artcraft/models/tts/NeMo/tutorials/nlp/02_NLP_Tokenizers.ipynb

In [ ]:

BRANCH = 'v1.0.2'

In [ ]:

"""
You can run either this notebook locally (if you have all the dependencies and a GPU) or on Google Colab.

Instructions for setting up Colab are as follows:
1. Open a new Python 3 notebook.
2. Import this notebook from GitHub (File -> Upload Notebook -> "GITHUB" tab -> copy/paste GitHub URL)
3. Connect to an instance with a GPU (Runtime -> Change runtime type -> select "GPU" for hardware accelerator)
4. Run this cell to set up dependencies.
"""
# If you're using Google Colab and not running locally, run this cell

# install NeMo
BRANCH = 'v1.0.2'
!python -m pip install git+https://github.com/NVIDIA/NeMo.git@$BRANCH#egg=nemo_toolkit[nlp]

In [ ]:

import os
import wget
from nemo.collections import nlp as nemo_nlp
from nemo.collections import common as nemo_common
from omegaconf import OmegaConf

Tokenizers Background

For Natural Language Processing, tokenization is an essential part of data preprocessing. It is the process of splitting a string into a list of tokens. One can think of token as parts like a word is a token in a sentence. Depending on the application, different tokenizers are more suitable than others.

For example, a WordTokenizer that splits the string on any whitespace, would tokenize the following string

"My first program, Hello World." -> ["My", "first", "program,", "Hello", "World."]

To turn the tokens into numerical model input, the standard method is to use a vocabulary and one-hot vectors for word embeddings. If a token appears in the vocabulary, its index is returned, if not the index of the unknown token is returned to mitigate out-of-vocabulary (OOV).

Tokenizers in NeMo

In NeMo, we support the most used tokenization algorithms. We offer a wrapper around Hugging Faces's AutoTokenizer - a factory class that gives access to all Hugging Face tokenizers. This includes particularly all BERT-like model tokenizers, such as BertTokenizer, AlbertTokenizer, RobertaTokenizer, GPT2Tokenizer. Apart from that, we also support other tokenizers such as WordTokenizer, CharTokenizer, and Google's SentencePieceTokenizer.

We make sure that all tokenizers are compatible with BERT-like models, e.g. BERT, Roberta, Albert, and Megatron. For that, we provide a high-level user API get_tokenizer(), which allows the user to instantiate a tokenizer model with only four input arguments:

• tokenizer_name: str
• tokenizer_model: Optional[str] = None
• vocab_file: Optional[str] = None
• special_tokens: Optional[Dict[str, str]] = None

Hugging Face and Megatron tokenizers (which uses Hugging Face underneath) can be automatically instantiated by only tokenizer_name, which downloads the corresponding vocab_file from the internet.

For SentencePieceTokenizer, WordTokenizer, and CharTokenizers tokenizer_model or/and vocab_file can be generated offline in advance using scripts/tokenizers/process_asr_text_tokenizer.py

The tokenizers in NeMo are designed to be used interchangeably, especially when used in combination with a BERT-based model.

Let's take a look at the list of available tokenizers:

In [ ]:

nemo_nlp.modules.get_tokenizer_list()

Hugging Face AutoTokenizer

In [ ]:

# instantiate tokenizer wrapper using pretrained model name only
tokenizer1 = nemo_nlp.modules.get_tokenizer(tokenizer_name="bert-base-cased")

# the wrapper has a reference to the original HuggingFace tokenizer
print(tokenizer1.tokenizer)

In [ ]:

# check vocabulary (this can be very long)
print(tokenizer1.tokenizer.vocab)

In [ ]:

# show all special tokens if it has any
print(tokenizer1.tokenizer.all_special_tokens)

In [ ]:

# instantiate tokenizer using custom vocabulary
vocab_file = "myvocab.txt"
vocab = ["he", "llo", "world"]
with open(vocab_file, 'w') as vocab_fp:
  vocab_fp.write("\n".join(vocab))

In [ ]:

tokenizer2 = nemo_nlp.modules.get_tokenizer(tokenizer_name="bert-base-cased", vocab_file=vocab_file)

In [ ]:

# Since we did not overwrite special tokens they should be the same as before
print(tokenizer1.tokenizer.all_special_tokens == tokenizer2.tokenizer.all_special_tokens )

Adding Special tokens

We do not recommend overwriting special tokens for Hugging Face pretrained models, since these are the commonly used default values.

If a user still wants to overwrite the special tokens, specify some of the following keys:

In [ ]:

special_tokens_dict = {"unk_token": "<UNK>", 
                       "sep_token": "<SEP>", 
                       "pad_token": "<PAD>", 
                       "bos_token": "<CLS>", 
                       "mask_token": "<MASK>",
                       "eos_token": "<SEP>",
                       "cls_token": "<CLS>"}
tokenizer3 = nemo_nlp.modules.get_tokenizer(tokenizer_name="bert-base-cased",
                                            vocab_file=vocab_file,
                                            special_tokens=special_tokens_dict)

# print newly set special tokens
print(tokenizer3.tokenizer.all_special_tokens)
# the special tokens should be different from the previous special tokens
print(tokenizer3.tokenizer.all_special_tokens != tokenizer1.tokenizer.all_special_tokens )

Notice, that if you specify tokens that were not previously included in the tokenizer's vocabulary file, new tokens will be added to the vocabulary file. You will see a message like this:

['<MASK>', '<CLS>', '<SEP>', '<PAD>', '<SEP>', '<CLS>', '<UNK>'] will be added to the vocabulary. Please resize your model accordingly

In [ ]:

# A safer way to add special tokens is the following:

# define your model
pretrained_model_name = 'bert-base-uncased'
model = nemo_nlp.modules.get_lm_model(pretrained_model_name=pretrained_model_name)

# define pretrained tokenizer
tokenizer_default = nemo_nlp.modules.get_tokenizer(tokenizer_name=pretrained_model_name)

In [ ]:

tokenizer_default.text_to_tokens('<MY_NEW_TOKEN> and another word')

As you can see in the above, the tokenizer splits <MY_NEW_TOKEN> into subtokens. Let's add this to the special tokens to make sure the tokenizer does not split this into subtokens.

In [ ]:

special_tokens = {'bos_token': '<BOS>',
                  'cls_token': '<CSL>',
                  'additional_special_tokens': ['<MY_NEW_TOKEN>', '<ANOTHER_TOKEN>']}
tokenizer_default.add_special_tokens(special_tokens_dict=special_tokens)

# resize your model so that the embeddings for newly added tokens are updated during training/finetuning
model.resize_token_embeddings(tokenizer_default.vocab_size)

# let's make sure the tokenizer doesn't split our special tokens into subtokens
tokenizer_default.text_to_tokens('<MY_NEW_TOKEN> and another word')

Now, the model doesn't break down our special token into the subtokens.

Megatron model tokenizer

In [ ]:

# Megatron tokenizers are instances of the Hugging Face BertTokenizer. 
tokenizer4 = nemo_nlp.modules.get_tokenizer(tokenizer_name="megatron-bert-cased")

Train custom tokenizer model and vocabulary from text file

We use the scripts/tokenizers/process_asr_text_tokenizer.py script to create a custom tokenizer model with its own vocabulary from an input file

In [ ]:

# download tokenizer script
script_file = "process_asr_text_tokenizer.py"

if not os.path.exists(script_file):
    print('Downloading script file...')
    wget.download(f'https://raw.githubusercontent.com/NVIDIA/NeMo/{BRANCH}/scripts/tokenizers/process_asr_text_tokenizer.py')
else:
    print ('Script already exists')

In [ ]:

# Let's prepare some small text data for the tokenizer
data_text = "NeMo is a toolkit for creating Conversational AI applications. \
NeMo toolkit makes it possible for researchers to easily compose complex neural network architectures \
for conversational AI using reusable components - Neural Modules. \
Neural Modules are conceptual blocks of neural networks that take typed inputs and produce typed outputs. \
Such modules typically represent data layers, encoders, decoders, language models, loss functions, or methods of combining activations. \
The toolkit comes with extendable collections of pre-built modules and ready-to-use models for automatic speech recognition (ASR), \
natural language processing (NLP) and text synthesis (TTS). \
Built for speed, NeMo can utilize NVIDIA's Tensor Cores and scale out training to multiple GPUs and multiple nodes."

In [ ]:

# Write the text data into a file
data_file="data.txt"

with open(data_file, 'w') as data_fp:
  data_fp.write(data_text)

In [ ]:

# Some additional parameters for the tokenizer
# To tokenize at unigram, char or word boundary instead of using bpe, change --spe_type accordingly. 
# More details see https://github.com/google/sentencepiece#train-sentencepiece-model

tokenizer_spe_type = "bpe"  # <-- Can be `bpe`, `unigram`, `word` or `char`
vocab_size = 32

In [ ]:

! python process_asr_text_tokenizer.py --data_file=$data_file --data_root=. --vocab_size=$vocab_size --tokenizer=spe --spe_type=$tokenizer_spe_type

In [ ]:

# See created tokenizer model and vocabulary
spe_model_dir=f"tokenizer_spe_{tokenizer_spe_type}_v{vocab_size}"
! ls $spe_model_dir

Use custom tokenizer for data preprocessing

Example: SentencePiece for BPE

In [ ]:

# initialize tokenizer with created tokenizer model, which inherently includes the vocabulary and specify optional special tokens
tokenizer_spe = nemo_nlp.modules.get_tokenizer(tokenizer_name="sentencepiece", tokenizer_model=spe_model_dir+"/tokenizer.model", special_tokens=special_tokens_dict)

# specified special tokens are added to the vocabuary
print(tokenizer_spe.vocab_size)

Example: WordTokenizer from Vocabulary

In [ ]:

# If you want to use a simple tokenizer like WordTokenizer without first generating the tokenizer.model first 
# we provide the alternative class WordTokenizer or CharTokenizer that takes a user vocabulary as input

# initialize tokenizer with vocabulary and specify optional special tokens
tokenizer_word = nemo_nlp.modules.get_tokenizer(tokenizer_name="word", vocab_file=vocab_file, special_tokens=special_tokens_dict)

# specified special tokens are added to the vocabulary
print(tokenizer_word.vocab_size)

Using any tokenizer to tokenize text into BERT compatible input

In [ ]:

text="hello world"

# create tokens
tokenized = [tokenizer_word.bos_token] + tokenizer_word.text_to_tokens(text) + [tokenizer_word.eos_token]
print(tokenized)

# turn token into input_ids for a neural model, such as BERTModule

print(tokenizer_word.tokens_to_ids(tokenized))