mathiascreutz
commited on
Commit
•
3e9ef65
1
Parent(s):
e81e681
Data loader handles new training set files
Browse files- opusparcus.py +16 -11
opusparcus.py
CHANGED
@@ -41,15 +41,14 @@ _HOMEPAGE = ""
|
|
41 |
|
42 |
_LICENSE = ""
|
43 |
|
44 |
-
# The HuggingFace dataset library
|
45 |
# This can be an arbitrary nested dict/list of URLs (see below in `_split_generators` method)
|
46 |
|
47 |
_URLs = {
|
48 |
-
|
49 |
-
"train": None, # actual value set in the `_split_generators` method
|
50 |
"validation": "validation.jsonl",
|
51 |
"test": "test.jsonl"
|
52 |
|
|
|
53 |
}
|
54 |
|
55 |
_VERSION = datasets.Version("1.0.0", "")
|
@@ -136,12 +135,18 @@ class Opusparcus(datasets.GeneratorBasedBuilder):
|
|
136 |
# It can accept any type or nested list/dict and will give back the same structure with the url replaced with path to local files.
|
137 |
# By default the archives will be extracted and a path to a cached folder where they are extracted is returned instead of the archive
|
138 |
|
139 |
-
if self.config.quality
|
140 |
-
#
|
141 |
-
|
142 |
-
|
143 |
-
|
144 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
145 |
data_dir = dl_manager.download_and_extract(_URLs)
|
146 |
|
147 |
splits = [
|
@@ -176,7 +181,7 @@ class Opusparcus(datasets.GeneratorBasedBuilder):
|
|
176 |
gen_kwargs={
|
177 |
"lang": self.config.lang,
|
178 |
"quality": self.config.quality,
|
179 |
-
"filepath":
|
180 |
"split": "train",
|
181 |
},
|
182 |
)
|
@@ -192,7 +197,7 @@ class Opusparcus(datasets.GeneratorBasedBuilder):
|
|
192 |
# This method handles input defined in _split_generators to yield (key, example) tuples from the dataset.
|
193 |
# The `key` is here for legacy reason (tfds) and is not important in itself.
|
194 |
if split == datasets.Split.TRAIN:
|
195 |
-
with bz2.open(filepath
|
196 |
# We know that this file only contains the desired language,
|
197 |
# because for the training sets the languages are in separate
|
198 |
# files, and only the desired language has been downloaded
|
|
|
41 |
|
42 |
_LICENSE = ""
|
43 |
|
44 |
+
# The HuggingFace dataset library doesn't host the datasets but only point to the original files
|
45 |
# This can be an arbitrary nested dict/list of URLs (see below in `_split_generators` method)
|
46 |
|
47 |
_URLs = {
|
|
|
|
|
48 |
"validation": "validation.jsonl",
|
49 |
"test": "test.jsonl"
|
50 |
|
51 |
+
# NB: the "train" split file is defined dynamically inside the `_split_generators` method
|
52 |
}
|
53 |
|
54 |
_VERSION = datasets.Version("1.0.0", "")
|
|
|
135 |
# It can accept any type or nested list/dict and will give back the same structure with the url replaced with path to local files.
|
136 |
# By default the archives will be extracted and a path to a cached folder where they are extracted is returned instead of the archive
|
137 |
|
138 |
+
if self.config.quality < 70:
|
139 |
+
# We need to retrieve the largest training set file
|
140 |
+
# containing the full training set for the desired language
|
141 |
+
_URLs["train"] = "train_{0}.60.jsonl.bz2".format(self.config.lang)
|
142 |
+
elif self.config.quality <= 95:
|
143 |
+
# We can do with a smaller version of the training set
|
144 |
+
# for the desired language
|
145 |
+
_URLs["train"] = "train_{0}.70.jsonl.bz2".format(self.config.lang)
|
146 |
+
|
147 |
+
# Otherwise, if the desired quality is above 95, we do not
|
148 |
+
# download any training data, because there is no matching data
|
149 |
+
|
150 |
data_dir = dl_manager.download_and_extract(_URLs)
|
151 |
|
152 |
splits = [
|
|
|
181 |
gen_kwargs={
|
182 |
"lang": self.config.lang,
|
183 |
"quality": self.config.quality,
|
184 |
+
"filepath": data_dir["train"],
|
185 |
"split": "train",
|
186 |
},
|
187 |
)
|
|
|
197 |
# This method handles input defined in _split_generators to yield (key, example) tuples from the dataset.
|
198 |
# The `key` is here for legacy reason (tfds) and is not important in itself.
|
199 |
if split == datasets.Split.TRAIN:
|
200 |
+
with bz2.open(filepath, "rt", encoding="utf-8") as f:
|
201 |
# We know that this file only contains the desired language,
|
202 |
# because for the training sets the languages are in separate
|
203 |
# files, and only the desired language has been downloaded
|