Commit
•
e1feddc
1
Parent(s):
00a53ae
Refactor wmt_utils to add language_pair to SubDataset
Browse files- wmt_utils.py +29 -28
wmt_utils.py
CHANGED
@@ -53,7 +53,7 @@ CWMT_SUBSET_NAMES = ["casia2015", "casict2011", "casict2015", "datum2015", "datu
|
|
53 |
class SubDataset:
|
54 |
"""Class to keep track of information on a sub-dataset of WMT."""
|
55 |
|
56 |
-
def __init__(self, name, target, sources,
|
57 |
"""Sub-dataset of WMT.
|
58 |
|
59 |
Args:
|
@@ -83,33 +83,34 @@ class SubDataset:
|
|
83 |
self._manual_dl_files = manual_dl_files if manual_dl_files else []
|
84 |
self.name = name
|
85 |
self.target = target
|
86 |
-
self.sources = set(sources)
|
|
|
87 |
|
88 |
-
def _inject_language(self, src, strings):
|
89 |
"""Injects languages into (potentially) template strings."""
|
90 |
-
if src not in self.
|
91 |
-
raise ValueError(f"Invalid source for '{self.name}': {src}")
|
92 |
|
93 |
def _format_string(s):
|
94 |
-
if "{0}" in s and "{1}" and "{src}" in s:
|
95 |
-
return s.format(*sorted([src,
|
96 |
elif "{0}" in s and "{1}" in s:
|
97 |
-
return s.format(*sorted([src,
|
98 |
elif "{src}" in s:
|
99 |
-
return s.format(src=src)
|
100 |
else:
|
101 |
return s
|
102 |
|
103 |
return [_format_string(s) for s in strings]
|
104 |
|
105 |
-
def get_url(self, src):
|
106 |
-
return self._inject_language(src, self._urls)
|
107 |
|
108 |
-
def get_manual_dl_files(self, src):
|
109 |
-
return self._inject_language(src, self._manual_dl_files)
|
110 |
|
111 |
-
def get_path(self, src):
|
112 |
-
return self._inject_language(src, self._paths)
|
113 |
|
114 |
|
115 |
# Subsets used in the training sets for various years of WMT.
|
@@ -677,14 +678,14 @@ class Wmt(datasets.GeneratorBasedBuilder):
|
|
677 |
@property
|
678 |
def subsets(self):
|
679 |
"""Subsets that make up each split of the dataset for the language pair."""
|
680 |
-
|
681 |
filtered_subsets = {}
|
682 |
subsets = self._subsets if self.config.subsets is None else self.config.subsets
|
683 |
for split, ss_names in subsets.items():
|
684 |
filtered_subsets[split] = []
|
685 |
for ss_name in ss_names:
|
686 |
dataset = DATASET_MAP[ss_name]
|
687 |
-
if
|
688 |
logger.info("Skipping sub-dataset that does not include language pair: %s", ss_name)
|
689 |
else:
|
690 |
filtered_subsets[split].append(ss_name)
|
@@ -708,7 +709,7 @@ class Wmt(datasets.GeneratorBasedBuilder):
|
|
708 |
yield ex[language]
|
709 |
|
710 |
def _split_generators(self, dl_manager):
|
711 |
-
source,
|
712 |
manual_paths_dict = {}
|
713 |
urls_to_download = {}
|
714 |
for ss_name in itertools.chain.from_iterable(self.subsets.values()):
|
@@ -716,30 +717,30 @@ class Wmt(datasets.GeneratorBasedBuilder):
|
|
716 |
# CzEng1.7 is CzEng1.6 with some blocks filtered out. We must download
|
717 |
# the filtering script so we can parse out which blocks need to be
|
718 |
# removed.
|
719 |
-
urls_to_download[_CZENG17_FILTER.name] = _CZENG17_FILTER.get_url(source)
|
720 |
|
721 |
# get dataset
|
722 |
dataset = DATASET_MAP[ss_name]
|
723 |
-
if dataset.get_manual_dl_files(source):
|
724 |
# TODO(PVP): following two lines skip configs that are incomplete for now
|
725 |
# +++++++++++++++++++++
|
726 |
logger.info("Skipping {dataset.name} for now. Incomplete dataset for {self.config.name}")
|
727 |
continue
|
728 |
# +++++++++++++++++++++
|
729 |
|
730 |
-
manual_dl_files = dataset.get_manual_dl_files(source)
|
731 |
manual_paths = [
|
732 |
os.path.join(os.path.abspath(os.path.expanduser(dl_manager.manual_dir)), fname)
|
733 |
for fname in manual_dl_files
|
734 |
]
|
735 |
assert all(
|
736 |
os.path.exists(path) for path in manual_paths
|
737 |
-
), f"For {dataset.name}, you must manually download the following file(s) from {dataset.get_url(source)} and place them in {dl_manager.manual_dir}: {', '.join(manual_dl_files)}"
|
738 |
|
739 |
# set manual path for correct subset
|
740 |
manual_paths_dict[ss_name] = manual_paths
|
741 |
else:
|
742 |
-
urls_to_download[ss_name] = dataset.get_url(source)
|
743 |
|
744 |
# Download and extract files from URLs.
|
745 |
downloaded_files = dl_manager.download_and_extract(urls_to_download)
|
@@ -759,10 +760,10 @@ class Wmt(datasets.GeneratorBasedBuilder):
|
|
759 |
|
760 |
def _generate_examples(self, split_subsets, extraction_map, with_translation=True):
|
761 |
"""Returns the examples in the raw (text) form."""
|
762 |
-
source,
|
763 |
|
764 |
def _get_local_paths(dataset, extract_dirs):
|
765 |
-
rel_paths = dataset.get_path(source)
|
766 |
if len(extract_dirs) == 1:
|
767 |
extract_dirs = extract_dirs * len(rel_paths)
|
768 |
return [
|
@@ -771,8 +772,8 @@ class Wmt(datasets.GeneratorBasedBuilder):
|
|
771 |
]
|
772 |
|
773 |
def _get_filenames(dataset):
|
774 |
-
rel_paths = dataset.get_path(source)
|
775 |
-
urls = dataset.get_url(source)
|
776 |
if len(urls) == 1:
|
777 |
urls = urls * len(rel_paths)
|
778 |
return [rel_path if rel_path else os.path.basename(url) for url, rel_path in zip(urls, rel_paths)]
|
@@ -782,7 +783,7 @@ class Wmt(datasets.GeneratorBasedBuilder):
|
|
782 |
# +++++++++++++++++++++
|
783 |
dataset = DATASET_MAP[ss_name]
|
784 |
source, _ = self.config.language_pair
|
785 |
-
if dataset.get_manual_dl_files(source):
|
786 |
logger.info(f"Skipping {dataset.name} for now. Incomplete dataset for {self.config.name}")
|
787 |
continue
|
788 |
# +++++++++++++++++++++
|
|
|
53 |
class SubDataset:
|
54 |
"""Class to keep track of information on a sub-dataset of WMT."""
|
55 |
|
56 |
+
def __init__(self, name, url, path="", target=None, sources=None, language_pairs=None, manual_dl_files=None):
|
57 |
"""Sub-dataset of WMT.
|
58 |
|
59 |
Args:
|
|
|
83 |
self._manual_dl_files = manual_dl_files if manual_dl_files else []
|
84 |
self.name = name
|
85 |
self.target = target
|
86 |
+
self.sources = set(sources) if sources else sources
|
87 |
+
self.language_pairs = language_pairs if language_pairs else {(src, target) for src in self.sources}
|
88 |
|
89 |
+
def _inject_language(self, src, tgt, strings):
|
90 |
"""Injects languages into (potentially) template strings."""
|
91 |
+
if (src, tgt) not in self.language_pairs:
|
92 |
+
raise ValueError(f"Invalid source for '{self.name}': ({src}-{tgt})")
|
93 |
|
94 |
def _format_string(s):
|
95 |
+
if "{0}" in s and "{1}" in s and "{src}" in s:
|
96 |
+
return s.format(*sorted([src, tgt]), src=src)
|
97 |
elif "{0}" in s and "{1}" in s:
|
98 |
+
return s.format(*sorted([src, tgt]))
|
99 |
elif "{src}" in s:
|
100 |
+
return s.format(src=src, tgt=tgt)
|
101 |
else:
|
102 |
return s
|
103 |
|
104 |
return [_format_string(s) for s in strings]
|
105 |
|
106 |
+
def get_url(self, src, tgt):
|
107 |
+
return self._inject_language(src=src, tgt=tgt, strings=self._urls)
|
108 |
|
109 |
+
def get_manual_dl_files(self, src, tgt):
|
110 |
+
return self._inject_language(src=src, tgt=tgt, strings=self._manual_dl_files)
|
111 |
|
112 |
+
def get_path(self, src, tgt):
|
113 |
+
return self._inject_language(src=src, tgt=tgt, strings=self._paths)
|
114 |
|
115 |
|
116 |
# Subsets used in the training sets for various years of WMT.
|
|
|
678 |
@property
|
679 |
def subsets(self):
|
680 |
"""Subsets that make up each split of the dataset for the language pair."""
|
681 |
+
language_pair = self.config.language_pair
|
682 |
filtered_subsets = {}
|
683 |
subsets = self._subsets if self.config.subsets is None else self.config.subsets
|
684 |
for split, ss_names in subsets.items():
|
685 |
filtered_subsets[split] = []
|
686 |
for ss_name in ss_names:
|
687 |
dataset = DATASET_MAP[ss_name]
|
688 |
+
if language_pair not in dataset.language_pairs:
|
689 |
logger.info("Skipping sub-dataset that does not include language pair: %s", ss_name)
|
690 |
else:
|
691 |
filtered_subsets[split].append(ss_name)
|
|
|
709 |
yield ex[language]
|
710 |
|
711 |
def _split_generators(self, dl_manager):
|
712 |
+
source, target = self.config.language_pair
|
713 |
manual_paths_dict = {}
|
714 |
urls_to_download = {}
|
715 |
for ss_name in itertools.chain.from_iterable(self.subsets.values()):
|
|
|
717 |
# CzEng1.7 is CzEng1.6 with some blocks filtered out. We must download
|
718 |
# the filtering script so we can parse out which blocks need to be
|
719 |
# removed.
|
720 |
+
urls_to_download[_CZENG17_FILTER.name] = _CZENG17_FILTER.get_url(source, target)
|
721 |
|
722 |
# get dataset
|
723 |
dataset = DATASET_MAP[ss_name]
|
724 |
+
if dataset.get_manual_dl_files(source, target):
|
725 |
# TODO(PVP): following two lines skip configs that are incomplete for now
|
726 |
# +++++++++++++++++++++
|
727 |
logger.info("Skipping {dataset.name} for now. Incomplete dataset for {self.config.name}")
|
728 |
continue
|
729 |
# +++++++++++++++++++++
|
730 |
|
731 |
+
manual_dl_files = dataset.get_manual_dl_files(source, target)
|
732 |
manual_paths = [
|
733 |
os.path.join(os.path.abspath(os.path.expanduser(dl_manager.manual_dir)), fname)
|
734 |
for fname in manual_dl_files
|
735 |
]
|
736 |
assert all(
|
737 |
os.path.exists(path) for path in manual_paths
|
738 |
+
), f"For {dataset.name}, you must manually download the following file(s) from {dataset.get_url(source, target)} and place them in {dl_manager.manual_dir}: {', '.join(manual_dl_files)}"
|
739 |
|
740 |
# set manual path for correct subset
|
741 |
manual_paths_dict[ss_name] = manual_paths
|
742 |
else:
|
743 |
+
urls_to_download[ss_name] = dataset.get_url(source, target)
|
744 |
|
745 |
# Download and extract files from URLs.
|
746 |
downloaded_files = dl_manager.download_and_extract(urls_to_download)
|
|
|
760 |
|
761 |
def _generate_examples(self, split_subsets, extraction_map, with_translation=True):
|
762 |
"""Returns the examples in the raw (text) form."""
|
763 |
+
source, target = self.config.language_pair
|
764 |
|
765 |
def _get_local_paths(dataset, extract_dirs):
|
766 |
+
rel_paths = dataset.get_path(source, target)
|
767 |
if len(extract_dirs) == 1:
|
768 |
extract_dirs = extract_dirs * len(rel_paths)
|
769 |
return [
|
|
|
772 |
]
|
773 |
|
774 |
def _get_filenames(dataset):
|
775 |
+
rel_paths = dataset.get_path(source, target)
|
776 |
+
urls = dataset.get_url(source, target)
|
777 |
if len(urls) == 1:
|
778 |
urls = urls * len(rel_paths)
|
779 |
return [rel_path if rel_path else os.path.basename(url) for url, rel_path in zip(urls, rel_paths)]
|
|
|
783 |
# +++++++++++++++++++++
|
784 |
dataset = DATASET_MAP[ss_name]
|
785 |
source, _ = self.config.language_pair
|
786 |
+
if dataset.get_manual_dl_files(source, target):
|
787 |
logger.info(f"Skipping {dataset.name} for now. Incomplete dataset for {self.config.name}")
|
788 |
continue
|
789 |
# +++++++++++++++++++++
|