Merge branch 'master' into BPE-tests

bindings/python/Cargo.lock

@@ -448,15 +448,15 @@ dependencies = [
 
 [[package]]
 name = "tokenizers"
-version = "0.0.7"
+version = "0.0.8"
 dependencies = [
  "pyo3 0.8.2 (registry+https://github.com/rust-lang/crates.io-index)",
- "tokenizers-lib 0.0.7",
+ "tokenizers-lib 0.0.8",
 ]
 
 [[package]]
 name = "tokenizers-lib"
-version = "0.0.7"
+version = "0.0.8"
 dependencies = [
  "clap 2.33.0 (registry+https://github.com/rust-lang/crates.io-index)",
  "lazy_static 1.4.0 (registry+https://github.com/rust-lang/crates.io-index)",

bindings/python/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "tokenizers"
-version = "0.0.7"
+version = "0.0.8"
 authors = ["Anthony MOI <m.anthony.moi@gmail.com>"]
 edition = "2018"
 

bindings/python/README.md

@@ -61,7 +61,7 @@ bpe = models.BPE.from_files(vocab, merges)
 tokenizer = Tokenizer(bpe)
 
 # Customize pre-tokenization and decoding
-tokenizer.with_pre_tokenizer(pre_tokenizers.ByteLevel.new())
+tokenizer.with_pre_tokenizer(pre_tokenizers.ByteLevel.new(True))
 tokenizer.with_decoder(decoders.ByteLevel.new())
 
 # And then encode:
@@ -85,7 +85,7 @@ from tokenizers import Tokenizer, models, pre_tokenizers, decoders, trainers
 tokenizer = Tokenizer(models.BPE.empty())
 
 # Customize pre-tokenization and decoding
-tokenizer.with_pre_tokenizer(pre_tokenizers.ByteLevel.new())
+tokenizer.with_pre_tokenizer(pre_tokenizers.ByteLevel.new(True))
 tokenizer.with_decoder(decoders.ByteLevel.new())
 
 # And then train

bindings/python/setup.py

@@ -3,7 +3,7 @@ from setuptools_rust import Binding, RustExtension
 
 setup(
     name="tokenizers",
-    version="0.0.7",
+    version="0.0.8",
     description="Fast and Customizable Tokenizers",
     long_description=open("README.md", "r", encoding="utf-8").read(),
     long_description_content_type="text/markdown",

bindings/python/src/tokenizer.rs

@@ -36,9 +36,8 @@ impl Tokenizer {
         }
     }
 
-    #[getter]
-    fn get_vocab_size(&self) -> usize {
-        self.tokenizer.get_vocab_size()
+    fn get_vocab_size(&self, with_added_tokens: bool) -> usize {
+        self.tokenizer.get_vocab_size(with_added_tokens)
     }
 
     fn with_model(&mut self, model: &mut Model) -> PyResult<()> {
@@ -192,12 +191,16 @@ impl Tokenizer {
         .into()
     }
 
-    fn decode(&self, ids: Vec<u32>) -> PyResult<String> {
-        ToPyResult(self.tokenizer.decode(ids)).into()
+    fn decode(&self, ids: Vec<u32>, skip_special_tokens: bool) -> PyResult<String> {
+        ToPyResult(self.tokenizer.decode(ids, skip_special_tokens)).into()
     }
 
-    fn decode_batch(&self, sentences: Vec<Vec<u32>>) -> PyResult<Vec<String>> {
-        ToPyResult(self.tokenizer.decode_batch(sentences)).into()
+    fn decode_batch(
+        &self,
+        sentences: Vec<Vec<u32>>,
+        skip_special_tokens: bool,
+    ) -> PyResult<Vec<String>> {
+        ToPyResult(self.tokenizer.decode_batch(sentences, skip_special_tokens)).into()
     }
 
     fn token_to_id(&self, token: &str) -> Option<u32> {
@@ -233,6 +236,10 @@ impl Tokenizer {
         Ok(self.tokenizer.add_tokens(&tokens))
     }
 
+    fn add_special_tokens(&mut self, tokens: Vec<&str>) -> PyResult<usize> {
+        Ok(self.tokenizer.add_special_tokens(&tokens))
+    }
+
     fn train(&mut self, trainer: &Trainer, files: Vec<String>) -> PyResult<()> {
         trainer.trainer.execute(|trainer| {
             if let Err(e) = self.tokenizer.train(trainer, files) {

tokenizers/Cargo.lock

@@ -351,7 +351,7 @@ dependencies = [
 
 [[package]]
 name = "tokenizers-lib"
-version = "0.0.7"
+version = "0.0.8"
 dependencies = [
  "clap 2.33.0 (registry+https://github.com/rust-lang/crates.io-index)",
  "lazy_static 1.4.0 (registry+https://github.com/rust-lang/crates.io-index)",

tokenizers/Cargo.toml

@@ -2,7 +2,7 @@
 authors = ["Anthony MOI <m.anthony.moi@gmail.com>"]
 edition = "2018"
 name = "tokenizers-lib"
-version = "0.0.7"
+version = "0.0.8"
 
 [lib]
 name = "tokenizers"

tokenizers/src/cli.rs

@@ -53,7 +53,7 @@ fn shell(matches: &ArgMatches) -> Result<()> {
         println!("Offsets:\t{:?}", encoded.get_offsets());
         println!(
             "Decoded:\t{}",
-            tokenizer.decode(encoded.get_ids().to_vec()).unwrap()
+            tokenizer.decode(encoded.get_ids().to_vec(), true).unwrap()
         );
         println!("Tokenized in {:?}", elapsed);
     }

tokenizers/src/tokenizer/mod.rs

@@ -93,6 +93,14 @@ pub struct AddedToken {
     /// Whether this token must be a single word or can break words
     pub single_word: bool,
 }
+impl AddedToken {
+    fn from(content: String) -> Self {
+        AddedToken {
+            content,
+            ..Default::default()
+        }
+    }
+}
 impl Default for AddedToken {
     fn default() -> Self {
         AddedToken {
@@ -132,6 +140,7 @@ pub struct Tokenizer {
     added_tokens: HashMap<AddedToken, u32>,
     added_tokens_r: HashMap<u32, AddedToken>,
     split_re: Option<regex::Regex>,
+    special_tokens: HashMap<String, u32>,
 
     // General processing parameters
     trunc: Option<TruncationParams>,
@@ -151,6 +160,7 @@ impl Tokenizer {
             added_tokens: HashMap::new(),
             added_tokens_r: HashMap::new(),
             split_re: None,
+            special_tokens: HashMap::new(),
 
             trunc: None,
             padding: None,
@@ -200,18 +210,31 @@ impl Tokenizer {
     }
 
     /// Get the size of the vocabulary
-    pub fn get_vocab_size(&self) -> usize {
+    pub fn get_vocab_size(&self, with_added_tokens: bool) -> usize {
         self.model.get_vocab_size()
+            + if with_added_tokens {
+                self.added_tokens.len()
+            } else {
+                0
+            }
     }
 
     /// Converts a token in the corresponding id.
     pub fn token_to_id(&self, token: &str) -> Option<u32> {
-        self.model.token_to_id(token)
+        if let Some(id) = self.added_tokens.get(&AddedToken::from(token.to_owned())) {
+            Some(*id)
+        } else {
+            self.model.token_to_id(token)
+        }
     }
 
     /// Converts an id to the corresponding token.
     pub fn id_to_token(&self, id: u32) -> Option<String> {
-        self.model.id_to_token(id)
+        if let Some(token) = self.added_tokens_r.get(&id) {
+            Some(token.content.clone())
+        } else {
+            self.model.id_to_token(id)
+        }
     }
 
     /// Encode the given sentence
@@ -324,15 +347,19 @@ impl Tokenizer {
     }
 
     /// Decode the given ids, back to a String
-    pub fn decode(&self, ids: Vec<u32>) -> Result<String> {
+    pub fn decode(&self, ids: Vec<u32>, skip_special_tokens: bool) -> Result<String> {
         let tokens = ids
             .into_iter()
             .map(|id| {
-                if let Some(token) = self.added_tokens_r.get(&id) {
+                let token = if let Some(token) = self.added_tokens_r.get(&id) {
                     Some(token.content.to_owned())
                 } else {
                     self.model.id_to_token(id)
-                }
+                };
+
+                token.filter(|token| {
+                    !skip_special_tokens || !self.special_tokens.contains_key(token)
+                })
             })
             .filter(|token| token.is_some())
             .map(|id| id.unwrap())
@@ -346,10 +373,14 @@ impl Tokenizer {
     }
 
     /// Decode all sentences in parallel
-    pub fn decode_batch(&self, sentences: Vec<Vec<u32>>) -> Result<Vec<String>> {
+    pub fn decode_batch(
+        &self,
+        sentences: Vec<Vec<u32>>,
+        skip_special_tokens: bool,
+    ) -> Result<Vec<String>> {
         sentences
             .into_par_iter()
-            .map(|sentence| self.decode(sentence))
+            .map(|sentence| self.decode(sentence, skip_special_tokens))
             .collect()
     }
 
@@ -467,11 +498,29 @@ impl Tokenizer {
         Ok(final_encoding)
     }
 
+    /// Register the given tokens as special tokens. This is especially useful for removing
+    /// these special tokens while decoding
+    pub fn add_special_tokens(&mut self, tokens: &[&str]) -> usize {
+        let added_tokens = tokens
+            .iter()
+            .map(|t| AddedToken::from((*t).to_owned()))
+            .collect::<Vec<_>>();
+
+        let added = self.add_tokens(&added_tokens);
+        for token in tokens {
+            if let Some(id) = self.token_to_id(token) {
+                self.special_tokens.entry((*token).to_owned()).or_insert(id);
+            }
+        }
+
+        added
+    }
+
     /// Add the given tokens to the added vocabulary
     pub fn add_tokens(&mut self, tokens: &[AddedToken]) -> usize {
         let mut ignored = 0;
         for token in tokens {
-            if token.content.is_empty() {
+            if token.content.is_empty() || self.token_to_id(&token.content).is_some() {
                 ignored += 1;
                 continue;
             }