u107/nemo

list models

QuartzNet15x5

---

• 2023-04-28
  • language: ru
  • from_pretrained: stt_en_quartznet15x5
  • Epoch: 50
  • Step: 738649
  • val_loss: 4.19466 (best 4.19466)
  • model: quartznet15x5/2023-04-28_00-24-37/stt_ru_quartznet15x5.nemo
  • last checkpoint: quartznet15x5/2023-04-28_00-24-37/stt_ru_quartznet15x5--val_loss=4.2732-epoch=50-last.ckpt
  • Dataset train: 709100 files totalling 680.21 hours
  • Dataset validation: 14670 files totalling 15.63 hours
• 2023-04-04
  • language: uk
  • from_pretrained: stt_en_quartznet15x5
  • Epoch: 34
  • Step: 566076
  • val_loss: 1.90886 (best 1.89907)
  • model: quartznet15x5/2023-04-04_02-57-02/stt_uk_quartznet15x5.nemo
  • last checkpoint: quartznet15x5/2023-04-04_02-57-02/QuartzNet15x5--val_loss=1.9543-epoch=34-last.ckpt
  • Dataset train: 1042635 files totalling 1028.14 hours
  • Dataset validation: 50640 files totalling 50.71 hours
• 2023-02-23
  • language: uk-RU
  • from_pretrained: stt_en_quartznet15x5
  • Epoch: 7
  • Step: 217287
  • val_loss: 1.66200 (best 1.66200)
  • model: quartznet15x5/2023-02-23_11-43-23/stt_uk_ru_quartznet15x5.nemo
  • last checkpoint: quartznet15x5/2023-02-23_11-43-23/QuartzNet15x5--val_loss=1.6620-epoch=7-last.ckpt
  • Dataset train: 1738252 files totalling 1645.74 hours
  • Dataset validation: 11373 files totalling 12.63 hours

Jasper10x5

---

• 2023-04-07
  • language: uk
  • from_pretrained: stt_en_jasper10x5dr
  • Epoch: 17
  • Step: 293256
  • val_loss: 2.10919 (best 2.10919)
  • model: jasper10x5/2023-04-07_05-01-29/stt_uk_jasper10x5.nemo
  • last checkpoint: jasper10x5/2023-04-07_05-01-29/Jasper10x5--val_loss=2.1092-epoch=17-last.ckpt
  • Dataset train: 1042635 files totalling 1028.14 hours
  • Dataset validation: 50640 files totalling 50.71 hours
• 2023-02-22
  • language: uk-RU
  • from_pretrained: stt_en_jasper10x5dr
  • Epoch: 4
  • Step: 271604
  • val_loss: 1.76058 (best 1.76058)
  • model: jasper10x5/2023-02-22_06-46-21/stt_uk_ru_jasper10x5.nemo
  • last checkpoint: jasper10x5/2023-02-22_06-46-21/Jasper10x5--val_loss=1.7606-epoch=4-last.ckpt
  • Dataset train: 1738252 files totalling 1645.74 hours
  • Dataset validation: 11373 files totalling 12.63 hours

Punctuation and Capitalization Model

punctuation_uk_bert

---

• 2023-06-28
  • language: uk
  • Epoch: 3
  • val_loss: 0.0319
  • last checkpoint: punctuation_uk_bert/2023-06-28_04-10-06/punctuation_uk_bert--val_loss=0.0385-epoch=3-last.ckpt

punctuation_ru_bert

---

• 2023-06-28
  • language: rU
  • Epoch: 5
  • val_loss: 0.0196
  • last checkpoint: punctuation_ru_bert/2023-06-28_15-12-29/punctuation_ru_bert--val_loss=0.0712-epoch=5-last.ckpt

Використання ASR:

from huggingface_hub import hf_hub_download
from nemo.collections.asr.models import ASRModel
from omegaconf import OmegaConf


# filename_model = 'quartznet15x5/2023-02-23_11-43-23/stt_uk_ru_quartznet15x5.nemo'
# filename_model = 'quartznet15x5/2023-04-04_02-57-02/stt_uk_quartznet15x5.nemo'
# filename_model = 'jasper10x5/2023-02-22_06-46-21/stt_uk_ru_jasper10x5.nemo'
filename_model = 'jasper10x5/2023-04-07_05-01-29/stt_uk_jasper10x5.nemo'

model_path = hf_hub_download(
  repo_id="u107/nemo",
  filename=filename_model,
  repo_type="model",
  cache_dir='/content',
  use_auth_token='hf_...')

nemo_model = ASRModel.restore_from(
  restore_path=model_path,
  # map_location='CPU',
  # return_config=True,
)  # type: ASRModel

# зберегти файл налаштувань
OmegaConf.save(nemo_model.cfg, 'model_config.yaml')

# відобразити версію nemo на якій було навчання
print(nemo_model.cfg.nemo_version)

# відобразити алфавіт
print(nemo_model.cfg.labels)

Використання NLP:

from huggingface_hub import hf_hub_download
from nemo.collections import nlp as nemo_nlp


# filename_model = 'punctuation_ru_bert/2023-06-28_15-12-29/punctuation_ru_bert.nemo'
filename_model = 'punctuation_uk_bert/2023-06-28_04-10-06/punctuation_uk_bert.nemo'

model_path = hf_hub_download(
  repo_id="u107/nemo",
  filename=filename_model,
  repo_type="model",
  cache_dir='/content',
  use_auth_token='hf_...')

pretrained_model = nemo_nlp.models.PunctuationCapitalizationModel.restore_from(
  restore_path=model_path
)

queries = [
  'ніхто ж не каже що б давати так',
]

result = pretrained_model.add_punctuation_capitalization(queries=queries)

print(result)