tokenizers/pr_2139/en/pipeline.md

# The tokenization pipeline

When calling `Tokenizer.encode` or
`Tokenizer.encode_batch`, the input
text(s) go through the following pipeline:

-   `normalization`
-   `pre-tokenization`
-   `model`
-   `post-processing`

We'll see in details what happens during each of those steps in detail,
as well as when you want to `decode <decoding>` some token ids, and how the 🤗 Tokenizers library allows you
to customize each of those steps to your needs. If you're already
familiar with those steps and want to learn by seeing some code, jump to
`our BERT from scratch example <example>`.

For the examples that require a `Tokenizer` we will use the tokenizer we trained in the
`quicktour`, which you can load with:

```python
from tokenizers import Tokenizer
tokenizer = Tokenizer.from_file("data/tokenizer-wiki.json")
```

```rust
use tokenizers::Tokenizer;
let mut tokenizer = Tokenizer::from_file("data/tokenizer-wiki.json")?;
```

```js
let { Tokenizer } = require("tokenizers");
let tokenizer = Tokenizer.fromFile("data/tokenizer-wiki.json");
```

## Normalization

Normalization is, in a nutshell, a set of operations you apply to a raw
string to make it less random or "cleaner". Common operations include
stripping whitespace, removing accented characters or lowercasing all
text. If you're familiar with [Unicode
normalization](https://unicode.org/reports/tr15), it is also a very
common normalization operation applied in most tokenizers.

Each normalization operation is represented in the 🤗 Tokenizers library
by a `Normalizer`, and you can combine
several of those by using a `normalizers.Sequence`. Here is a normalizer applying NFD Unicode normalization
and removing accents as an example:

```python
from tokenizers import normalizers
from tokenizers.normalizers import NFD, StripAccents
normalizer = normalizers.Sequence([NFD(), StripAccents()])
```

```rust
use tokenizers::normalizers::{
    strip::StripAccents, unicode::NFD, utils::Sequence as NormalizerSequence,
};
let normalizer = NormalizerSequence::new(vec![NFD.into(), StripAccents.into()]);
```

```js
let { sequenceNormalizer, nfdNormalizer, stripAccentsNormalizer } = require("tokenizers");
let normalizer = sequenceNormalizer([nfdNormalizer(), stripAccentsNormalizer()]);
```

You can manually test that normalizer by applying it to any string:

```python
normalizer.normalize_str("Héllò hôw are ü?")
# "Hello how are u?"
```

```rust
use tokenizers::{NormalizedString, Normalizer};
let mut normalized = NormalizedString::from("Héllò hôw are ü?");
normalizer.normalize(&mut normalized)?;
println!("{}", normalized.get());
// "Hello how are u?"
```

```js
let normalized = normalizer.normalizeString("Héllò hôw are ü?")
// "Hello how are u?"
```

When building a `Tokenizer`, you can
customize its normalizer by just changing the corresponding attribute:

```python
tokenizer.normalizer = normalizer
```

```rust
tokenizer.with_normalizer(Some(normalizer)).unwrap();
```

```js
tokenizer.setNormalizer(normalizer)
```

Of course, if you change the way a tokenizer applies normalization, you
should probably retrain it from scratch afterward.

## Pre-Tokenization

Pre-tokenization is the act of splitting a text into smaller objects
that give an upper bound to what your tokens will be at the end of
training. A good way to think of this is that the pre-tokenizer will
split your text into "words" and then, your final tokens will be parts
of those words.

An easy way to pre-tokenize inputs is to split on spaces and
punctuations, which is done by the
`pre_tokenizers.Whitespace`
pre-tokenizer:

```python
from tokenizers.pre_tokenizers import Whitespace
pre_tokenizer = Whitespace()
pre_tokenizer.pre_tokenize_str("Hello! How are you? I'm fine, thank you.")
# [("Hello", (0, 5)), ("!", (5, 6)), ("How", (7, 10)), ("are", (11, 14)), ("you", (15, 18)),
#  ("?", (18, 19)), ("I", (20, 21)), ("'", (21, 22)), ('m', (22, 23)), ("fine", (24, 28)),
#  (",", (28, 29)), ("thank", (30, 35)), ("you", (36, 39)), (".", (39, 40))]
```

```rust
use tokenizers::pre_tokenizers::whitespace::Whitespace;
use tokenizers::{OffsetReferential, OffsetType, PreTokenizedString, PreTokenizer};
let pre_tokenizer = Whitespace {};
let mut pre_tokenized = PreTokenizedString::from("Hello! How are you? I'm fine, thank you.");
pre_tokenizer.pre_tokenize(&mut pre_tokenized)?;
println!(
    "{:?}",
    pre_tokenized.get_splits(OffsetReferential::Original, OffsetType::Byte)
);
// [("Hello", (0, 5), None), ("!", (5, 6), None), ("How", (7, 10), None),
//  ("are", (11, 14), None), ("you", (15, 18), None), ("?", (18, 19), None),
//  ("I", (20, 21), None), ("\'", (21, 22), None), ("m", (22, 23), None),
//  ("fine", (24, 28), None), (",", (28, 29), None), ("thank", (30, 35), None),
//  ("you", (36, 39), None), (".", (39, 40), None)]
```

```js
let { whitespacePreTokenizer } = require("tokenizers");
var preTokenizer = whitespacePreTokenizer();
var preTokenized = preTokenizer.preTokenizeString("Hello! How are you? I'm fine, thank you.");
```

The output is a list of tuples, with each tuple containing one word and
its span in the original sentence (which is used to determine the final
`offsets` of our `Encoding`). Note that splitting on
punctuation will split contractions like `"I'm"` in this example.

You can combine together any `PreTokenizer` together. For instance, here is a pre-tokenizer that will
split on space, punctuation and digits, separating numbers in their
individual digits:

```python
from tokenizers import pre_tokenizers
from tokenizers.pre_tokenizers import Digits
pre_tokenizer = pre_tokenizers.Sequence([Whitespace(), Digits(individual_digits=True)])
pre_tokenizer.pre_tokenize_str("Call 911!")
# [("Call", (0, 4)), ("9", (5, 6)), ("1", (6, 7)), ("1", (7, 8)), ("!", (8, 9))]
```

```rust
use tokenizers::pre_tokenizers::{digits::Digits, sequence::Sequence};
let pre_tokenizer = Sequence::new(vec![Whitespace {}.into(), Digits::new(true).into()]);
let mut pre_tokenized = PreTokenizedString::from("Call 911!");
pre_tokenizer.pre_tokenize(&mut pre_tokenized)?;
println!(
    "{:?}",
    pre_tokenized.get_splits(OffsetReferential::Original, OffsetType::Byte)
);
```

```js
let { sequencePreTokenizer, digitsPreTokenizer } = require("tokenizers");
var preTokenizer = sequencePreTokenizer([whitespacePreTokenizer(), digitsPreTokenizer(true)]);
var preTokenized = preTokenizer.preTokenizeString("Call 911!");
```

As we saw in the `quicktour`, you can
customize the pre-tokenizer of a `Tokenizer` by just changing the corresponding attribute:

```python
tokenizer.pre_tokenizer = pre_tokenizer
```

```rust
tokenizer.with_pre_tokenizer(Some(pre_tokenizer));
```

```js
tokenizer.setPreTokenizer(preTokenizer)
```

Of course, if you change the way the pre-tokenizer, you should probably
retrain your tokenizer from scratch afterward.

## Model

Once the input texts are normalized and pre-tokenized, the
`Tokenizer` applies the model on the
pre-tokens. This is the part of the pipeline that needs training on your
corpus (or that has been trained if you are using a pretrained
tokenizer).

The role of the model is to split your "words" into tokens, using the
rules it has learned. It's also responsible for mapping those tokens to
their corresponding IDs in the vocabulary of the model.

This model is passed along when initializing the
`Tokenizer` so you already know how to
customize this part. Currently, the 🤗 Tokenizers library supports:

-   `models.BPE`
-   `models.Unigram`
-   `models.WordLevel`
-   `models.WordPiece`

For more details about each model and its behavior, you can check
[here](components#models)

## Post-Processing

Post-processing is the last step of the tokenization pipeline, to
perform any additional transformation to the
`Encoding` before it's returned, like
adding potential special tokens.

As we saw in the quick tour, we can customize the post processor of a
`Tokenizer` by setting the
corresponding attribute. For instance, here is how we can post-process
to make the inputs suitable for the BERT model:

```python
from tokenizers.processors import TemplateProcessing
tokenizer.post_processor = TemplateProcessing(
    single="[CLS] $A [SEP]",
    pair="[CLS] $A [SEP] $B:1 [SEP]:1",
    special_tokens=[("[CLS]", 1), ("[SEP]", 2)],
)
```

```rust
use tokenizers::processors::template::TemplateProcessing;
tokenizer.with_post_processor(Some(
    TemplateProcessing::builder()
        .try_single("[CLS] $A [SEP]")
        .unwrap()
        .try_pair("[CLS] $A [SEP] $B:1 [SEP]:1")
        .unwrap()
        .special_tokens(vec![("[CLS]", 1), ("[SEP]", 2)])
        .build()
        .unwrap(),
));
```

```js
let { templateProcessing } = require("tokenizers");
tokenizer.setPostProcessor(templateProcessing(
    "[CLS] $A [SEP]",
    "[CLS] $A [SEP] $B:1 [SEP]:1",
    [["[CLS]", 1], ["[SEP]", 2]]
));
```

Note that contrarily to the pre-tokenizer or the normalizer, you don't
need to retrain a tokenizer after changing its post-processor.

## All together: a BERT tokenizer from scratch

Let's put all those pieces together to build a BERT tokenizer. First,
BERT relies on WordPiece, so we instantiate a new
`Tokenizer` with this model:

```python
from tokenizers import Tokenizer
from tokenizers.models import WordPiece
bert_tokenizer = Tokenizer(WordPiece(unk_token="[UNK]"))
```

```rust
use tokenizers::models::wordpiece::WordPiece;
use tokenizers::Tokenizer;
let mut bert_tokenizer = Tokenizer::new(
    WordPiece::builder()
        .unk_token("[UNK]".to_string())
        .build()
        .unwrap(),
);
```

```js
let { Tokenizer } = require("tokenizers");
let { WordPiece } = require("tokenizers");
let bertTokenizer = new Tokenizer(WordPiece.init({}, { unkToken: "[UNK]" }));
```

Then we know that BERT preprocesses texts by removing accents and
lowercasing. We also use a unicode normalizer:

```python
from tokenizers import normalizers
from tokenizers.normalizers import NFD, Lowercase, StripAccents
bert_tokenizer.normalizer = normalizers.Sequence([NFD(), Lowercase(), StripAccents()])
```

```rust
use tokenizers::normalizers::utils::Sequence as NormalizerSequence;
use tokenizers::normalizers::{strip::StripAccents, unicode::NFD, utils::Lowercase};
bert_tokenizer
    .with_normalizer(Some(NormalizerSequence::new(vec![
        NFD.into(),
        Lowercase.into(),
        StripAccents.into(),
    ])))
    .unwrap();
```

```js
let { sequenceNormalizer, lowercaseNormalizer, nfdNormalizer, stripAccentsNormalizer }
    = require("tokenizers");
bertTokenizer.setNormalizer(sequenceNormalizer([
    nfdNormalizer(), lowercaseNormalizer(), stripAccentsNormalizer()
]))
```

The pre-tokenizer is just splitting on whitespace and punctuation:

```python
from tokenizers.pre_tokenizers import Whitespace
bert_tokenizer.pre_tokenizer = Whitespace()
```

```rust
use tokenizers::pre_tokenizers::whitespace::Whitespace;
bert_tokenizer.with_pre_tokenizer(Some(Whitespace {}));
```

```js
let { whitespacePreTokenizer } = require("tokenizers");
bertTokenizer.setPreTokenizer(whitespacePreTokenizer());
```

And the post-processing uses the template we saw in the previous
section:

```python
from tokenizers.processors import TemplateProcessing
bert_tokenizer.post_processor = TemplateProcessing(
    single="[CLS] $A [SEP]",
    pair="[CLS] $A [SEP] $B:1 [SEP]:1",
    special_tokens=[
        ("[CLS]", 1),
        ("[SEP]", 2),
    ],
)
```

```rust
use tokenizers::processors::template::TemplateProcessing;
bert_tokenizer.with_post_processor(Some(
    TemplateProcessing::builder()
        .try_single("[CLS] $A [SEP]")
        .unwrap()
        .try_pair("[CLS] $A [SEP] $B:1 [SEP]:1")
        .unwrap()
        .special_tokens(vec![("[CLS]", 1), ("[SEP]", 2)])
        .build()
        .unwrap(),
));
```

```js
let { templateProcessing } = require("tokenizers");
bertTokenizer.setPostProcessor(templateProcessing(
    "[CLS] $A [SEP]",
    "[CLS] $A [SEP] $B:1 [SEP]:1",
    [["[CLS]", 1], ["[SEP]", 2]]
));
```

We can use this tokenizer and train on it on wikitext like in the
`quicktour`:

```python
from tokenizers.trainers import WordPieceTrainer
trainer = WordPieceTrainer(vocab_size=30522, special_tokens=["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"])
files = [f"data/wikitext-103-raw/wiki.{split}.raw" for split in ["test", "train", "valid"]]
bert_tokenizer.train(files, trainer)
bert_tokenizer.save("data/bert-wiki.json")
```

```rust
use tokenizers::models::{wordpiece::WordPieceTrainer, TrainerWrapper};
let mut trainer: TrainerWrapper = WordPieceTrainer::builder()
    .vocab_size(30_522)
    .special_tokens(vec![
        AddedToken::from("[UNK]", true),
        AddedToken::from("[CLS]", true),
        AddedToken::from("[SEP]", true),
        AddedToken::from("[PAD]", true),
        AddedToken::from("[MASK]", true),
    ])
    .build()
    .into();
let files = vec![
    "data/wikitext-103-raw/wiki.train.raw".into(),
    "data/wikitext-103-raw/wiki.test.raw".into(),
    "data/wikitext-103-raw/wiki.valid.raw".into(),
];
bert_tokenizer.train_from_files(&mut trainer, files)?;
bert_tokenizer.save("data/bert-wiki.json", false)?;
```

```js
let { wordPieceTrainer } = require("tokenizers");
let trainer = wordPieceTrainer({
    vocabSize: 30522,
    specialTokens: ["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"]
});
let files = ["test", "train", "valid"].map(split => `data/wikitext-103-raw/wiki.${split}.raw`);
bertTokenizer.train(files, trainer);
bertTokenizer.save("data/bert-wiki.json")
```

## Decoding

On top of encoding the input texts, a `Tokenizer` also has an API for decoding, that is converting IDs
generated by your model back to a text. This is done by the methods
`Tokenizer.decode` (for one predicted text) and `Tokenizer.decode_batch` (for a batch of predictions).

The `decoder` will first convert the IDs back to tokens
(using the tokenizer's vocabulary) and remove all special tokens, then
join those tokens with spaces:

```python
output = tokenizer.encode("Hello, y'all! How are you 😁 ?")
print(output.ids)
# [1, 27253, 16, 93, 11, 5097, 5, 7961, 5112, 6218, 0, 35, 2]
tokenizer.decode([1, 27253, 16, 93, 11, 5097, 5, 7961, 5112, 6218, 0, 35, 2])
# "Hello , y ' all ! How are you ?"
```

```rust
let output = tokenizer.encode("Hello, y'all! How are you 😁 ?", true)?;
println!("{:?}", output.get_ids());
// [1, 27253, 16, 93, 11, 5097, 5, 7961, 5112, 6218, 0, 35, 2]
let decoded = tokenizer.decode(
    &[1, 27253, 16, 93, 11, 5097, 5, 7961, 5112, 6218, 0, 35, 2],
    true,
)?;
println!("{decoded}");
// "Hello , y ' all ! How are you ?"
```

```js
let output = await tokenizer.encode("Hello, y'all! How are you 😁 ?");
console.log(output.getIds());
// [1, 27253, 16, 93, 11, 5097, 5, 7961, 5112, 6218, 0, 35, 2]
let decoded = await tokenizer.decode([1, 27253, 16, 93, 11, 5097, 5, 7961, 5112, 6218, 0, 35, 2], true);
// "Hello , y ' all ! How are you ?"
```

If you used a model that added special characters to represent subtokens
of a given "word" (like the `"##"` in
WordPiece) you will need to customize the `decoder` to treat
them properly. If we take our previous `bert_tokenizer` for instance the
default decoding will give:

```python
output = bert_tokenizer.encode("Welcome to the 🤗 Tokenizers library.")
print(output.tokens)
# ["[CLS]", "welcome", "to", "the", "[UNK]", "tok", "##eni", "##zer", "##s", "library", ".", "[SEP]"]
bert_tokenizer.decode(output.ids)
# "welcome to the tok ##eni ##zer ##s library ."
```

```rust
let output = bert_tokenizer.encode("Welcome to the 🤗 Tokenizers library.", true)?;
println!("{:?}", output.get_tokens());
// ["[CLS]", "welcome", "to", "the", "[UNK]", "tok", "##eni", "##zer", "##s", "library", ".", "[SEP]"]
let decoded = bert_tokenizer.decode(output.get_ids(), true)?;
println!("{decoded}");
// "welcome to the tok ##eni ##zer ##s library ."
```

```js
let output = await bertTokenizer.encode("Welcome to the 🤗 Tokenizers library.");
console.log(output.getTokens());
// ["[CLS]", "welcome", "to", "the", "[UNK]", "tok", "##eni", "##zer", "##s", "library", ".", "[SEP]"]
var decoded = await bertTokenizer.decode(output.getIds(), true);
// "welcome to the tok ##eni ##zer ##s library ."
```

But by changing it to a proper decoder, we get:

```python
from tokenizers import decoders
bert_tokenizer.decoder = decoders.WordPiece()
bert_tokenizer.decode(output.ids)
# "welcome to the tokenizers library."
```

```rust
use tokenizers::decoders::wordpiece::WordPiece as WordPieceDecoder;
bert_tokenizer.with_decoder(Some(WordPieceDecoder::default()));
let decoded = bert_tokenizer.decode(output.get_ids(), true)?;
// "welcome to the tokenizers library."
```

```js
let { wordPieceDecoder } = require("tokenizers");
bertTokenizer.setDecoder(wordPieceDecoder());
var decoded = await bertTokenizer.decode(output.getIds(), true);
// "welcome to the tokenizers library."
```