From 8e0d8500c895e24d913ee196e19bf6e968284fae Mon Sep 17 00:00:00 2001
From: piotrmp <piotr.m.przybyla@gmail.com>
Date: Tue, 22 Nov 2022 15:52:36 +0100
Subject: [PATCH] Modifications for larger models.
---
src/lambo/examples/run_pretraining.py | 4 +-
src/lambo/learning/model.py | 8 +--
src/lambo/learning/model_pretraining.py | 4 +-
src/lambo/learning/train.py | 4 +-
src/lambo/resources/languages.txt | 65 ++++++++++++------
src/lambo/utils/generate_languages_txt.py | 82 +++++++++++++++++++++++
6 files changed, 138 insertions(+), 29 deletions(-)
create mode 100644 src/lambo/utils/generate_languages_txt.py
diff --git a/src/lambo/examples/run_pretraining.py b/src/lambo/examples/run_pretraining.py
index ddcb05b..a1a14c5 100644
--- a/src/lambo/examples/run_pretraining.py
+++ b/src/lambo/examples/run_pretraining.py
@@ -26,8 +26,8 @@ if __name__=='__main__':
line[0] != '#' and line.split(' ')[1] != '?']
languages = list(dict.fromkeys(languages))
- MAX_DOCUMENTS = 10
- CONTEXT_LEN = 256
+ MAX_DOCUMENTS = 100
+ CONTEXT_LEN = 1024
for language in languages:
if (outpath / ('oscar_' + language + '.pth')).exists():
diff --git a/src/lambo/learning/model.py b/src/lambo/learning/model.py
index 6383d3e..9e69067 100644
--- a/src/lambo/learning/model.py
+++ b/src/lambo/learning/model.py
@@ -7,8 +7,8 @@ class LamboNetwork(Module):
"""
LAMBO neural network model. The network has four layers:
- * embedding layers for characters, representing each as a 32-long vector,
- * bidirectional LSTM layer, taking a concatenation of (1) character embedding and (2) one-hot UTF category vector as input and outputting 2*128-long state vector,
+ * embedding layers for characters, representing each as a 32-long vector (or 64-long),
+ * bidirectional LSTM layer, taking a concatenation of (1) character embedding and (2) one-hot UTF category vector as input and outputting 2*128-long state vector (or 2*256),
* dense linear layer, converting LSTM state vectors to class scores
* softmax layer, computing probability of eight events for any character:
@@ -34,8 +34,8 @@ class LamboNetwork(Module):
self.embedding_layer = copy.deepcopy(pretrained.embedding_layer)
self.lstm_layer = copy.deepcopy(pretrained.lstm_layer)
else:
- self.embedding_layer = Embedding(len(dict), 32, dict['<PAD>'])
- self.lstm_layer = LSTM(input_size=self.embedding_layer.embedding_dim + utf_categories_num, hidden_size=128,
+ self.embedding_layer = Embedding(len(dict), 64, dict['<PAD>'])
+ self.lstm_layer = LSTM(input_size=self.embedding_layer.embedding_dim + utf_categories_num, hidden_size=256,
batch_first=True,
bidirectional=True)
self.linear_layer = Linear(self.lstm_layer.hidden_size * 2, 8)
diff --git a/src/lambo/learning/model_pretraining.py b/src/lambo/learning/model_pretraining.py
index e37d08b..fa07693 100644
--- a/src/lambo/learning/model_pretraining.py
+++ b/src/lambo/learning/model_pretraining.py
@@ -17,8 +17,8 @@ class LamboPretrainingNetwork(Module):
"""
super(LamboPretrainingNetwork, self).__init__()
self.max_len = max_len
- self.embedding_layer = Embedding(len(dict), 32, dict['<PAD>'])
- self.lstm_layer = LSTM(input_size=self.embedding_layer.embedding_dim + utf_categories_num, hidden_size=128,
+ self.embedding_layer = Embedding(len(dict), 64, dict['<PAD>'])
+ self.lstm_layer = LSTM(input_size=self.embedding_layer.embedding_dim + utf_categories_num, hidden_size=256,
batch_first=True,
bidirectional=True)
self.linear_layer = Linear(self.lstm_layer.hidden_size * 2, len(dict))
diff --git a/src/lambo/learning/train.py b/src/lambo/learning/train.py
index c45c500..f33c88b 100644
--- a/src/lambo/learning/train.py
+++ b/src/lambo/learning/train.py
@@ -114,7 +114,7 @@ def train_new_and_save(model_name, treebank_path, save_path, epochs=10):
BATCH_SIZE = 32
print("Initiating the model.")
- MAX_LEN = 100
+ MAX_LEN = 1024
dict, train_dataloader, test_dataloader = prepare_dataloaders_withdict([train_doc, dev_doc], [test_doc],
MAX_LEN,
BATCH_SIZE)
@@ -156,7 +156,7 @@ def train_pretrained_and_save(language, treebank_path, save_path, pretrained_pat
train_doc, dev_doc, test_doc = read_treebank(treebank_path, True)
print("Initiating the model.")
- MAX_LEN = 100
+ MAX_LEN = 1024
model = LamboNetwork(MAX_LEN, dict, len(utf_category_dictionary), pretrained=pretrained_model)
print("Preparing data")
diff --git a/src/lambo/resources/languages.txt b/src/lambo/resources/languages.txt
index ee5981c..9e213ff 100644
--- a/src/lambo/resources/languages.txt
+++ b/src/lambo/resources/languages.txt
@@ -1,17 +1,20 @@
# Format: <UD training corpus> <ISO 639-1 code (for OSCAR pretraining)> <Language name> <Recommended (chosen by size)>
UD_Afrikaans-AfriBooms af Afrikaans
-UD_Ancient_Greek-Perseus ? Ancient_Greek
UD_Ancient_Greek-PROIEL ? Ancient_Greek *
-UD_Arabic-PADT ar Arabic
-UD_Armenian-ArmTDP hy Armenian
-UD_Western_Armenian-ArmTDP hy Western_Armenian
+UD_Ancient_Greek-Perseus ? Ancient_Greek
+UD_Ancient_Hebrew-PTNK ? Ancient_Hebrew
+UD_Arabic-NYUAD ar Arabic
+UD_Arabic-PADT ar Arabic *
+UD_Armenian-ArmTDP hy Armenian *
+UD_Armenian-BSUT hy Armenian
UD_Basque-BDT eu Basque
UD_Belarusian-HSE be Belarusian
UD_Bulgarian-BTB bg Bulgarian
UD_Catalan-AnCora ca Catalan
-UD_Chinese-GSD zh Chinese *
-UD_Chinese-GSDSimp zh Chinese
+UD_Chinese-GSD zh Chinese
+UD_Chinese-GSDSimp zh Chinese *
UD_Classical_Chinese-Kyoto ? Classical_Chinese
+UD_Coptic-Scriptorium ? Coptic
UD_Croatian-SET hr Croatian
UD_Czech-CAC cs Czech
UD_Czech-CLTT cs Czech
@@ -21,54 +24,77 @@ UD_Danish-DDT da Danish
UD_Dutch-Alpino nl Dutch *
UD_Dutch-LassySmall nl Dutch
UD_English-Atis en English
+UD_English-ESL en English
UD_English-EWT en English *
UD_English-GUM en English
+UD_English-GUMReddit en English
UD_English-LinES en English
UD_English-ParTUT en English
UD_Estonian-EDT et Estonian *
UD_Estonian-EWT et Estonian
+UD_Faroese-FarPaHC fo Faroese
UD_Finnish-FTB fi Finnish
UD_Finnish-TDT fi Finnish *
+UD_French-FTB fr French
UD_French-GSD fr French *
UD_French-ParTUT fr French
+UD_French-ParisStories fr French
UD_French-Rhapsodie fr French
UD_French-Sequoia fr French
UD_Galician-CTG gl Galician
UD_German-GSD de German
UD_German-HDT de German *
+UD_Gothic-PROIEL ? Gothic
UD_Greek-GDT el Greek
-UD_Hebrew-HTB he Hebrew
+UD_Hebrew-HTB he Hebrew *
+UD_Hebrew-IAHLTwiki he Hebrew
UD_Hindi-HDTB hi Hindi
+UD_Hindi_English-HIENCS ? Hindi_English
UD_Hungarian-Szeged hu Hungarian
+UD_Icelandic-GC is Icelandic
UD_Icelandic-IcePaHC is Icelandic *
UD_Icelandic-Modern is Icelandic
UD_Indonesian-GSD id Indonesian
UD_Irish-IDT ga Irish
UD_Italian-ISDT it Italian *
+UD_Italian-MarkIT it Italian
UD_Italian-ParTUT it Italian
UD_Italian-PoSTWITA it Italian
UD_Italian-TWITTIRO it Italian
UD_Italian-VIT it Italian
-UD_Japanese-GSD ja Japanese *
+UD_Japanese-BCCWJ ja Japanese *
+UD_Japanese-BCCWJLUW ja Japanese
+UD_Japanese-GSD ja Japanese
UD_Japanese-GSDLUW ja Japanese
-UD_Korean-Kaist ko Korean
+UD_Korean-GSD ko Korean
+UD_Korean-Kaist ko Korean *
UD_Latin-ITTB la Latin *
UD_Latin-LLCT la Latin
UD_Latin-PROIEL la Latin
UD_Latin-UDante la Latin
UD_Latvian-LVTB lv Latvian
-UD_Lithuanian-ALKSNIS lt Lithuanian
+UD_Lithuanian-ALKSNIS lt Lithuanian *
+UD_Lithuanian-HSE lt Lithuanian
UD_Maltese-MUDT mt Maltese
-UD_Norwegian-Bokmaal no Norwegian_Bokmål
-UD_Norwegian-Nynorsk nn Norwegian_Nynorsk *
-UD_Norwegian-NynorskLIA nn Norwegian_Nynorsk
+UD_Marathi-UFAL mr Marathi
+UD_Naija-NSC ? Naija
+UD_Norwegian-Bokmaal no Norwegian
+UD_Norwegian-Nynorsk nn Norwegian *
+UD_Norwegian-NynorskLIA nn Norwegian
+UD_Old_Church_Slavonic-PROIEL ? Old_Church_Slavonic
+UD_Old_East_Slavic-Birchbark ? Old_East_Slavic
+UD_Old_East_Slavic-RNC ? Old_East_Slavic
+UD_Old_East_Slavic-TOROT ? Old_East_Slavic *
UD_Old_French-SRCMF ? Old_French
UD_Persian-PerDT fa Persian *
UD_Persian-Seraji fa Persian
-UD_Polish-PDB pl Polish *
UD_Polish-LFG pl Polish
+UD_Polish-PDB pl Polish *
+UD_Pomak-Philotis ? Pomak
UD_Portuguese-Bosque pt Portuguese
-UD_Portuguese-GSD pt Portuguese *
+UD_Portuguese-CINTIL pt Portuguese *
+UD_Portuguese-GSD pt Portuguese
+UD_Portuguese-PetroGold pt Portuguese
UD_Romanian-Nonstandard ro Romanian *
UD_Romanian-RRT ro Romanian
UD_Romanian-SiMoNERo ro Romanian
@@ -83,8 +109,9 @@ UD_Spanish-AnCora es Spanish *
UD_Spanish-GSD es Spanish
UD_Swedish-LinES sv Swedish
UD_Swedish-Talbanken sv Swedish *
+UD_Swedish_Sign_Language-SSLC ? Swedish_Sign_Language
UD_Tamil-TTB ta Tamil
-UD_Telugu-MTG te Telegu
+UD_Telugu-MTG te Telugu
UD_Turkish-Atis tr Turkish
UD_Turkish-BOUN tr Turkish
UD_Turkish-FrameNet tr Turkish
@@ -92,11 +119,11 @@ UD_Turkish-IMST tr Turkish
UD_Turkish-Kenet tr Turkish
UD_Turkish-Penn tr Turkish *
UD_Turkish-Tourism tr Turkish
+UD_Turkish_German-SAGT ? Turkish_German
UD_Ukrainian-IU uk Ukrainian
UD_Urdu-UDTB ur Urdu
UD_Uyghur-UDT ug Uyghur
UD_Vietnamese-VTB vi Vietnamese
UD_Welsh-CCG cy Welsh
-#NKJP_Polish-byName pl Polish
-#NKJP_Polish-byType pl Polish
-#UD_Polish-PDBnoMW pl Polish
\ No newline at end of file
+UD_Western_Armenian-ArmTDP hy Western_Armenian
+UD_Wolof-WTB wo Wolof
\ No newline at end of file
diff --git a/src/lambo/utils/generate_languages_txt.py b/src/lambo/utils/generate_languages_txt.py
new file mode 100644
index 0000000..d3982e4
--- /dev/null
+++ b/src/lambo/utils/generate_languages_txt.py
@@ -0,0 +1,82 @@
+from pathlib import Path
+
+old_languages_txt = ''
+new_ud_treebanks = ''
+
+codedict = {}
+for line in open(old_languages_txt):
+ if line.startswith('#'):
+ continue
+ parts = line.strip().split(' ')
+ lang = parts[2]
+ code = parts[1]
+ codedict[lang] = code
+
+ud11path = Path(new_ud_treebanks)
+
+subdirs = [x for x in ud11path.iterdir() if x.is_dir()]
+subdirs.sort()
+
+sizes = {}
+
+for subdir in subdirs:
+ hasTrain = False
+ hasDev = False
+ hasTest = False
+ trainfile = None
+ for file in subdir.iterdir():
+ if file.name.endswith('train.txt'):
+ hasTrain = True
+ trainfile = file
+ elif file.name.endswith('test.txt'):
+ hasTest = True
+ elif file.name.endswith('dev.txt'):
+ hasDev = True
+ if (not hasTrain) or (not hasTest) or (not hasDev):
+ continue
+ treebank_name = subdir.name
+ language_name = treebank_name[3:].split('-')[0]
+ code = '@@@@@'
+ if language_name in codedict:
+ code = codedict[language_name]
+ if language_name not in sizes:
+ sizes[language_name] = {}
+ sizes[language_name][treebank_name] = trainfile.stat().st_size
+
+for language_name in sizes:
+ maxlen = 0
+ best = None
+ for treebank_name in sizes[language_name]:
+ if sizes[language_name][treebank_name] > maxlen:
+ best = treebank_name
+ maxlen = sizes[language_name][treebank_name]
+ if len(sizes[language_name]) > 1:
+ sizes[language_name]['preferred'] = best
+
+print(
+ "# Format: <UD training corpus> <ISO 639-1 code (for OSCAR pretraining)> <Language name> <Recommended (chosen by size)>")
+for subdir in subdirs:
+ hasTrain = False
+ hasDev = False
+ hasTest = False
+ trainfile = None
+ for file in subdir.iterdir():
+ if file.name.endswith('train.txt'):
+ hasTrain = True
+ trainfile = file
+ elif file.name.endswith('test.txt'):
+ hasTest = True
+ elif file.name.endswith('dev.txt'):
+ hasDev = True
+ if (not hasTrain) or (not hasTest) or (not hasDev):
+ continue
+ treebank_name = subdir.name
+ language_name = treebank_name[3:].split('-')[0]
+ code = '@@@@@'
+ if language_name in codedict:
+ code = codedict[language_name]
+ preferred = ''
+ if 'preferred' in sizes[language_name] and sizes[language_name]['preferred'] == treebank_name:
+ preferred = ' *'
+ print(treebank_name + ' ' + code + ' ' + language_name + preferred)
+
--
GitLab