explosion/spaCy
1
1
Fork 0
mirror of https://github.com/explosion/spaCy.git synced 2024-12-25 09:26:27 +03:00
Code Issues Packages Projects Releases Wiki Activity

Add support for fiddly hyper-parameters to train func

Browse Source
This commit is contained in:
Matthew Honnibal 2017-05-22 04:47:14 -05:00
parent 80e19a2399
commit bc2294d7f1
1 changed files with 16 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
spacy/cli/train.py
View File

@ -7,6 +7,7 @@ import cytoolz
from pathlib import Path from pathlib import Path
import dill import dill
import tqdm import tqdm
from thinc.neural.optimizers import linear_decay
from ..tokens.doc import Doc from ..tokens.doc import Doc
from ..scorer import Scorer from ..scorer import Scorer
@ -40,23 +41,34 @@ def train(lang_id, output_dir, train_data, dev_data, n_iter, n_sents,
corpus = GoldCorpus(train_path, dev_path) corpus = GoldCorpus(train_path, dev_path)
dropout = util.env_opt('dropout', 0.0) dropout = util.env_opt('dropout', 0.0)
dropout_decay = util.env_opt('dropout_decay', 0.0)
optimizer = nlp.begin_training(lambda: corpus.train_tuples, use_gpu=use_gpu) optimizer = nlp.begin_training(lambda: corpus.train_tuples, use_gpu=use_gpu)
n_train_docs = corpus.count_train() n_train_docs = corpus.count_train()
batch_size = float(util.env_opt('min_batch_size', 4))
max_batch_size = util.env_opt('max_batch_size', 64)
batch_accel = util.env_opt('batch_accel', 1.001)
print("Itn.\tDep. Loss\tUAS\tNER F.\tTag %\tToken %") print("Itn.\tDep. Loss\tUAS\tNER F.\tTag %\tToken %")
for i in range(n_iter): for i in range(n_iter):
with tqdm.tqdm(total=n_train_docs) as pbar: with tqdm.tqdm(total=n_train_docs) as pbar:
train_docs = corpus.train_docs(nlp, shuffle=i, projectivize=True) train_docs = corpus.train_docs(nlp, shuffle=i, projectivize=True)
for batch in cytoolz.partition_all(20, train_docs): idx = 0
while idx < n_train_docs:
batch = list(cytoolz.take(int(batch_size), train_docs))
if not batch:
break
docs, golds = zip(*batch) docs, golds = zip(*batch)
docs = list(docs)
golds = list(golds)
nlp.update(docs, golds, drop=dropout, sgd=optimizer) nlp.update(docs, golds, drop=dropout, sgd=optimizer)
pbar.update(len(docs)) pbar.update(len(docs))
idx += len(docs)
batch_size *= batch_accel
batch_size = min(int(batch_size), max_batch_size)
dropout = linear_decay(dropout, dropout_decay, i*n_train_docs+idx)
with nlp.use_params(optimizer.averages): with nlp.use_params(optimizer.averages):
scorer = nlp.evaluate(corpus.dev_docs(nlp)) scorer = nlp.evaluate(corpus.dev_docs(nlp))
print_progress(i, {}, scorer.scores) print_progress(i, {}, scorer.scores)
with (output_path / 'model.bin').open('wb') as file_: with (output_path / 'model.bin').open('wb') as file_:
with nlp.use_params(optimizer.averages):
dill.dump(nlp, file_, -1) dill.dump(nlp, file_, -1)