nlp, gold_tuples):
self.nlp = nlp
self.gold_tuples = gold_tuples
2017-03-11 17:12:48 +00:00
self.nr_epoch = 0
2016-10-13 01:24:53 +00:00
def epochs(self, nr_epoch, augment_data=None, gold_preproc=False):
2016-11-25 14:55:33 +00:00
cached_golds = {}
def _epoch(indices):
2017-03-11 17:12:48 +00:00
for i in tqdm.tqdm(indices):
2016-11-25 14:55:33 +00:00
raw_text, paragraph_tuples = self.gold_tuples[i]
2016-10-13 01:24:53 +00:00
if gold_preproc:
raw_text = None
else:
paragraph_tuples = merge_sents(paragraph_tuples)
2016-11-25 14:55:33 +00:00
if augment_data is None:
docs = self.make_docs(raw_text, paragraph_tuples)
if i in cached_golds:
golds = cached_golds[i]
else:
golds = self.make_golds(docs, paragraph_tuples)
else:
raw_text, paragraph_tuples = augment_data(raw_text, paragraph_tuples)
2016-11-25 14:55:33 +00:00
docs = self.make_docs(raw_text, paragraph_tuples)
golds = self.make_golds(docs, paragraph_tuples)
for doc, gold in zip(docs, golds):
yield doc, gold
2016-11-25 14:55:33 +00:00
indices = list(range(len(self.gold_tuples)))
for itn in range(nr_epoch):
2016-11-25 14:55:33 +00:00
random.shuffle(indices)
yield _epoch(indices)
2017-03-11 17:12:48 +00:00
self.nr_epoch += 1
def update(self, docs, golds, drop=0.):
2016-10-15 21:53:46 +00:00
for process in self.nlp.pipeline:
if hasattr(process, 'update'):
loss = process.update(doc, gold, sgd=self.sgd, drop=drop,
itn=self.nr_epoch)
self.sgd.finish_update()
else:
process(doc)
return doc
2016-10-13 01:24:53 +00:00
def evaluate(self, dev_sents, gold_preproc=False):
scorer = Scorer()
for raw_text, paragraph_tuples in dev_sents:
2016-10-13 01:24:53 +00:00
if gold_preproc:
raw_text = None
else:
paragraph_tuples = merge_sents(paragraph_tuples)
docs = self.make_docs(raw_text, paragraph_tuples)
golds = self.make_golds(docs, paragraph_tuples)
for doc, gold in zip(docs, golds):
2016-11-25 14:55:33 +00:00
for process in self.nlp.pipeline:
process(doc)
scorer.score(doc, gold)
return scorer
def make_docs(self, raw_text, paragraph_tuples):
if raw_text is not None:
return [self.nlp.tokenizer(raw_text)]
else:
return [self.nlp.tokenizer.tokens_from_list(sent_tuples[0][1])
for sent_tuples in paragraph_tuples]
def make_golds(self, docs, paragraph_tuples):
if len(docs) == 1:
2016-11-25 14:55:33 +00:00
return [GoldParse.from_annot_tuples(docs[0], sent_tuples[0])
for sent_tuples in paragraph_tuples]
else:
2016-11-25 14:55:33 +00:00
return [GoldParse.from_annot_tuples(doc, sent_tuples[0])
for doc, sent_tuples in zip(docs, paragraph_tuples)]