This article is the third day article of Natural Language Processing # 2 Advent Calendar 2019.
This article deals with learning document categorization using spaCy's command-line interface. Eventually, I would like to use GiNZA to classify Japanese documents, but unfortunately I have exhausted my efforts so I will only handle English documents.
In the following articles I wrote earlier, Japanese documents are classified without using the CLI. https://qiita.com/kyamamoto9120/items/cd36017be13529366767
In the above article, the problem was that although categorization could be learned, it was very slow. The cause is that it uses only one core CPU and does not support GPU.
Looking back at the documentation to address this issue, I noticed that when learning with spaCy, it was suggested that using the CLI was the best way to do it for most purposes. https://spacy.io/usage/training#spacy-train-cli
On the other hand, there is not much talk about how to learn using the spaCy CLI, and there is almost no information in Japanese. So I decided to try it to help someone.
We are very welcome, so thank you.
This article has been validated with spaCy 2.2.3. The execution environment uses the GPU runtime of google Colab, and the GPU is Tesla P100.
Also, use fetch_20newsgroups of scikit-learn for the dataset. I think that it is similar to livedoor news corpus in Japanese in 20 categories of news articles.
In order to learn with the spaCy CLI, the data must be in JSON format as shown below.
[{
"id": int, # ID of the document within the corpus
"paragraphs": [{ # list of paragraphs in the corpus
"raw": string, # raw text of the paragraph
"sentences": [{ # list of sentences in the paragraph
"tokens": [{ # list of tokens in the sentence
"id": int, # index of the token in the document
"dep": string, # dependency label
"head": int, # offset of token head relative to token index
"tag": string, # part-of-speech tag
"orth": string, # verbatim text of the token
"ner": string # BILUO label, e.g. "O" or "B-ORG"
}],
"brackets": [{ # phrase structure (NOT USED by current models)
"first": int, # index of first token
"last": int, # index of last token
"label": string # phrase label
}]
}],
"cats": [{ # new in v2.2: categories for text classifier
"label": string, # text category label
"value": float / bool # label applies (1.0/true) or not (0.0/false)
}]
}]
}]
Source: https://spacy.io/api/annotation#json-input
Conversion from spaCy Doc to the above JSON format can be done by using gold.docs_to_json.
The following is the process to convert the fetch_20newsgroups dataset to the above JSON format with reference to example.
import spacy
import srsly
from spacy.gold import docs_to_json
from sklearn.datasets import fetch_20newsgroups
from sklearn.model_selection import train_test_split
def save_to_json(model, data, targets, target_names, output_file, n_texts=0):
def get_categories(target):
return dict([(key, int(target == i)) for i, key in enumerate(target_names)])
nlp = spacy.load(model)
nlp.disable_pipes(*nlp.pipe_names)
sentencizer = nlp.create_pipe("sentencizer")
nlp.add_pipe(sentencizer, first=True)
docs = []
count = 0
for i, doc in enumerate(nlp.pipe(data)):
doc.cats = get_categories(targets[i])
docs.append(doc)
if n_texts > 0 and count == n_texts:
break
count += 1
srsly.write_json(output_file, [docs_to_json(docs)])
return count
newsgroups_train = fetch_20newsgroups(subset='train')
newsgroups_test = fetch_20newsgroups(subset='test')
X_train, X_dev, y_train, y_dev = train_test_split(newsgroups_train.data, newsgroups_train.target, test_size=0.20, random_state=42)
save_to_json(
'en_core_web_sm',
X_train,
y_train,
newsgroups_train.target_names,
'train.json',
)
save_to_json(
'en_core_web_sm',
X_dev,
y_dev,
newsgroups_train.target_names,
'dev.json',
)
save_to_json(
'en_core_web_sm',
newsgroups_test.data,
newsgroups_test.target,
newsgroups_test.target_names,
'test.json',
)
Learning is done with the python -m spacy train command.
I am using -v instead of -b to specify the base model, which is based on the following issue.
Reference: https://github.com/explosion/spaCy/issues/4504
$ time python -m spacy train en output train.json dev.json -v en_core_web_sm -p textcat -ta simple_cnn -g 0
Training pipeline: ['textcat']
Starting with blank model 'en'
Loading vector from model 'en_core_web_sm'
Counting training words (limit=0)
tcmalloc: large alloc 2035253248 bytes == 0x65048000 @ 0x7fdf2bb321e7 0x5acc1b 0x7fdf2115d5db 0x7fdf2115dbf0 0x7fdf2115de36 0x7fdf2115b5c1 0x50abc5 0x50c549 0x7fdec8648b20 0x7fdec864d98f 0x7fdec86406c5 0x7fdec868fc47 0x7fdecb4f172a 0x7fdec867cfce 0x7fdecb4f172a 0x7fdec86a91e7 0x7fdecb4f172a 0x7fdec8670148 0x7fdec867a24b 0x594f4c 0x54a2c5 0x551761 0x5aa69c 0x50ab53 0x50d320 0x5081d5 0x5895e1 0x5a04ce 0x50d8f5 0x5081d5 0x50a020
Textcat evaluation score: F1-score macro-averaged across the labels
'alt.atheism, comp.graphics, comp.os.ms-windows.misc, comp.sys.ibm.pc.hardware,
comp.sys.mac.hardware, comp.windows.x, misc.forsale, rec.autos, rec.motorcycles,
rec.sport.baseball, rec.sport.hockey, sci.crypt, sci.electronics, sci.med,
sci.space, soc.religion.christian, talk.politics.guns, talk.politics.mideast,
talk.politics.misc, talk.religion.misc'
Itn Textcat Loss Textcat Token % CPU WPS GPU WPS
--- ------------ ------- ------- ------- -------
1 1300.851 45.507 100.000 34746 125805
2 210.326 62.555 100.000 34629 125121
3 155.350 71.460 100.000 35131 125583
4 120.228 76.617 100.000 35244 126553
5 104.154 78.516 100.000 35371 126499
6 64.690 80.724 100.000 35491 126734
7 57.947 82.024 100.000 35750 123946
8 49.666 82.662 100.000 35840 126354
9 38.450 83.410 100.000 35538 126965
10 36.798 83.496 100.000 35435 125634
11 24.351 83.933 100.000 35467 125578
12 24.935 83.989 100.000 35490 125223
13 26.003 84.009 100.000 35329 125631
14 16.517 84.013 100.000 35255 123643
15 16.210 84.367 100.000 35234 124040
16 16.640 84.601 100.000 35554 126235
17 17.365 84.494 100.000 35436 126539
18 18.860 84.866 100.000 35596 126301
19 11.555 84.910 100.000 35644 126256
20 17.691 84.916 100.000 35481 125505
21 16.697 85.556 100.000 35767 125510
22 13.508 84.815 100.000 35795 126961
23 12.806 84.610 100.000 36021 126313
24 6.776 84.754 100.000 36155 126526
25 10.806 85.122 100.000 35866 126933
26 10.999 85.007 100.000 35972 126758
27 8.463 85.224 100.000 36065 125702
28 7.601 84.920 100.000 35606 127107
29 10.220 85.088 100.000 36242 127223
30 7.911 84.889 100.000 36157 128428
✔ Saved model to output directory
output/model-final
✔ Created best model
output/model-best
real 97m9.901s
user 87m25.982s
sys 9m32.144s
It seems that atmosphere learning is being done, and it seems that GPU is also used. In fact, the Tesla P100 allows you to learn much faster than with a CPU alone. With the K80, there wasn't much difference with the CPU alone, so I think it's better to carefully select the GPU when using Google Colab.
Learning is done with the python -m spacy evaluate command.
I wasn't addicted to this, and I was able to execute it smoothly.
$ time python -m spacy evaluate output/model-best/ test.json -g 0
tcmalloc: large alloc 1610211328 bytes == 0x4ce4c000 @ 0x7fce167961e7 0x5acc1b 0x7fce0bdc15db 0x7fce0bdc1bf0 0x7fce0bdc1e36 0x7fce0bdbf5c1 0x50abc5 0x50c549 0x7fcdb32acb20 0x7fcdb32b198f 0x7fcdb32a46c5 0x7fcdb32f3c47 0x7fcdb615572a 0x7fcdb32e0fce 0x7fcdb615572a 0x7fcdb330d1e7 0x7fcdb615572a 0x7fcdb32d4148 0x7fcdb32de24b 0x594f4c 0x54a2c5 0x551761 0x5aa69c 0x50ab53 0x50c549 0x5081d5 0x5895e1 0x5a04ce 0x50d8f5 0x5081d5 0x50a020
tcmalloc: large alloc 1610211328 bytes == 0x4ce4c000 @ 0x7fce167961e7 0x5acc1b 0x7fce0bdc15db 0x7fce0bdc1bf0 0x7fce0bdc1e36 0x7fce0bdbf5c1 0x50abc5 0x50c549 0x7fcdb32acb20 0x7fcdb32b198f 0x7fcdb32a46c5 0x7fcdb32f3c47 0x7fcdb615572a 0x7fcdb32e0fce 0x7fcdb615572a 0x7fcdb330d1e7 0x7fcdb615572a 0x7fcdb32d4148 0x7fcdb32de3fe 0x594f4c 0x54a2c5 0x551761 0x5aa69c 0x50ab53 0x50c549 0x5081d5 0x5895e1 0x5a04ce 0x50d8f5 0x5081d5 0x50a020
================================== Results ==================================
Time 24.76 s
Words 3124908
Words/s 126185
TOK 100.00
POS 0.00
UAS 0.00
LAS 0.00
NER P 0.00
NER R 0.00
NER F 0.00
Textcat 70.99
real 1m22.570s
user 1m13.028s
sys 0m9.359s
The score at the time of evaluation is much lower than the score at the time of learning. The cause has not been investigated, but I think this is an issue when working on an actual task.
I feel that there is a lot of room for improvement because we have not performed any pretreatment this time.
In this article, we used spaCy's CLI to classify a dataset called fetch_20newsgroups. It's a poor article, but I hope it helps someone.
Next, I would like to categorize Japanese documents using the same method.
Recommended Posts