Commit
·
2d3e792
1
Parent(s):
2253c00
Move parsing fb2s to the self._generate_examples func. Use self.base_path to find fb2s in the repo
Browse files- murakami.py +91 -120
murakami.py
CHANGED
@@ -1,5 +1,6 @@
|
|
1 |
"""
|
2 |
-
Parse all paragraphs from all *.fb2 files in the input directory, create a Huggingface Dataset and push it
|
|
|
3 |
"""
|
4 |
|
5 |
|
@@ -7,14 +8,13 @@ import os
|
|
7 |
from pathlib import Path
|
8 |
from lxml import etree
|
9 |
import datasets
|
10 |
-
from datasets import Dataset
|
11 |
-
from huggingface_hub import create_repo
|
12 |
|
13 |
datasets.logging.set_verbosity_info()
|
14 |
|
15 |
|
16 |
_DESCRIPTION = """\
|
17 |
-
Russian translations of Murakami novels, to fine-tune a generative language model. Source is FB2 files
|
|
|
18 |
"""
|
19 |
|
20 |
|
@@ -23,20 +23,7 @@ class Builder(datasets.GeneratorBasedBuilder):
|
|
23 |
|
24 |
VERSION = datasets.Version("1.1.0")
|
25 |
|
26 |
-
# This is an example of a dataset with multiple configurations.
|
27 |
-
# If you don't want/need to define several sub-sets in your dataset,
|
28 |
-
# just remove the BUILDER_CONFIG_CLASS and the BUILDER_CONFIGS attributes.
|
29 |
-
|
30 |
-
# If you need to make complex sub-parts in the datasets with configurable options
|
31 |
-
# You can create your own builder configuration class to store attribute, inheriting from datasets.BuilderConfig
|
32 |
-
# BUILDER_CONFIG_CLASS = MyBuilderConfig
|
33 |
-
|
34 |
-
# You will be able to load one or the other configurations in the following list with
|
35 |
-
# data = datasets.load_dataset('my_dataset', 'first_domain')
|
36 |
-
# data = datasets.load_dataset('my_dataset', 'second_domain')
|
37 |
-
|
38 |
def _info(self):
|
39 |
-
# This method specifies the datasets.DatasetInfo object which contains informations and typings for the dataset
|
40 |
return datasets.DatasetInfo(
|
41 |
# This is the description that will appear on the datasets page.
|
42 |
description=_DESCRIPTION,
|
@@ -44,9 +31,95 @@ class Builder(datasets.GeneratorBasedBuilder):
|
|
44 |
features=datasets.Features({"text": datasets.Value("string")}),
|
45 |
)
|
46 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
47 |
# Number of initial <p> element to take from each fb2, by number. This allows to skip
|
48 |
# intros and other junk in the beginning of an fb2. This is built semi-manually using
|
49 |
-
# the `helper_to_find_first_paragraphs`
|
50 |
START_PARAGRAPHS = {
|
51 |
3: 5,
|
52 |
6: 27,
|
@@ -88,105 +161,3 @@ class Builder(datasets.GeneratorBasedBuilder):
|
|
88 |
print("❌")
|
89 |
for i, p in enumerate(list(paragraphs)[:30]):
|
90 |
print(f" {i} {p.text}")
|
91 |
-
|
92 |
-
def _split_generators(self, dl_manager):
|
93 |
-
# This method is tasked with downloading/extracting the data and defining the splits depending on the configuration
|
94 |
-
# If several configurations are possible (listed in BUILDER_CONFIGS), the configuration selected by the user is in self.config.name
|
95 |
-
data_dir = "data"
|
96 |
-
text_by_name = {}
|
97 |
-
fb2s = list(Path(data_dir).glob("*.fb2"))
|
98 |
-
if len(fb2s) > 0:
|
99 |
-
print(f"Found {len(fb2s)} fb2 files in {data_dir}")
|
100 |
-
else:
|
101 |
-
raise ValueError(f"No fb2 files found in {data_dir}")
|
102 |
-
|
103 |
-
for bi, path in enumerate(fb2s):
|
104 |
-
print(bi, path)
|
105 |
-
|
106 |
-
# Load the FB2 format file
|
107 |
-
with path.open("rb") as file:
|
108 |
-
fb2_data = file.read()
|
109 |
-
|
110 |
-
# Print structure of the FB2 format file
|
111 |
-
# print(etree.tostring(etree.fromstring(fb2_data), pretty_print=True))
|
112 |
-
|
113 |
-
# Parse the FB2 format file using lxml
|
114 |
-
root = etree.fromstring(fb2_data)
|
115 |
-
|
116 |
-
# Get the title of the book
|
117 |
-
title = root.xpath(
|
118 |
-
"//fb:title-info/fb:book-title",
|
119 |
-
namespaces={"fb": "http://www.gribuser.ru/xml/fictionbook/2.0"},
|
120 |
-
)[0].text
|
121 |
-
print(title)
|
122 |
-
|
123 |
-
# Get all book paragraphs
|
124 |
-
paragraphs = root.xpath(
|
125 |
-
"//fb:p",
|
126 |
-
namespaces={"fb": "http://www.gribuser.ru/xml/fictionbook/2.0"},
|
127 |
-
)
|
128 |
-
|
129 |
-
# UNCOMMENT THE LINE BELOW TO BUILD `START_PARAGRAPHS`:
|
130 |
-
# self.helper_to_find_first_paragraphs(paragraphs, title, bi)
|
131 |
-
|
132 |
-
found_paragraphs = []
|
133 |
-
skipping = True
|
134 |
-
for pi, p in enumerate(paragraphs):
|
135 |
-
if p.text is None:
|
136 |
-
continue
|
137 |
-
if (
|
138 |
-
bi in Builder.START_PARAGRAPHS
|
139 |
-
and pi >= Builder.START_PARAGRAPHS[bi]
|
140 |
-
):
|
141 |
-
skipping = False
|
142 |
-
if skipping and p.text.lower() == title.lower():
|
143 |
-
skipping = False
|
144 |
-
if not skipping:
|
145 |
-
found_paragraphs.append(p)
|
146 |
-
print(f"Found {len(found_paragraphs)} paragraphs")
|
147 |
-
|
148 |
-
text_by_name[title] = ""
|
149 |
-
for p in found_paragraphs:
|
150 |
-
text_by_name[title] += p.text.replace(" ", " ") + "\n"
|
151 |
-
text_by_name[title] += "\n"
|
152 |
-
|
153 |
-
print("Novel by size:")
|
154 |
-
for title, text in text_by_name.items():
|
155 |
-
print(f" {title}: {len(text):,} characters")
|
156 |
-
|
157 |
-
smallest_title = min(text_by_name, key=lambda k: len(text_by_name[k]))
|
158 |
-
print(
|
159 |
-
f"Using smallest novel {smallest_title} "
|
160 |
-
f"({len(text_by_name[smallest_title]):,} characters) as a test set"
|
161 |
-
)
|
162 |
-
|
163 |
-
test_titles = [smallest_title]
|
164 |
-
train_titles = [t for t in text_by_name if t not in test_titles]
|
165 |
-
|
166 |
-
return [
|
167 |
-
datasets.SplitGenerator(
|
168 |
-
name=datasets.Split.TRAIN,
|
169 |
-
# These kwargs will be passed to _generate_examples
|
170 |
-
gen_kwargs={
|
171 |
-
"titles": train_titles,
|
172 |
-
"texts": [text_by_name[t] for t in train_titles],
|
173 |
-
"split": "train",
|
174 |
-
},
|
175 |
-
),
|
176 |
-
datasets.SplitGenerator(
|
177 |
-
name=datasets.Split.TEST,
|
178 |
-
# These kwargs will be passed to _generate_examples
|
179 |
-
gen_kwargs={
|
180 |
-
"titles": test_titles,
|
181 |
-
"texts": [text_by_name[t] for t in test_titles],
|
182 |
-
"split": "test",
|
183 |
-
},
|
184 |
-
),
|
185 |
-
]
|
186 |
-
|
187 |
-
# method parameters are unpacked from `gen_kwargs` as given in `_split_generators`
|
188 |
-
def _generate_examples(self, titles, texts, split):
|
189 |
-
# This method handles input defined in _split_generators to yield (key, example) tuples from the dataset.
|
190 |
-
# The `key` is for legacy reasons (tfds) and is not important in itself, but must be unique for each example.
|
191 |
-
for title, text in zip(titles, texts):
|
192 |
-
yield title, {"text": text}
|
|
|
1 |
"""
|
2 |
+
Parse all paragraphs from all *.fb2 files in the input directory, create a Huggingface Dataset and push it
|
3 |
+
to the Hub as `vldsavelyev/murakami`.
|
4 |
"""
|
5 |
|
6 |
|
|
|
8 |
from pathlib import Path
|
9 |
from lxml import etree
|
10 |
import datasets
|
|
|
|
|
11 |
|
12 |
datasets.logging.set_verbosity_info()
|
13 |
|
14 |
|
15 |
_DESCRIPTION = """\
|
16 |
+
Russian translations of Murakami novels, to fine-tune a generative language model. Source is FB2 files
|
17 |
+
from http://flibusta.is/a/8570.
|
18 |
"""
|
19 |
|
20 |
|
|
|
23 |
|
24 |
VERSION = datasets.Version("1.1.0")
|
25 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
26 |
def _info(self):
|
|
|
27 |
return datasets.DatasetInfo(
|
28 |
# This is the description that will appear on the datasets page.
|
29 |
description=_DESCRIPTION,
|
|
|
31 |
features=datasets.Features({"text": datasets.Value("string")}),
|
32 |
)
|
33 |
|
34 |
+
def _split_generators(self, dl_manager: datasets.DownloadManager):
|
35 |
+
data_dir = Path(self.base_path) / "data"
|
36 |
+
fb2_paths = list(data_dir.glob("*.fb2"))
|
37 |
+
if len(fb2_paths) > 0:
|
38 |
+
print(f"Found {len(fb2_paths)} fb2 files in {data_dir}")
|
39 |
+
else:
|
40 |
+
raise ValueError(f"No fb2 files found in {data_dir}")
|
41 |
+
|
42 |
+
smallest_path = min(fb2_paths, key=os.path.getsize)
|
43 |
+
print(f"Using smallest title as a training example: {smallest_path}")
|
44 |
+
|
45 |
+
return [
|
46 |
+
datasets.SplitGenerator(
|
47 |
+
name=datasets.Split.TRAIN,
|
48 |
+
# These kwargs will be passed to _generate_examples
|
49 |
+
gen_kwargs={
|
50 |
+
"filepaths": [p for p in fb2_paths if p != smallest_path],
|
51 |
+
},
|
52 |
+
),
|
53 |
+
datasets.SplitGenerator(
|
54 |
+
name=datasets.Split.TEST,
|
55 |
+
# These kwargs will be passed to _generate_examples
|
56 |
+
gen_kwargs={
|
57 |
+
"filepaths": [smallest_path],
|
58 |
+
},
|
59 |
+
),
|
60 |
+
]
|
61 |
+
|
62 |
+
def _generate_examples(self, filepaths):
|
63 |
+
for fileidx, filepath in enumerate(filepaths):
|
64 |
+
print(fileidx, filepath)
|
65 |
+
title, text = self._extract_text_from_fb2(filepath, fileidx)
|
66 |
+
yield title, {"text": text}
|
67 |
+
|
68 |
+
@staticmethod
|
69 |
+
def _extract_text_from_fb2(filepath: Path, fileidx: int) -> tuple[str, str]:
|
70 |
+
"""
|
71 |
+
Parse FB2 file and return the concatenation of its paragraphs, along with the title.
|
72 |
+
"""
|
73 |
+
# Load the FB2 format file
|
74 |
+
with filepath.open("rb") as file:
|
75 |
+
fb2_data = file.read()
|
76 |
+
|
77 |
+
# Print structure of the FB2 format file
|
78 |
+
# print(etree.tostring(etree.fromstring(fb2_data), pretty_print=True))
|
79 |
+
|
80 |
+
# Parse the FB2 format file using lxml
|
81 |
+
root = etree.fromstring(fb2_data)
|
82 |
+
|
83 |
+
# Get the title of the book
|
84 |
+
title = root.xpath(
|
85 |
+
"//fb:title-info/fb:book-title",
|
86 |
+
namespaces={"fb": "http://www.gribuser.ru/xml/fictionbook/2.0"},
|
87 |
+
)[0].text
|
88 |
+
print(title)
|
89 |
+
|
90 |
+
# Get all book paragraphs
|
91 |
+
paragraphs = root.xpath(
|
92 |
+
"//fb:p",
|
93 |
+
namespaces={"fb": "http://www.gribuser.ru/xml/fictionbook/2.0"},
|
94 |
+
)
|
95 |
+
|
96 |
+
# UNCOMMENT THE LINE BELOW TO BUILD `START_PARAGRAPHS`:
|
97 |
+
# self.helper_to_find_first_paragraphs(paragraphs, title, bi)
|
98 |
+
|
99 |
+
found_paragraphs = []
|
100 |
+
skipping = True
|
101 |
+
for pi, p in enumerate(paragraphs):
|
102 |
+
if p.text is None:
|
103 |
+
continue
|
104 |
+
if (
|
105 |
+
fileidx in Builder.START_PARAGRAPHS
|
106 |
+
and pi >= Builder.START_PARAGRAPHS[fileidx]
|
107 |
+
):
|
108 |
+
skipping = False
|
109 |
+
if skipping and p.text.lower() == title.lower():
|
110 |
+
skipping = False
|
111 |
+
if not skipping:
|
112 |
+
found_paragraphs.append(p)
|
113 |
+
print(f"Found {len(found_paragraphs)} paragraphs")
|
114 |
+
text = ""
|
115 |
+
for p in found_paragraphs:
|
116 |
+
text += p.text.replace(" ", " ") + "\n"
|
117 |
+
text += "\n"
|
118 |
+
return title, text
|
119 |
+
|
120 |
# Number of initial <p> element to take from each fb2, by number. This allows to skip
|
121 |
# intros and other junk in the beginning of an fb2. This is built semi-manually using
|
122 |
+
# the `self.helper_to_find_first_paragraphs` function.
|
123 |
START_PARAGRAPHS = {
|
124 |
3: 5,
|
125 |
6: 27,
|
|
|
161 |
print("❌")
|
162 |
for i, p in enumerate(list(paragraphs)[:30]):
|
163 |
print(f" {i} {p.text}")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|