YouTokenToMe
YouTokenToMe is an unsupervised text tokenizer focused on computational efficiency. It currently implements fast Byte Pair Encoding (BPE) [Sennrich et al.]. Our implementation is much faster in training and tokenization than Hugging Face, fastBPE and SentencePiece. In some test cases, it is 90 times faster. Check out our benchmark results.
Key advantages:
- Multithreading for training and tokenization
- The algorithm has
O(N)complexity, whereNis the length of training data - Highly efficient implementation in C++
- Python wrapper and command-line interface
Extra features:
- BPE-dropout (as described in Provilkov et al, 2019)
As well as in the algorithm from the original paper, ours does not consider tokens that cross word boundaries. Just like in SentencePiece, all space symbols were replaced by meta symbol "▁" (U+2581). It allows sequences of tokens to be converted back to text and for word boundaries to be restored.
For example, the phrase Blazingly fast tokenization! can be tokenized into
['▁Bl', 'az', 'ingly', '▁fast', '▁token', 'ization', '!']
Installation
pip install youtokentome
Python interface
Example
Let's start with a self-contained example.
import random
import youtokentome as yttm
train_data_path = "train_data.txt"
model_path = "example.model"
# Generating random file with training data
# 10000 lines with 100 characters in each line
n_lines = 10000
n_characters = 100
with open(train_data_path, "w") as fout:
for _ in range(n_lines):
print("".join([random.choice("abcd ") for _ in range(n_characters)]), file=fout)
# Generating random text
test_text = "".join([random.choice("abcde ") for _ in range(100)])
# Training model
yttm.BPE.train(data=train_data_path, vocab_size=5000, model=model_path)
# Loading model
bpe = yttm.BPE(model=model_path)
# Two types of tokenization
print(bpe.encode([test_text], output_type=yttm.OutputType.ID))
print(bpe.encode([test_text], output_type=yttm.OutputType.SUBWORD))
Training model
youtokentome.BPE.train(data, model, vocab_size, coverage, n_threads=-1, pad_id=0, unk_id=1, bos_id=2, eos_id=3)
Trains BPE model and saves to file.
Args:
data: string, path to file with training datamodel: string, path to where the trained model will be savedvocab_size: int, number of tokens in the final vocabularycoverage: float, fraction of characters covered by the model. Must be in the range [0, 1]. A good value to use is about 0.9999.n_threads: int, number of parallel threads used to run. If -1 is passed, then all available threads are going to be used. Note that the number of threads is limited by 8 (see benchmark).pad_id: int, reserved id for paddingunk_id: int, reserved id for unknown symbolsbos_id: int, reserved id for begin of sentence tokeneos_id: int, reserved id for end of sentence token
Returns: Class youtokentome.BPE with the loaded model.
Model loading
youtokentome.BPE(model, n_threads=-1)
Class constructor. Loads the trained model.
model: string, path to the trained modeln_threads: int, number of parallel threads used to run. If equal to -1, then the maximum number of threads available will be used.
Methods
Class youtokentome.BPE has the following methods:
encode
encode(self, sentences, output_type=yttm.OutputType.ID, bos=False, eos=False, reverse=False, dropout_prob=0)
Args:
sentences: list of strings, sentences for tokenization.output_type: enum, sentence can be tokenized to ids or subwords. UseOutputType.IDfor ids andOutputType.SUBWORDfor subwords.bos: bool, if True then token “beginning of sentence” will be addedeos: bool, if True then token “end of sentence” will be addedreverse: bool, if True the output sequence of tokens will be reverseddropout_prob: float, BPE-dropout probability (the probability of a merge being dropped). Must be in the range [0, 1].
Returns: If output_type is equal to youtokentome.OutputType.ID or youtokentome.OutputType.SUBWORD then a list of lists of integers or list of lists of strings will be returned respectively.
vocab
vocab(self)
Returns: A list vocab_size strings. The i-th string in the list corresponds to i-th subword.
vocab_size
vocab_size(self)
Returns: int. Size of vocabulary.
subword_to_id
subword_to_id(self, subword)
Args:
subword: string.
Returns: Integer from the range [0, vocab_size-1]. Id of subword or, if there is no such subword in the vocabulary, unk_id will be returned.
id_to_subword
id_to_subword(self, id)
Args:
id: int, must be in the range [0, vocab_size-1]
Returns: string. Subword from vocabulary by id.
decode
decode(self, ids, ignore_ids=None)
Convert each id to subword and concatenate with space symbol.
Args:
ids: list of lists of integers. All integers must be in the range [0, vocab_size-1]ignore_ids: collection of integers. These indices would be ignored during the decoding. All integers must be in the range [0, vocab_size-1] [default: None]
Returns: List of strings.
Command line interface
Example
$ yttm bpe --data TRAINING_DATA_FILE --model OUTPUT_MODEL_FILE --vocab_size 2000
$ yttm encode --model OUTPUT_MODEL_FILE --output_type subword < TEST_DATA_FILE > ENCODED_DATA
Supported commands
YouTokenToMe supports the following commands:
$ yttm --help
Usage: yttm [OPTIONS] COMMAND [ARGS]...
Options:
--help Show this message and exit.
Commands:
bpe Train BPE model.
decode Decode ids to text.
encode Encode text to ids or subwords.
vocab Print list of learned subwords.
Command bpe allows you to train Byte Pair Encoding model based on a text file.
$ yttm bpe --help
Usage: yttm bpe [OPTIONS]
Train BPE model.
Options:
--data PATH Training data file path. [required]
--model PATH Output model file path. [required]
--vocab_size INTEGER Number of tokens in the final vocabulary. [required]
--coverage FLOAT Fraction of characters covered by the model. [default: 1.0]
--n_threads INTEGER Number of threads. [default: -1]
--pad_id INTEGER Padding token id. [default: 0]
--unk_id INTEGER Unknown token id. [default: 1]
--bos_id INTEGER 'Begin of sentence' token id. [default: 2]
--eos_id INTEGER 'End of sentence' token id. [default: 3]
--help Show this message and exit.
Apply BPE encoding for a corpus of sentences. Use stdin for input and stdout for output.
By default, encoding works in parallel using n_threads threads. Number of threads is limited by 8 (see benchmark).
With the --stream option, --n_threads will be ignored and all sentences will be processed one by one. Each sentence will be tokenized and written to the stdout before the next sentence is read.
$ yttm encode --help
Usage: yttm encode [OPTIONS]
Encode text to ids or subwords.
Options:
--model PATH Path to file with learned model. [required]
--output_type TEXT 'id' or 'subword'. [required]
--n_threads INTEGER Number of threads. [default: -1]
--bos Add tab 'begin of sentence'.
--eos Add tab 'end of sentence'.
--reverse Reverse output sequence of tokens.
--stream Process each line before reading the next one.
--dropout_prob BPE-dropout probability (the probability of a merge being dropped). [default: 0]
--help Show this message and exit.
Print vocabulary. This can be useful for understanding the model.
$ yttm vocab --help
Usage: yttm vocab [OPTIONS]
Print list of learned subwords.
Options:
--model PATH Path to file with learned model. [required]
--verbose Add merging rules.
--help Show this message and exit.
Convert ids back to text. Use stdin for input and stdout for output.
$ yttm decode --help
Usage: yttm decode [OPTIONS]
Decode ids to text.
Options:
--model PATH Path to file with learned model. [required]
--ignore_ids List of indices to ignore for decoding. Example: --ignore_ids=1,2,3
--help Show this message and exit.
![[Question] How to learn joint bpe and vocabulary](https://avatars.githubusercontent.com/u/11013023?v=4)
12 Sep 05, 2022
552 Nov 28, 2022
9 Dec 20, 2022
1 Dec 30, 2021
13.8k Jan 02, 2023
3 Nov 17, 2022
1 Nov 15, 2021
1 Jan 18, 2022
457 Dec 23, 2022
6 May 22, 2022
11 Nov 11, 2022
8 Aug 11, 2022
0 Nov 29, 2022
1k Nov 17, 2022
6 Jan 12, 2022
570 Dec 31, 2022
478 Dec 25, 2022
24 Nov 30, 2022
12.3k Jan 02, 2023
4 Jan 20, 2022