T5 Russian Spell

🚀 t5-russian-spell

A model for correcting text from recognized audio. You can feed the results of my audio recognition model into this model.

🚀 Quick Start

This model, t5-russian-spell, is designed to correct text from recognized audio. You can use the results of the audio recognition model UrukHan/wav2vec2-russian as input for this model.

Example Comparison

Datasets for Training

• UrukHan/t5-russian-spell_I
• UrukHan/t5-russian-spell_II
• UrukHan/t5-russian-spell_III

💻 Usage Examples

Basic Usage

# Install the transformers library
!pip install transformers

# Import libraries
from transformers import AutoModelForSeq2SeqLM, T5TokenizerFast

# Set the name of the selected model from the hub
MODEL_NAME = 'UrukHan/t5-russian-spell'
MAX_INPUT = 256

# Load the model and tokenizer
tokenizer = T5TokenizerFast.from_pretrained(MODEL_NAME)
model = AutoModelForSeq2SeqLM.from_pretrained(MODEL_NAME)

# Input data (can be an array of phrases or text)
input_sequences = ['сеглдыя хорош ден', 'когд а вы прдет к нам в госи']   # or you can use single phrases:  input_sequences = 'сеглдыя хорош ден'

task_prefix = "Spell correct: "                 # Tokenize the data
if type(input_sequences) != list: input_sequences = [input_sequences]
encoded = tokenizer(
  [task_prefix + sequence for sequence in input_sequences],
  padding="longest",
  max_length=MAX_INPUT,
  truncation=True,
  return_tensors="pt",
)

predicts = model.generate(encoded)    # # Prediction

tokenizer.batch_decode(predicts, skip_special_tokens=True)  # Decode the data

Advanced Usage

You can use the following Colab notebook to run the training and save the model to your own repository on the Hugging Face Hub: Colab Notebook

# Install libraries
!pip install datasets
!apt install git-lfs
!pip install transformers
!pip install sentencepiece 
!pip install rouge_score

# Import libraries
import numpy as np
from datasets import Dataset
import tensorflow as 
import nltk
from transformers import T5TokenizerFast, Seq2SeqTrainingArguments, Seq2SeqTrainer, AutoModelForSeq2SeqLM, DataCollatorForSeq2Seq
import torch
from transformers.optimization import Adafactor, AdafactorSchedule
from datasets import load_dataset, load_metric

# Load parameters
raw_datasets = load_dataset("xsum")
metric = load_metric("rouge")
nltk.download('punkt')

# Enter your Hugging Face Hub key
from huggingface_hub import notebook_login
notebook_login()

# Define parameters
REPO = "t5-russian-spell"  # Enter the name of your repository
MODEL_NAME = "UrukHan/t5-russian-spell" # Enter the name of the selected model from the hub
MAX_INPUT = 256  # Enter the maximum length of input data in tokens (the length of input phrases in words (you can count half a word as a token))
MAX_OUTPUT  = 256 # Enter the maximum length of predictions in tokens (you can reduce it for summarization tasks or other tasks where the output is shorter)
BATCH_SIZE = 8 
DATASET = 'UrukHan/t5-russian-spell_I'   # Enter the name of the dataset

# Load the dataset. I will describe the use of other types of data below
data = load_dataset(DATASET)

# Load the model and tokenizer
tokenizer = T5TokenizerFast.from_pretrained(MODEL_NAME)
model = AutoModelForSeq2SeqLM.from_pretrained(MODEL_NAME)

model.config.max_length = MAX_OUTPUT  # By default, it is 20, so the output sequences are truncated in all models
# Comment out after the first save to the repository. It is optional
tokenizer.push_to_hub(repo_name) 

train = data['train']
test = data['test'].train_test_split(0.02)['test']  # I reduced the test set so that I don't have to wait too long for the error calculation between epochs

data_collator = DataCollatorForSeq2Seq(tokenizer, model=model) #return_tensors="tf"

def compute_metrics(eval_pred):
  predictions, labels = eval_pred
  decoded_preds = tokenizer.batch_decode(predictions, skip_special_tokens=True)
  # Replace -100 in the labels as we can't decode them.
  labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
  decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)
  
  # Rouge expects a newline after each sentence
  decoded_preds = ["\n".join(nltk.sent_tokenize(pred.strip())) for pred in decoded_preds]
  decoded_labels = ["\n".join(nltk.sent_tokenize(label.strip())) for label in decoded_labels]
  
  result = metric.compute(predictions=decoded_preds, references=decoded_labels, use_stemmer=True)
  # Extract a few results
  result = {key: value.mid.fmeasure * 100 for key, value in result.items()}
  
  # Add mean generated length
  prediction_lens = [np.count_nonzero(pred != tokenizer.pad_token_id) for pred in predictions]
  result["gen_len"] = np.mean(prediction_lens)
  
  return {k: round(v, 4) for k, v in result.items()}
  
training_args = Seq2SeqTrainingArguments(
  output_dir = REPO,
  #overwrite_output_dir=True,
  evaluation_strategy='steps',
  #learning_rate=2e-5,
  eval_steps=5000,
  save_steps=5000,
  num_train_epochs=1,
  predict_with_generate=True,
  per_device_train_batch_size=BATCH_SIZE,
  per_device_eval_batch_size=BATCH_SIZE,
  fp16=True,
  save_total_limit=2,
  #generation_max_length=256,
  #generation_num_beams=4,
  weight_decay=0.005,
  #logging_dir='logs',
  push_to_hub=True,
)

# Manually select the optimizer. The original architecture of T5 uses the Adafactor optimizer
optimizer = Adafactor(
    model.parameters(),
    lr=1e-5,
    eps=(1e-30, 1e-3),
    clip_threshold=1.0,
    decay_rate=-0.8,
    beta1=None,
    weight_decay=0.0,
    relative_step=False,
    scale_parameter=False,
    warmup_init=False,
)
lr_scheduler = AdafactorSchedule(optimizer)

trainer = Seq2SeqTrainer(
  model=model,
  args=training_args,
  train_dataset = train,
  eval_dataset = test,
  optimizers = (optimizer, lr_scheduler),
  tokenizer = tokenizer,
  compute_metrics=compute_metrics
)

trainer.train()

trainer.push_to_hub()

Example of Converting Arrays for this Network

input_data = ['удач почти отнее отвернулась', 'в хааоде проведения чемпиониавта мира дветысячивосемнандцтая лгодаа']
output_data = ['Удача почти от нее отвернулась', 'в ходе проведения чемпионата мира две тысячи восемнадцатого года']

# Tokenize the input data
task_prefix = "Spell correct: "
input_sequences = input_data 
encoding = tokenizer(
  [task_prefix + sequence for sequence in input_sequences],
  padding="longest",
  max_length=MAX_INPUT,
  truncation=True,
  return_tensors="pt",
)
input_ids, attention_mask = encoding.input_ids, encoding.attention_mask

# Tokenize the output data
target_encoding = tokenizer(output_data, padding="longest", max_length=MAX_OUTPUT, truncation=True)
labels = target_encoding.input_ids
# replace padding token id's of the labels by -100
labels = torch.tensor(labels)
labels[labels == tokenizer.pad_token_id] = -100

# Convert our data to the dataset format   

data = Dataset.from_pandas(pd.DataFrame({'input_ids': list(np.array(input_ids)), 'attention_mask': list(np.array(attention_mask)), 'labels': list(np.array(labels))}))
data = data.train_test_split(0.02)
# and we will get the input for our network for our trainer:   train_dataset = data['train'],  eval_dataset = data['test']