Merge pull request #197 from huggingface/remove-normalized

Remove NormalizedString from Encoding

.github/workflows/rust.yml

@@ -60,7 +60,22 @@ jobs:
           args: --manifest-path ./tokenizers/Cargo.toml --all-targets --all-features -- -D warnings
 
       - name: Run Tests
+        if: matrix.os != 'windows-latest'
+        shell: bash
+        working-directory: ./tokenizers
+        run: make test
+
+      # Skip integration tests for now on Windows
+      - name: Run lib Tests on Windows
+        if: matrix.os == 'windows-latest'
         uses: actions-rs/cargo@v1
         with:
           command: test
-          args: --verbose --manifest-path ./tokenizers/Cargo.toml
+          args: --verbose --manifest-path ./tokenizers/Cargo.toml --lib
+
+      - name: Run doc Tests on Windows
+        if: matrix.os == 'windows-latest'
+        uses: actions-rs/cargo@v1
+        with:
+          command: test
+          args: --verbose --manifest-path ./tokenizers/Cargo.toml --doc

.gitignore

@@ -7,6 +7,7 @@ target
 Cargo.lock
 
 /data
+tokenizers/data
 /docs
 
 __pycache__

bindings/node/lib/bindings/raw-encoding.d.ts

@@ -44,17 +44,6 @@ export interface RawEncoding {
    */
   getTypeIds(): number[];
 
-  /**
-   * Returns the original string
-   *
-   * @param [begin] The index from which to start (can be negative).
-   * @param [end] The index (excluded) to which to stop (can be negative).
-   * Stopping at the end of the string if not provided.
-   * @returns The full original string if no parameter is provided,
-   * otherwise the original string between `begin` and `end`
-   */
-  getOriginalString(begin?: number, end?: number): string;
-
   /**
    * Pad the current Encoding at the given length
    *

bindings/node/lib/bindings/raw-encoding.test.ts

@@ -22,7 +22,6 @@ describe("RawEncoding", () => {
     expect(typeof encoding.getIds).toBe("function");
     expect(typeof encoding.getLength).toBe("function");
     expect(typeof encoding.getOffsets).toBe("function");
-    expect(typeof encoding.getOriginalString).toBe("function");
     expect(typeof encoding.getOverflowing).toBe("function");
     expect(typeof encoding.getSpecialTokensMask).toBe("function");
     expect(typeof encoding.getTokens).toBe("function");
@@ -31,109 +30,6 @@ describe("RawEncoding", () => {
     expect(typeof encoding.truncate).toBe("function");
   });
 
-  describe("getOriginalString", () => {
-    it("returns the full original string when no params", () => {
-      const original = encoding.getOriginalString();
-      expect(original).toEqual(originalString);
-    });
-
-    it("accepts `undefined` as first parameter", () => {
-      const original = encoding.getOriginalString(undefined);
-      expect(original).toEqual(originalString);
-    });
-
-    it("accepts `undefined` as second parameter", () => {
-      const original = encoding.getOriginalString(0, undefined);
-      expect(original).toEqual(originalString);
-    });
-
-    it("throws an error when `begin` is out of range", () => {
-      expect(() => encoding.getOriginalString(1000)).toThrow();
-    });
-
-    it("returns the original string starting at the specified index", () => {
-      const original = encoding.getOriginalString(3);
-      expect(original).toEqual("name is john");
-    });
-
-    it("throws an error when `end` is out of range", () => {
-      expect(() => encoding.getOriginalString(0, 1000)).toThrow();
-    });
-
-    it("returns the original string between the two specified indexes", () => {
-      const original = encoding.getOriginalString(3, 7);
-      expect(original).toEqual("name");
-    });
-
-    describe("with only a negative `begin`", () => {
-      it("returns the original string counting from the end when in the range", () => {
-        const original = encoding.getOriginalString(-4);
-        expect(original).toEqual("john");
-      });
-
-      it("throws an error when out of range", () => {
-        expect(() => encoding.getOriginalString(-1000)).toThrow();
-      });
-    });
-
-    describe("with a positive `begin` and a negative `end`", () => {
-      it("returns the original string when resulting range is valid", () => {
-        const original = encoding.getOriginalString(3, -5);
-        expect(original).toEqual("name is");
-      });
-
-      it("throws an error when resulting `end` index is lower than `begin`", () => {
-        expect(() => encoding.getOriginalString(7, -10)).toThrow();
-      });
-
-      it("throws an error when `begin` is out of range", () => {
-        expect(() => encoding.getOriginalString(1000, -10)).toThrow();
-      });
-
-      it("throws an error when resulting `end` index is out of range", () => {
-        expect(() => encoding.getOriginalString(7, -1000)).toThrow();
-      });
-    });
-
-    describe("with a negative `begin` and a positive `end`", () => {
-      it("returns the original string when resulting range is valid", () => {
-        const original = encoding.getOriginalString(-7, 10);
-        expect(original).toEqual("is");
-      });
-
-      it("throws an error when resulting `begin` index is upper than `end`", () => {
-        expect(() => encoding.getOriginalString(-3, 5)).toThrow();
-      });
-
-      it("throws an error when `end` is out of range", () => {
-        expect(() => encoding.getOriginalString(-5, 1000)).toThrow();
-      });
-
-      it("throws an error when resulting `begin` index is out of range", () => {
-        expect(() => encoding.getOriginalString(-1000, 10)).toThrow();
-      });
-    });
-
-    describe("with negatives `begin` and `end`", () => {
-      it("returns the original string when resulting range is valid", () => {
-        const original = encoding.getOriginalString(-7, -5);
-        expect(original).toEqual("is");
-      });
-
-      it("throws an error when resulting `end` index is lower than `begin`", () => {
-        expect(() => encoding.getOriginalString(-5, -10)).toThrow();
-      });
-
-      it("throws an error when resulting `begin` index is out of range", () => {
-        expect(() => encoding.getOriginalString(-1000, -10)).toThrow();
-      });
-
-      it("throws an error when resulting `end` index is out of range", () => {
-        expect(() => encoding.getOriginalString(-10, -1000)).toThrow();
-      });
-    });
-  });
-
   describe("truncate", () => {
     it("accepts `undefined` as second parameter", () => {
       expect(encoding.truncate(10, undefined)).toBeUndefined();

bindings/node/lib/bindings/tokenizer.test.ts

@@ -1,3 +1,4 @@
+/* eslint-disable @typescript-eslint/no-explicit-any */
 /* eslint-disable @typescript-eslint/no-empty-function */
 
 import { promisify } from "util";
@@ -112,7 +113,7 @@ describe("Tokenizer", () => {
         [2, 6],
         [6, 8],
         [8, 12],
-        [12, 16]
+        [0, 4]
       ]);
       expect(encoding.getOverflowing()).toEqual([]);
       expect(encoding.getSpecialTokensMask()).toEqual([0, 0, 0, 0, 0]);

bindings/node/lib/bindings/utils.d.ts

@@ -0,0 +1,11 @@
+/**
+ * Returns a subpart of a string according to specified indexes, and respecting unicode characters
+ *
+ * @param text The text for which to return a subpart
+ * @param [begin] The index from which to start (can be negative).
+ * @param [end] The index (excluded) to which to stop (can be negative).
+ * Stopping at the end of the string if not provided.
+ * @returns The full string if no start/end indexes are provided,
+ * otherwise the original string between `begin` (included) and `end` (excluded)
+ */
+export function slice(text: string, start?: number, end?: number): string;

bindings/node/lib/bindings/utils.js

@@ -0,0 +1,5 @@
+const native = require("./native");
+
+module.exports = {
+  slice: native.utils_slice
+};

bindings/node/lib/bindings/utils.test.ts

@@ -0,0 +1,107 @@
+import { slice } from "./utils";
+
+describe("slice", () => {
+  const text = "My name is John 👋";
+  const sliceText = slice.bind({}, text);
+
+  it("returns the full text when no params", () => {
+    const sliced = sliceText();
+    expect(sliced).toEqual(text);
+  });
+
+  it("accepts `undefined` as second parameter", () => {
+    const original = sliceText(undefined);
+    expect(original).toEqual(text);
+  });
+
+  it("accepts `undefined` as third parameter", () => {
+    const original = sliceText(0, undefined);
+    expect(original).toEqual(text);
+  });
+
+  it("throws an error when `begin` is out of range", () => {
+    expect(() => sliceText(1000)).toThrow();
+  });
+
+  it("returns slice starting at the specified index", () => {
+    const original = sliceText(3);
+    expect(original).toEqual("name is John 👋");
+  });
+
+  it("throws an error when `end` is out of range", () => {
+    expect(() => sliceText(0, 1000)).toThrow();
+  });
+
+  it("returns the text between the two specified indexes", () => {
+    const original = sliceText(3, 7);
+    expect(original).toEqual("name");
+  });
+
+  describe("with only a negative `begin`", () => {
+    it("returns the original string counting from the end when in the range", () => {
+      const original = sliceText(-1);
+      expect(original).toEqual("👋");
+    });
+
+    it("throws an error when out of range", () => {
+      expect(() => sliceText(-1000)).toThrow();
+    });
+  });
+
+  describe("with a positive `begin` and a negative `end`", () => {
+    it("returns correct slice when resulting range is valid", () => {
+      const original = sliceText(3, -7);
+      expect(original).toEqual("name is");
+    });
+
+    it("throws an error when resulting `end` index is lower than `begin`", () => {
+      expect(() => sliceText(7, -12)).toThrow();
+    });
+
+    it("throws an error when `begin` is out of range", () => {
+      expect(() => sliceText(1000, -12)).toThrow();
+    });
+
+    it("throws an error when resulting `end` index is out of range", () => {
+      expect(() => sliceText(7, -1000)).toThrow();
+    });
+  });
+
+  describe("with a negative `begin` and a positive `end`", () => {
+    it("returns correct slice when resulting range is valid", () => {
+      const original = sliceText(-9, 10);
+      expect(original).toEqual("is");
+    });
+
+    it("throws an error when resulting `begin` index is upper than `end`", () => {
+      expect(() => sliceText(-3, 5)).toThrow();
+    });
+
+    it("throws an error when `end` is out of range", () => {
+      expect(() => sliceText(-5, 1000)).toThrow();
+    });
+
+    it("throws an error when resulting `begin` index is out of range", () => {
+      expect(() => sliceText(-1000, 10)).toThrow();
+    });
+  });
+
+  describe("with negatives `begin` and `end`", () => {
+    it("returns correct slice when resulting range is valid", () => {
+      const original = sliceText(-9, -7);
+      expect(original).toEqual("is");
+    });
+
+    it("throws an error when resulting `end` index is lower than `begin`", () => {
+      expect(() => sliceText(-5, -10)).toThrow();
+    });
+
+    it("throws an error when resulting `begin` index is out of range", () => {
+      expect(() => sliceText(-1000, -10)).toThrow();
+    });
+
+    it("throws an error when resulting `end` index is out of range", () => {
+      expect(() => sliceText(-10, -1000)).toThrow();
+    });
+  });
+});

bindings/node/lib/implementations/encoding.ts

@@ -5,7 +5,6 @@ export class Encoding {
   private _ids?: number[];
   private _length?: number;
   private _offsets?: [number, number][];
-  private _originalString?: string;
   private _overflowing?: Encoding[];
   private _specialTokensMask?: number[];
   private _tokens?: string[];
@@ -103,27 +102,6 @@ export class Encoding {
     return (this._typeIds = this.rawEncoding.getTypeIds());
   }
 
-  /**
-   * Returns the original string
-   *
-   * @param [begin] The index from which to start (can be negative).
-   * @param [end] The index (excluded) to which to stop (can be negative).
-   * Stopping at the end of the string if not provided.
-   * @returns The full original string if no parameter is provided,
-   * otherwise the original string between `begin` and `end`
-   */
-  getOriginalString(begin?: number, end?: number): string {
-    if (begin === undefined && end === undefined) {
-      if (this._originalString !== undefined) {
-        return this._originalString;
-      } else {
-        return (this._originalString = this.rawEncoding.getOriginalString());
-      }
-    }
-
-    return this.rawEncoding.getOriginalString(begin, end);
-  }
-
   /**
    * Pad the current Encoding at the given length
    *
@@ -153,7 +131,6 @@ export class Encoding {
       "_ids",
       "_length",
       "_offsets",
-      "_originalString",
       "_overflowing",
       "_specialTokensMask",
       "_tokens",

bindings/node/lib/index.ts

@@ -1,5 +1,6 @@
 // export * from "./bindings";
 export * from "./implementations/tokenizers";
 export * from "./bindings/enums";
+export { slice } from "./bindings/utils";
 export { PaddingOptions, TruncationOptions } from "./bindings/tokenizer";
 export { Encoding } from "./implementations/encoding";

bindings/node/native/src/container.rs

@@ -0,0 +1,102 @@
+/// A Container type
+///
+/// Provides an interface to allow transfer of ownership between Node and Rust.
+/// It either contains a Box with full ownership of the content, or a pointer to the content.
+///
+/// The main goal here is to allow Node calling into Rust to initialize some objects. Later
+/// these objects may need to be used by Rust who will expect to take ownership. Since Node
+/// does not allow any sort of ownership transfer, it will keep a reference to this object
+/// until it gets cleaned up by the GC. In this case, we actually give the ownership to Rust,
+/// and just keep a pointer in the Node object.
+pub enum Container<T: ?Sized> {
+    Owned(Box<T>),
+    Pointer(*mut T),
+    Empty,
+}
+
+impl<T> Container<T>
+where
+    T: ?Sized,
+{
+    pub fn from_ref(reference: &Box<T>) -> Self {
+        let content: *const T = &**reference;
+        Container::Pointer(content as *mut _)
+    }
+
+    pub fn is_owned(&self) -> bool {
+        if let Container::Owned(_) = &self {
+            true
+        } else {
+            false
+        }
+    }
+
+    /// Consumes ourself and return the Boxed element if we have the ownership, None otherwise.
+    pub fn take(self) -> Option<Box<T>> {
+        match self {
+            Container::Owned(obj) => Some(obj),
+            Container::Pointer(_) => None,
+            Container::Empty => None,
+        }
+    }
+
+    /// Replace an empty content by the new provided owned one, otherwise do nothing
+    pub fn to_owned(&mut self, o: Box<T>) {
+        if let Container::Empty = self {
+            unsafe {
+                let new_container = Container::Owned(o);
+                std::ptr::write(self, new_container);
+            }
+        }
+    }
+
+    /// Return the owned T, keeping a Pointer to it if we currently own it. None otherwise
+    pub fn to_pointer(&mut self) -> Option<Box<T>> {
+        if let Container::Owned(_) = self {
+            unsafe {
+                let old_container = std::ptr::read(self);
+                let ptr = Box::into_raw(old_container.take().unwrap());
+                let new_container = Container::Pointer(ptr);
+                std::ptr::write(self, new_container);
+
+                Some(Box::from_raw(ptr))
+            }
+        } else {
+            None
+        }
+    }
+
+    pub fn execute<F, U>(&self, closure: F) -> U
+    where
+        F: FnOnce(Option<&Box<T>>) -> U,
+    {
+        match self {
+            Container::Owned(val) => closure(Some(val)),
+            Container::Pointer(ptr) => unsafe {
+                let val = Box::from_raw(*ptr);
+                let res = closure(Some(&val));
+                // We call this to make sure we don't drop the Box
+                Box::into_raw(val);
+                res
+            },
+            Container::Empty => closure(None),
+        }
+    }
+
+    pub fn execute_mut<F, U>(&mut self, closure: F) -> U
+    where
+        F: FnOnce(Option<&mut Box<T>>) -> U,
+    {
+        match self {
+            Container::Owned(val) => closure(Some(val)),
+            Container::Pointer(ptr) => unsafe {
+                let mut val = Box::from_raw(*ptr);
+                let res = closure(Some(&mut val));
+                // We call this to make sure we don't drop the Box
+                Box::into_raw(val);
+                res
+            },
+            Container::Empty => closure(None),
+        }
+    }
+}

bindings/node/native/src/decoders.rs

@@ -1,6 +1,6 @@
 extern crate tokenizers as tk;
 
-use crate::utils::Container;
+use crate::container::Container;
 use neon::prelude::*;
 
 /// Decoder

bindings/node/native/src/encoding.rs

@@ -2,7 +2,7 @@ extern crate tokenizers as tk;
 
 use tk::tokenizer::PaddingDirection;
 
-use crate::utils::Container;
+use crate::container::Container;
 use neon::prelude::*;
 
 /// Encoding
@@ -159,62 +159,6 @@ declare_types! {
             Ok(js_overflowings.upcast())
         }
 
-        method getOriginalString(mut cx) {
-            // getOriginalString(begin?: number, end?: number)
-            let this = cx.this();
-
-            let len_original = {
-                let guard = cx.lock();
-                let len = this.borrow(&guard).encoding.execute(|encoding| {
-                    encoding.unwrap().get_normalized().len_original()
-                });
-                len
-            };
-
-            let get_index = |x: i32| -> usize {
-                if x >= 0 {
-                    x as usize
-                } else {
-                    (len_original as i32 + x) as usize
-                }
-            };
-
-            let begin_index = if let Some(begin_arg) = cx.argument_opt(0) {
-                if begin_arg.downcast::<JsUndefined>().is_err() {
-                    let begin = begin_arg.downcast::<JsNumber>().or_throw(&mut cx)?.value() as i32;
-                    get_index(begin)
-                } else {
-                    0
-                }
-            } else {
-                0
-            };
-
-            let end_index = if let Some(end_arg) = cx.argument_opt(1) {
-                if end_arg.downcast::<JsUndefined>().is_err() {
-                    let end = end_arg.downcast::<JsNumber>().or_throw(&mut cx)?.value() as i32;
-                    get_index(end)
-                } else {
-                    len_original
-                }
-            } else {
-                len_original
-            };
-
-            let original = {
-                let guard = cx.lock();
-                let original = this.borrow(&guard).encoding.execute(|encoding| {
-                    encoding.unwrap().get_normalized().get_range_original(begin_index..end_index)
-                });
-                original
-            };
-            if let Some(original) = original {
-                Ok(cx.string(original).upcast())
-            } else {
-                cx.throw_error("Error in offsets")
-            }
-        }
-
         method pad(mut cx) {
             // pad(length: number, options?: {
             //   direction?: 'left' | 'right' = 'right',

bindings/node/native/src/lib.rs

@@ -3,6 +3,7 @@
 extern crate neon;
 extern crate tokenizers as tk;
 
+mod container;
 mod decoders;
 mod encoding;
 mod models;
@@ -31,6 +32,8 @@ register_module!(mut m, {
     pre_tokenizers::register(&mut m, "pre_tokenizers")?;
     // Trainers
     trainers::register(&mut m, "trainers")?;
+    // Utils
+    utils::register(&mut m, "utils")?;
 
     Ok(())
 });

bindings/node/native/src/models.rs

@@ -1,7 +1,7 @@
 extern crate tokenizers as tk;
 
+use crate::container::Container;
 use crate::tasks::models::{BPEFromFilesTask, WordPieceFromFilesTask};
-use crate::utils::Container;
 use neon::prelude::*;
 use std::path::Path;
 

bindings/node/native/src/normalizers.rs

@@ -1,6 +1,6 @@
 extern crate tokenizers as tk;
 
-use crate::utils::Container;
+use crate::container::Container;
 use neon::prelude::*;
 
 /// Normalizer

bindings/node/native/src/pre_tokenizers.rs

@@ -1,6 +1,6 @@
 extern crate tokenizers as tk;
 
-use crate::utils::Container;
+use crate::container::Container;
 use neon::prelude::*;
 
 /// PreTokenizers

bindings/node/native/src/processors.rs

@@ -1,6 +1,6 @@
 extern crate tokenizers as tk;
 
-use crate::utils::Container;
+use crate::container::Container;
 use neon::prelude::*;
 
 /// Processor

bindings/node/native/src/tokenizer.rs

@@ -1,5 +1,6 @@
 extern crate tokenizers as tk;
 
+use crate::container::Container;
 use crate::decoders::JsDecoder;
 use crate::models::JsModel;
 use crate::normalizers::JsNormalizer;
@@ -7,7 +8,6 @@ use crate::pre_tokenizers::JsPreTokenizer;
 use crate::processors::JsPostProcessor;
 use crate::tasks::tokenizer::{DecodeTask, EncodeTask, WorkingTokenizer};
 use crate::trainers::JsTrainer;
-use crate::utils::Container;
 use neon::prelude::*;
 
 use tk::tokenizer::{
@@ -64,17 +64,43 @@ declare_types! {
             let mut with_added_tokens = true;
             if let Some(args) = cx.argument_opt(0) {
                 if args.downcast::<JsUndefined>().is_err() {
-                    with_added_tokens = args.downcast::<JsBoolean>().or_throw(&mut cx)?.value() as bool;
+                    with_added_tokens = args.downcast::<JsBoolean>()
+                        .or_throw(&mut cx)?
+                        .value() as bool;
                 }
             }
 
             let mut this = cx.this();
             let guard = cx.lock();
-            let size = this.borrow_mut(&guard).tokenizer.get_vocab_size(with_added_tokens);
+            let size = this.borrow_mut(&guard)
+                .tokenizer
+                .get_vocab_size(with_added_tokens);
 
             Ok(cx.number(size as f64).upcast())
         }
 
+        method normalize(mut cx) {
+            // normalize(sentence: String) -> String
+            let sentence = cx.argument::<JsString>(0)?.value();
+
+            let this = cx.this();
+            let guard = cx.lock();
+
+            let result = {
+                this.borrow(&guard)
+                    .tokenizer
+                    .normalize(&sentence)
+                    .map(|s| s.get().to_owned())
+            };
+            let normalized = result
+                .map_err(|e| {
+                    cx.throw_error::<_, ()>(format!("{}", e))
+                        .unwrap_err()
+                })?;
+
+            Ok(cx.string(normalized).upcast())
+        }
+
         method encode(mut cx) {
             // encode(
             //   sentence: String,

bindings/node/native/src/trainers.rs

@@ -1,6 +1,6 @@
 extern crate tokenizers as tk;
 
-use crate::utils::Container;
+use crate::container::Container;
 use neon::prelude::*;
 use std::collections::HashSet;
 

bindings/node/native/src/utils.rs

@@ -1,102 +1,51 @@
-/// A Container type
+extern crate tokenizers as tk;
-///
-/// Provides an interface to allow transfer of ownership between Node and Rust.
-/// It either contains a Box with full ownership of the content, or a pointer to the content.
-///
-/// The main goal here is to allow Node calling into Rust to initialize some objects. Later
-/// these objects may need to be used by Rust who will expect to take ownership. Since Node
-/// does not allow any sort of ownership transfer, it will keep a reference to this object
-/// until it gets cleaned up by the GC. In this case, we actually give the ownership to Rust,
-/// and just keep a pointer in the Node object.
-pub enum Container<T: ?Sized> {
-    Owned(Box<T>),
-    Pointer(*mut T),
-    Empty,
-}
 
-impl<T> Container<T>
+use neon::prelude::*;
-where
-    T: ?Sized,
-{
-    pub fn from_ref(reference: &Box<T>) -> Self {
-        let content: *const T = &**reference;
-        Container::Pointer(content as *mut _)
-    }
 
-    pub fn is_owned(&self) -> bool {
+/// slice(s: string, start?: number, end?: number)
-        if let Container::Owned(_) = &self {
+fn slice(mut cx: FunctionContext) -> JsResult<JsString> {
-            true
+    let s = cx.argument::<JsString>(0)?.value();
+    let len = s.chars().count();
+
+    let get_index = |x: i32| -> usize {
+        if x >= 0 {
+            x as usize
         } else {
-            false
+            (len as i32 + x) as usize
-        }
         }
+    };
 
-    /// Consumes ourself and return the Boxed element if we have the ownership, None otherwise.
+    let begin_index = if let Some(begin_arg) = cx.argument_opt(1) {
-    pub fn take(self) -> Option<Box<T>> {
+        if begin_arg.downcast::<JsUndefined>().is_err() {
-        match self {
+            let begin = begin_arg.downcast::<JsNumber>().or_throw(&mut cx)?.value() as i32;
-            Container::Owned(obj) => Some(obj),
+            get_index(begin)
-            Container::Pointer(_) => None,
+        } else {
-            Container::Empty => None,
+            0
-        }
-    }
-
-    /// Replace an empty content by the new provided owned one, otherwise do nothing
-    pub fn to_owned(&mut self, o: Box<T>) {
-        if let Container::Empty = self {
-            unsafe {
-                let new_container = Container::Owned(o);
-                std::ptr::write(self, new_container);
-            }
-        }
-    }
-
-    /// Return the owned T, keeping a Pointer to it if we currently own it. None otherwise
-    pub fn to_pointer(&mut self) -> Option<Box<T>> {
-        if let Container::Owned(_) = self {
-            unsafe {
-                let old_container = std::ptr::read(self);
-                let ptr = Box::into_raw(old_container.take().unwrap());
-                let new_container = Container::Pointer(ptr);
-                std::ptr::write(self, new_container);
-
-                Some(Box::from_raw(ptr))
         }
     } else {
-            None
+        0
+    };
+
+    let end_index = if let Some(end_arg) = cx.argument_opt(2) {
+        if end_arg.downcast::<JsUndefined>().is_err() {
+            let end = end_arg.downcast::<JsNumber>().or_throw(&mut cx)?.value() as i32;
+            get_index(end)
+        } else {
+            len
+        }
+    } else {
+        len
+    };
+
+    if let Some(slice) = tk::tokenizer::get_range_of(&s, begin_index..end_index) {
+        Ok(cx.string(slice))
+    } else {
+        cx.throw_error("Error in offsets")
     }
 }
 
-    pub fn execute<F, U>(&self, closure: F) -> U
+/// Register everything here
-    where
+pub fn register(m: &mut ModuleContext, prefix: &str) -> NeonResult<()> {
-        F: FnOnce(Option<&Box<T>>) -> U,
+    m.export_function(&format!("{}_slice", prefix), slice)?;
-    {
+    Ok(())
-        match self {
-            Container::Owned(val) => closure(Some(val)),
-            Container::Pointer(ptr) => unsafe {
-                let val = Box::from_raw(*ptr);
-                let res = closure(Some(&val));
-                // We call this to make sure we don't drop the Box
-                Box::into_raw(val);
-                res
-            },
-            Container::Empty => closure(None),
-        }
-    }
-
-    pub fn execute_mut<F, U>(&mut self, closure: F) -> U
-    where
-        F: FnOnce(Option<&mut Box<T>>) -> U,
-    {
-        match self {
-            Container::Owned(val) => closure(Some(val)),
-            Container::Pointer(ptr) => unsafe {
-                let mut val = Box::from_raw(*ptr);
-                let res = closure(Some(&mut val));
-                // We call this to make sure we don't drop the Box
-                Box::into_raw(val);
-                res
-            },
-            Container::Empty => closure(None),
-        }
-    }
 }

bindings/python/CHANGELOG.md

@@ -6,18 +6,28 @@ a high number of files as it avoids having too many progress bars on screen.
 - `ByteLevel` is also a `PostProcessor` now and handles trimming the offsets if activated. This
 avoids the unintuitive inclusion of the whitespaces in the produced offsets, even if these
 whitespaces are part of the actual token.
-It has been added to `ByteLevelBPETokenizer` and but it is off by default (`trim_offsets=False`).
+It has been added to `ByteLevelBPETokenizer` but it is off by default (`trim_offsets=False`).
-- `encode` and `encode_batch` no take a new optional argument, specifying whether we should add the
+([#188](https://github.com/huggingface/tokenizers/pull/188))
-special tokens. This stays activated by default.
+- `encode` and `encode_batch` now take a new optional argument, specifying whether we should add the
+special tokens. This is activated by default. ([#193](https://github.com/huggingface/tokenizers/pull/193))
+- `original_str` and `normalized_str` have been removed from the `Encoding` returned by `encode` and
+`encode_batch`. This brings a reduction of 70% the memory footprint.
+([#197](https://github.com/huggingface/tokenizers/pull/197))
 
 ## Fixes:
-- Fix some issues with the offsets being wrong with the `ByteLevel` BPE:
+- Fix some issues with the offsets being wrong with the `ByteLevel` BPE ([#193](https://github.com/huggingface/tokenizers/pull/193)):
 	- when `add_prefix_space=True`
 	- when a Unicode character gets split-up in multiple byte-level characters ([#156](https://github.com/huggingface/tokenizers/issues/156))
 
 ## How to migrate:
 - Add the `ByteLevel` `PostProcessor` to your byte-level BPE tokenizers if relevant. If you are
 using `ByteLevelBPETokenizer`, this option is disabled by default (`trim_offsets=False`).
+- Access to the `original_str` on the `Encoding` has been removed. The original string is the input
+of `encode` so it didn't make sense to keep it here.
+- No need to call `original_str.offsets(offsets[N])` to convert offsets to the original string. They
+are now relative to the original string by default.
+- Access to the `normalized_str` on the `Encoding` has been removed. Can be retrieved by calling
+`normalize(sequence)` on the `Tokenizer`
 
 # v0.6.0
 

bindings/python/src/encoding.rs

@@ -1,115 +1,11 @@
 extern crate tokenizers as tk;
 
 use crate::error::PyError;
-use pyo3::exceptions;
 use pyo3::prelude::*;
 use pyo3::types::*;
-use pyo3::{PyMappingProtocol, PyObjectProtocol, PySequenceProtocol};
+use pyo3::{PyObjectProtocol, PySequenceProtocol};
 use tk::tokenizer::PaddingDirection;
 
-fn get_range(item: PyObject, max_len: usize) -> PyResult<std::ops::Range<usize>> {
-    let gil = Python::acquire_gil();
-    let py = gil.python();
-
-    let slice = if let Ok(index) = item.extract::<isize>(py) {
-        if index >= max_len as isize || index < -(max_len as isize) {
-            Err(exceptions::IndexError::py_err("Index out of bounds"))
-        } else {
-            Ok(if index == -1 {
-                PySlice::new(py, index, max_len as isize, 1)
-            } else {
-                PySlice::new(py, index, index + 1, 1)
-            })
-        }
-    } else if let Ok(slice) = item.cast_as::<PySlice>(py) {
-        Ok(slice)
-    } else if let Ok(offset) = item.cast_as::<PyTuple>(py) {
-        if offset.len() == 2 {
-            let start = offset.get_item(0).extract::<isize>()?;
-            let end = offset.get_item(1).extract::<isize>()?;
-            Ok(PySlice::new(py, start, end, 1))
-        } else {
-            Err(exceptions::TypeError::py_err("Expected Tuple[int, int]"))
-        }
-    } else {
-        Err(exceptions::TypeError::py_err(
-            "Expected number or slice or Tuple[int, int]",
-        ))
-    }?;
-
-    // Find out range from the slice
-    let len: std::os::raw::c_long = (max_len as i32) as _;
-    let PySliceIndices { start, stop, .. } = slice.indices(len)?;
-
-    Ok(start as usize..stop as usize)
-}
-
-enum IndexableStringType {
-    Original,
-    Normalized,
-}
-
-#[pyclass(dict)]
-pub struct IndexableString {
-    s: tk::tokenizer::NormalizedString,
-    t: IndexableStringType,
-}
-#[pymethods]
-impl IndexableString {
-    fn offsets(&self, item: PyObject) -> PyResult<Option<(usize, usize)>> {
-        let range = get_range(item, self.s.len())?;
-
-        match self.t {
-            IndexableStringType::Original => Ok(self
-                .s
-                .get_original_offsets(range)
-                .map(|range| (range.start, range.end))),
-            IndexableStringType::Normalized => Ok(Some((range.start, range.end))),
-        }
-    }
-}
-
-#[pyproto]
-impl PyObjectProtocol for IndexableString {
-    fn __repr__(&self) -> PyResult<String> {
-        Ok(match self.t {
-            IndexableStringType::Original => self.s.get_original().to_owned(),
-            IndexableStringType::Normalized => self.s.get().to_owned(),
-        })
-    }
-
-    fn __str__(&self) -> PyResult<String> {
-        Ok(match self.t {
-            IndexableStringType::Original => self.s.get_original().to_owned(),
-            IndexableStringType::Normalized => self.s.get().to_owned(),
-        })
-    }
-}
-
-#[pyproto]
-impl PyMappingProtocol for IndexableString {
-    fn __getitem__(&self, item: PyObject) -> PyResult<String> {
-        // Find out the range
-        let range = get_range(item, self.s.len())?;
-
-        // Get the range from the relevant string
-        let s = match self.t {
-            IndexableStringType::Original => self.s.get_range_original(range),
-            IndexableStringType::Normalized => self.s.get_range(range),
-        };
-
-        s.map(|s| s.to_owned())
-            .ok_or_else(|| exceptions::IndexError::py_err("Wrong offsets"))
-    }
-
-    fn __len__(self) -> PyResult<usize> {
-        Ok(match self.t {
-            IndexableStringType::Original => self.s.len_original(),
-            IndexableStringType::Normalized => self.s.len(),
-        })
-    }
-}
-
 #[pyclass(dict)]
 #[repr(transparent)]
 pub struct Encoding {
@@ -127,7 +23,7 @@ impl PyObjectProtocol for Encoding {
     fn __repr__(&self) -> PyResult<String> {
         Ok(format!(
             "Encoding(num_tokens={}, attributes=[ids, type_ids, tokens, offsets, \
-             attention_mask, special_tokens_mask, overflowing, original_str, normalized_str])",
+             attention_mask, special_tokens_mask, overflowing])",
             self.encoding.get_ids().len()
         ))
     }
@@ -142,50 +38,6 @@ impl PySequenceProtocol for Encoding {
 
 #[pymethods]
 impl Encoding {
-    #[getter]
-    fn get_normalized_str(&self) -> IndexableString {
-        IndexableString {
-            s: self.encoding.get_normalized().clone(),
-            t: IndexableStringType::Normalized,
-        }
-    }
-
-    #[getter]
-    fn get_original_str(&self) -> IndexableString {
-        IndexableString {
-            s: self.encoding.get_normalized().clone(),
-            t: IndexableStringType::Original,
-        }
-    }
-
-    #[args(kwargs = "**")]
-    fn get_range(
-        &self,
-        range: (usize, usize),
-        kwargs: Option<&PyDict>,
-    ) -> PyResult<Option<String>> {
-        let mut original = false;
-        if let Some(kwargs) = kwargs {
-            if let Some(koriginal) = kwargs.get_item("original") {
-                original = koriginal.extract()?;
-            }
-        }
-
-        if original {
-            Ok(self
-                .encoding
-                .get_normalized()
-                .get_range_original(range.0..range.1)
-                .map(|s| s.to_owned()))
-        } else {
-            Ok(self
-                .encoding
-                .get_normalized()
-                .get_range(range.0..range.1)
-                .map(|s| s.to_owned()))
-        }
-    }
-
     #[getter]
     fn get_ids(&self) -> Vec<u32> {
         self.encoding.get_ids().to_vec()

bindings/python/src/tokenizer.rs

@@ -159,6 +159,15 @@ impl Tokenizer {
         self.tokenizer.with_padding(None);
     }
 
+    fn normalize(&self, sentence: &str) -> PyResult<String> {
+        ToPyResult(
+            self.tokenizer
+                .normalize(sentence)
+                .map(|s| s.get().to_owned()),
+        )
+        .into()
+    }
+
     #[args(add_special_tokens = true)]
     fn encode(
         &self,

bindings/python/tokenizers/__init__.pyi

@@ -16,32 +16,9 @@ from typing import Optional, Union, List, Tuple
 
 Offsets = Tuple[int, int]
 
-class IndexableString:
-    """
-    Works almost like a `str`, but allows indexing on offsets
-    provided on an `Encoding`
-    """
-
-    def offsets(self, offsets: Tuple[int, int]) -> Optional[Tuple[int, int]]:
-        """ Convert the Encoding's offsets to the current string.
-
-        `Encoding` provides a list of offsets that are actually offsets to the Normalized
-        version of text. Calling this method with the offsets provided by `Encoding` will make
-        sure that said offsets can be used to index the `str` directly.
-        """
-        pass
-
 class Encoding:
     """ An Encoding as returned by the Tokenizer """
 
-    @property
-    def normalized_str(self) -> IndexableString:
-        """ The normalized string """
-        pass
-    @property
-    def original_str(self) -> IndexableString:
-        """ The original string """
-        pass
     @property
     def ids(self) -> List[int]:
         """ The tokenized ids """
@@ -244,6 +221,17 @@ class Tokenizer:
     def no_padding(self):
         """ Disable padding """
         pass
+    def normalize(self, sequence: str) -> str:
+        """ Normalize the given sequence
+
+        Args:
+            sequence: str:
+                The sequence to normalize
+
+        Returns:
+            The normalized string
+        """
+        pass
     def encode(
         self, sequence: str, pair: Optional[str] = None, add_special_tokens: bool = True
     ) -> Encoding:

bindings/python/tokenizers/implementations/base_tokenizer.py

@@ -125,6 +125,18 @@ class BaseTokenizer:
         """
         return self._tokenizer.add_special_tokens(special_tokens)
 
+    def normalize(self, sequence: str) -> str:
+        """ Normalize the given sequence
+
+        Args:
+            sequence: str:
+                The sequence to normalize
+
+        Returns:
+            The normalized string
+        """
+        return self._tokenizer.normalize(sequence)
+
     def encode(
         self, sequence: str, pair: Optional[str] = None, add_special_tokens: bool = True
     ) -> Encoding:

tokenizers/CHANGELOG.md

@@ -6,9 +6,16 @@ a high number of files as it avoids having too many progress bars on screen.
 - Improve BPE and WordPiece builders.
 - `ByteLevel` is also a `PostProcessor` now and handles trimming the offsets if activated. This
 avoids the unintuitive inclusion of the whitespaces in the produced offsets, even if these
-whitespaces are part of the actual token.
+whitespaces are part of the actual token. ([#188](https://github.com/huggingface/tokenizers/pull/188))
 - `encode` and `encode_batch` now take a new argument, specifying whether we should add the
-special tokens.
+special tokens. ([#193](https://github.com/huggingface/tokenizers/pull/193))
+- The `NormalizedString` has been removed from the `Encoding`. It is now possible to retrieve it
+by calling `normalized` on the `Tokenizer`. This brings a reduction of 70% of the memory footprint
+([#197](https://github.com/huggingface/tokenizers/pull/197))
+- The `NormalizedString` API has been improved. It is now possible to retrieve part of both strings
+using both "normalized" or "original" offsets. ([#197](https://github.com/huggingface/tokenizers/pull/197))
+- The offsets provided on `Encoding` are now relative to the original string, and not the normalized
+one anymore. ([#197](https://github.com/huggingface/tokenizers/pull/197))
 
 ## Fixes:
 - Fix some issues with the offsets being wrong with the `ByteLevel` BPE:

tokenizers/Makefile

@@ -1,5 +1,12 @@
+DATA_DIR = data
 BENCHMARK_DIR = benches
-BENCHMARK_RESOURCES = $(BENCHMARK_DIR)/gpt2-vocab.json $(BENCHMARK_DIR)/gpt2-merges.txt $(BENCHMARK_DIR)/big.txt
+TESTS_DIR = tests
+
+dir_guard=@mkdir -p $(@D)
+
+SHARED_RESOURCES = $(DATA_DIR)/gpt2-vocab.json $(DATA_DIR)/gpt2-merges.txt
+BENCHMARK_RESOURCES = $(SHARED_RESOURCES) $(DATA_DIR)/big.txt
+TESTS_RESOURCES = $(SHARED_RESOURCES)
 
 .PHONY : build
 build :
@@ -20,7 +27,7 @@ lint :
 	cargo clippy --all-targets --all-features -- -D warnings
 
 .PHONY : test
-test :
+test : $(TESTS_RESOURCES)
 	cargo test
 
 .PHONY : doc
@@ -38,8 +45,10 @@ all-checks : lint test doc
 bench : $(BENCHMARK_RESOURCES)
 	cargo bench -- --verbose
 
-$(BENCHMARK_DIR)/gpt2-% :
+$(DATA_DIR)/gpt2-% :
+	$(dir_guard)
 	wget https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-$* -O $@
 
-$(BENCHMARK_DIR)/big.txt :
+$(DATA_DIR)/big.txt :
+	$(dir_guard)
 	wget https://norvig.com/big.txt -O $@

tokenizers/benches/.gitignore

@@ -1,2 +0,0 @@
-*.txt
-*.json

tokenizers/benches/bpe_benchmark.rs

@@ -68,13 +68,13 @@ fn iter_bench_encode_batch(
 }
 
 fn bench_gpt2(c: &mut Criterion) {
-    let bpe = BPE::from_files("benches/gpt2-vocab.json", "benches/gpt2-merges.txt")
+    let bpe = BPE::from_files("data/gpt2-vocab.json", "data/gpt2-merges.txt")
         .build()
         .unwrap();
     let tokenizer = create_gpt2_tokenizer(bpe);
     let mut lines: Vec<EncodeInput> = vec![];
     let mut batches: Vec<Vec<EncodeInput>> = vec![vec![]];
-    for line in BufReader::new(File::open(Path::new("benches/big.txt")).unwrap())
+    for line in BufReader::new(File::open(Path::new("data/big.txt")).unwrap())
         .lines()
         .map(line_to_input)
     {
@@ -93,7 +93,7 @@ fn bench_gpt2(c: &mut Criterion) {
         b.iter_custom(|iters| iter_bench_encode_batch(iters, &tokenizer, &batches))
     });
 
-    let bpe = BPE::from_files("benches/gpt2-vocab.json", "benches/gpt2-merges.txt")
+    let bpe = BPE::from_files("data/gpt2-vocab.json", "data/gpt2-merges.txt")
         .cache_capacity(0)
         .build()
         .unwrap();

tokenizers/src/pre_tokenizers/byte_level.rs

@@ -239,7 +239,7 @@ impl PostProcessor for ByteLevel {
             None => encoding,
             Some(mut pair) => {
                 process_offsets(&mut pair);
-                encoding.merge_with(pair);
+                encoding.merge_with(pair, false);
                 encoding
             }
         };
@@ -251,7 +251,9 @@ impl PostProcessor for ByteLevel {
 #[cfg(test)]
 mod tests {
     use super::ByteLevel;
-    use crate::tokenizer::{Decoder, Encoding, NormalizedString, PostProcessor, PreTokenizer};
+    use crate::tokenizer::{
+        Decoder, Encoding, NormalizedString, PostProcessor, PreTokenizer, Range,
+    };
 
     #[test]
     fn pre_tokenization() {
@@ -391,13 +393,12 @@ mod tests {
             ]
         );
         assert_eq!(input.get(), "iâŃ¢j");
-        assert_eq!(input.get_range_original(1..4), Some("⭢".into()));
+        assert_eq!(input.get_range_original(Range::Normalized(1..4)), Some("⭢"));
     }
 
     #[test]
     fn processor_trims_offsets() {
         let start = Encoding::new(
-            NormalizedString::from(""),
             vec![],
             vec![],
             vec![
@@ -412,7 +413,6 @@ mod tests {
             vec![],
         );
         let expected = Encoding::new(
-            NormalizedString::from(""),
             vec![],
             vec![],
             vec![
@@ -434,7 +434,7 @@ mod tests {
         );
 
         let mut pair_expected = expected.clone();
-        pair_expected.merge_with(expected);
+        pair_expected.merge_with(expected, false);
         assert_eq!(
             pair_expected,
             bytelevel

tokenizers/src/processors/bert.rs

@@ -43,7 +43,6 @@ impl PostProcessor for BertProcessing {
         let attention_mask = vec![1; ids.len()];
 
         let mut new_encoding = Encoding::new(
-            encoding.get_normalized().clone(),
             ids,
             type_ids,
             tokens,
@@ -68,7 +67,6 @@ impl PostProcessor for BertProcessing {
                     let attention_mask = vec![1; ids.len()];
 
                     Encoding::new(
-                        encoding.get_normalized().clone(),
                         ids,
                         type_ids,
                         tokens,
@@ -91,7 +89,6 @@ impl PostProcessor for BertProcessing {
             let pair_attention_mask = vec![1; pair_ids.len()];
 
             let new_pair_encoding = Encoding::new(
-                encoding.get_normalized().clone(),
                 pair_ids,
                 pair_type_ids,
                 pair_tokens,
@@ -112,7 +109,6 @@ impl PostProcessor for BertProcessing {
                         let pair_attention_mask = vec![1; pair_ids.len()];
 
                         Encoding::new(
-                            encoding.get_normalized().clone(),
                             pair_ids,
                             pair_type_ids,
                             pair_tokens,
@@ -125,7 +121,7 @@ impl PostProcessor for BertProcessing {
                     .collect(),
             );
 
-            new_encoding.merge_with(new_pair_encoding);
+            new_encoding.merge_with(new_pair_encoding, false);
         }
 
         Ok(new_encoding)

tokenizers/src/processors/roberta.rs

@@ -43,7 +43,6 @@ impl PostProcessor for RobertaProcessing {
         let attention_mask = vec![1; ids.len()];
 
         let mut new_encoding = Encoding::new(
-            encoding.get_normalized().clone(),
             ids,
             type_ids,
             tokens,
@@ -68,7 +67,6 @@ impl PostProcessor for RobertaProcessing {
             let pair_attention_mask = vec![1; pair_ids.len()];
 
             let new_pair_encoding = Encoding::new(
-                encoding.get_normalized().clone(),
                 pair_ids,
                 pair_type_ids,
                 pair_tokens,
@@ -78,7 +76,7 @@ impl PostProcessor for RobertaProcessing {
                 encoding.take_overflowing(),
             );
 
-            new_encoding.merge_with(new_pair_encoding);
+            new_encoding.merge_with(new_pair_encoding, false);
         }
 
         Ok(new_encoding)

tokenizers/src/tokenizer/encoding.rs

@@ -1,26 +1,9 @@
-use crate::tokenizer::NormalizedString;
+use crate::utils::padding::PaddingDirection;
 use rayon::prelude::*;
 
-/// The various possible padding directions.
-#[derive(Debug, Clone, Copy)]
-pub enum PaddingDirection {
-    Left,
-    Right,
-}
-
-impl std::convert::AsRef<str> for PaddingDirection {
-    fn as_ref(&self) -> &str {
-        match self {
-            PaddingDirection::Left => "left",
-            PaddingDirection::Right => "right",
-        }
-    }
-}
-
 /// Represents the output of a `Tokenizer`.
 #[derive(Default, PartialEq, Debug, Clone)]
 pub struct Encoding {
-    normalized: NormalizedString,
     ids: Vec<u32>,
     type_ids: Vec<u32>,
     tokens: Vec<String>,
@@ -32,7 +15,6 @@ pub struct Encoding {
 impl Encoding {
     #[allow(clippy::too_many_arguments)]
     pub fn new(
-        normalized: NormalizedString,
         ids: Vec<u32>,
         type_ids: Vec<u32>,
         tokens: Vec<String>,
@@ -42,7 +24,6 @@ impl Encoding {
         overflowing: Vec<Encoding>,
     ) -> Self {
         Encoding {
-            normalized,
             ids,
             type_ids,
             tokens,
@@ -53,10 +34,6 @@ impl Encoding {
         }
     }
 
-    pub fn get_normalized(&self) -> &NormalizedString {
-        &self.normalized
-    }
-
     pub fn get_tokens(&self) -> &[String] {
         &self.tokens[..]
     }
@@ -124,7 +101,6 @@ impl Encoding {
             }
 
             let o = Encoding {
-                normalized: self.normalized.clone(),
                 ids: get_current_part(&prev_encoding.ids, &o_ids, part_size, part_id, stride),
                 type_ids: get_current_part(
                     &prev_encoding.type_ids,
@@ -173,7 +149,7 @@ impl Encoding {
     }
 
     /// Merge ourself with the given `Encoding`. Happens in place.
-    pub fn merge_with(&mut self, pair: Encoding) {
+    pub fn merge_with(&mut self, pair: Encoding, growing_offsets: bool) {
         // Handle merging the overflowing parts too: Combine them all
         // In most of the cases, we expect `pair.overflowing.len() == 0`
         let mut overflowings = vec![];
@@ -182,33 +158,33 @@ impl Encoding {
         for self_o in &self.overflowing {
             // 1. The pair itself
             let mut n_encoding = self_o.clone();
-            n_encoding.merge_with(pair.clone());
+            n_encoding.merge_with(pair.clone(), growing_offsets);
             overflowings.push(n_encoding);
 
             // 2. Its overflowings (this should rarely happen...)
             for other_o in &pair.overflowing {
                 let mut n_encoding = self_o.clone();
-                n_encoding.merge_with(other_o.clone());
+                n_encoding.merge_with(other_o.clone(), growing_offsets);
                 overflowings.push(n_encoding);
             }
         }
         // 2. Ourself with all the other overflowings (this should rarely happen too...)
         for other_o in &pair.overflowing {
             let mut n_encoding = self.clone();
-            n_encoding.merge_with(other_o.clone());
+            n_encoding.merge_with(other_o.clone(), growing_offsets);
             overflowings.push(n_encoding);
         }
 
         // Finish by merging ourself with the other encoding
-        self.normalized.merge_with(&pair.normalized);
         self.ids.extend(pair.ids);
         self.type_ids.extend(pair.type_ids);
         self.tokens.extend(pair.tokens);
 
-        let starting_offset = self
+        let starting_offset = if growing_offsets {
-            .offsets
+            self.offsets.last().map_or(0, |o| o.1)
-            .iter()
+        } else {
-            .fold(0, |max, (_, end)| if *end > max { *end } else { max });
+            0
+        };
         self.offsets.extend(
             pair.offsets
                 .into_iter()
@@ -304,7 +280,6 @@ mod tests {
     #[test]
     fn merge_encodings() {
         let mut a = Encoding {
-            normalized: NormalizedString::from("Hello "),
             ids: vec![1],
             type_ids: vec![0],
             tokens: vec![String::from("Hello ")],
@@ -314,7 +289,6 @@ mod tests {
             overflowing: vec![],
         };
         let b = Encoding {
-            normalized: NormalizedString::from("World!"),
             ids: vec![2],
             type_ids: vec![1],
             tokens: vec![String::from("World!")],
@@ -323,12 +297,11 @@ mod tests {
             attention_mask: vec![1],
             overflowing: vec![],
         };
-        a.merge_with(b);
+        a.merge_with(b, true);
 
         assert_eq!(
             a,
             Encoding {
-                normalized: NormalizedString::from("Hello World!"),
                 ids: vec![1, 2],
                 type_ids: vec![0, 1],
                 tokens: vec![String::from("Hello "), String::from("World!")],
@@ -343,7 +316,6 @@ mod tests {
     #[test]
     fn truncate() {
         let mut a = Encoding {
-            normalized: NormalizedString::from("Hello World!"),
             ids: vec![1, 2, 3],
             type_ids: vec![0, 0, 0],
             tokens: vec![
@@ -361,7 +333,6 @@ mod tests {
         assert_eq!(
             a,
             Encoding {
-                normalized: NormalizedString::from("Hello World!"),
                 ids: vec![1, 2],
                 type_ids: vec![0, 0],
                 tokens: vec![String::from("Hello"), String::from("World")],
@@ -369,7 +340,6 @@ mod tests {
                 special_tokens_mask: vec![0, 0],
                 attention_mask: vec![1, 1],
                 overflowing: vec![Encoding {
-                    normalized: NormalizedString::from("Hello World!"),
                     ids: vec![3],
                     type_ids: vec![0],
                     tokens: vec![String::from("!")],

tokenizers/src/tokenizer/mod.rs

@@ -9,10 +9,9 @@
 //!   - [`PostProcessor`](trait.PostProcessor.html): Takes care of the processing after tokenization (like truncating, padding,
 //!   ...).
 
-pub use crate::utils::{
+use crate::utils::iter::ResultShunt;
-    pad_encodings, truncate_encodings, PaddingParams, PaddingStrategy, TruncationParams,
+pub use crate::utils::padding::{pad_encodings, PaddingDirection, PaddingParams, PaddingStrategy};
-    TruncationStrategy,
+pub use crate::utils::truncation::{truncate_encodings, TruncationParams, TruncationStrategy};
-};
 use indicatif::{ProgressBar, ProgressStyle};
 use rayon::prelude::*;
 use std::{
@@ -31,6 +30,11 @@ pub use normalizer::*;
 pub type Result<T> = std::result::Result<T, Box<dyn std::error::Error + Send + Sync>>;
 pub type Offsets = (usize, usize);
 
+/// Takes care of pre-processing strings.
+pub trait Normalizer {
+    fn normalize(&self, normalized: &mut NormalizedString) -> Result<()>;
+}
+
 /// The `PreTokenizer` is in charge of doing the pre-segmentation step. It splits the given string
 /// in multiple substrings, keeping track of the offsets of said substrings from the
 /// `NormalizedString`. In some occasions, the `PreTokenizer` might need to modify the given
@@ -71,7 +75,7 @@ impl dyn PostProcessor {
         match pair_encoding {
             None => Ok(encoding),
             Some(pair) => {
-                encoding.merge_with(pair);
+                encoding.merge_with(pair, false);
                 Ok(encoding)
             }
         }
@@ -290,25 +294,44 @@ impl Tokenizer {
         }
     }
 
-    pub fn num_added_tokens(&self, is_pair: bool) -> usize {
+    /// Normalize the given sentence and return the corresponding normalized string
-        self.post_processor
+    pub fn normalize(&self, sentence: &str) -> Result<NormalizedString> {
-            .as_ref()
+        let mut normalized = self
-            .map_or(0, |p| p.as_ref().added_tokens(is_pair))
+            .split_on_added_tokens(sentence)
+            .into_iter()
+            .map(|(sentence, id)| -> Result<NormalizedString> {
+                if id.is_some() {
+                    Ok(NormalizedString::from(&sentence))
+                } else {
+                    let mut normalized = self.do_normalize(&sentence)?;
+                    let _ = self.pre_tokenize(&mut normalized)?;
+
+                    Ok(normalized)
+                }
+            })
+            .collect::<Result<Vec<_>>>()?;
+
+        let others = normalized.split_off(1);
+        let mut normalized: NormalizedString = normalized.into_iter().next().unwrap();
+        for n in others {
+            normalized.merge_with(&n);
+        }
+
+        Ok(normalized)
     }
 
     /// Encode the given sentence
     pub fn encode(&self, input: EncodeInput, add_special_tokens: bool) -> Result<Encoding> {
-        let generate_output = move |sentence: String, type_id: u32| -> Result<Encoding> {
+        let generate_output =
+            move |sentence: String, type_id: u32| -> Result<(Encoding, NormalizedString)> {
                 // First we need to split into as many sequences as needed to avoid splitting
                 // on our added tokens
-            let mut encodings = self
+                let results = self.split_on_added_tokens(&sentence).into_iter().map(
-                .split_on_added_tokens(&sentence)
+                    |(sentence, id)| -> Result<(Encoding, NormalizedString)> {
-                .into_iter()
-                .map(|(sentence, id)| -> Result<Encoding> {
                         // If this is one of our added tokens, lets return an encoding directly
                         if let Some(id) = id {
-                        return Ok(Encoding::new(
+                            return Ok((
-                            NormalizedString::from(&sentence),
+                                Encoding::new(
                                     vec![id],
                                     vec![type_id],
                                     vec![sentence.to_owned()],
@@ -316,11 +339,13 @@ impl Tokenizer {
                                     vec![0],
                                     vec![1],
                                     vec![],
+                                ),
+                                NormalizedString::from(&sentence),
                             ));
                         }
 
                         // 1. Normalization
-                    let mut normalized = self.normalize(&sentence)?;
+                        let mut normalized = self.do_normalize(&sentence)?;
 
                         // 2. Pre tokenization
                         let pre_tokenized = self.pre_tokenize(&mut normalized)?;
@@ -343,8 +368,8 @@ impl Tokenizer {
                             },
                         );
 
-                    Ok(Encoding::new(
+                        Ok((
-                        normalized,
+                            Encoding::new(
                                 ids,
                                 vec![type_id; length],
                                 tokens,
@@ -352,22 +377,33 @@ impl Tokenizer {
                                 vec![0; length],
                                 vec![1; length],
                                 vec![],
+                            ),
+                            normalized,
                         ))
-                })
+                    },
-                .collect::<Result<Vec<Encoding>>>()?;
+                );
+
+                let (mut encodings, mut normalized) =
+                    ResultShunt::process(results, |iter| iter.unzip::<_, _, Vec<_>, Vec<_>>())?;
 
                 if encodings.is_empty() {
-                return Ok(Encoding::default());
+                    return Ok((Encoding::default(), NormalizedString::from("")));
                 }
 
                 let others = encodings.split_off(1);
                 let mut first: Encoding = encodings.into_iter().next().unwrap();
 
                 for encoding in others {
-                first.merge_with(encoding);
+                    first.merge_with(encoding, true);
                 }
 
-            Ok(first)
+                let others = normalized.split_off(1);
+                let mut normalized: NormalizedString = normalized.into_iter().next().unwrap();
+                for n in others {
+                    normalized.merge_with(&n);
+                }
+
+                Ok((first, normalized))
             };
 
         let (sentence, pair) = match input {
@@ -375,14 +411,37 @@ impl Tokenizer {
             EncodeInput::Dual(s1, s2) => (s1, Some(s2)),
         };
 
-        let encoding = generate_output(sentence, 0)?;
+        let (encoding, normalized) = generate_output(sentence, 0)?;
-        let pair_encoding = match pair {
+        let (pair_encoding, pair_normalized) = match pair {
-            Some(pair) => Some(generate_output(pair, 1)?),
+            Some(pair) => {
-            None => None,
+                let (e, n) = generate_output(pair, 1)?;
+                (Some(e), Some(n))
+            }
+            None => (None, None),
         };
 
         // 4. Post processing
-        self.post_process(encoding, pair_encoding, add_special_tokens)
+        let mut output = self.post_process(encoding, pair_encoding, add_special_tokens)?;
+
+        // 5. Convert offsets back to original string
+        let mut current_offset = (0, 0);
+        let mut n_source = &normalized;
+        output
+            .get_offsets_mut()
+            .iter_mut()
+            .for_each(|(start, end)| {
+                if (*start, *end) < current_offset {
+                    n_source = &pair_normalized.as_ref().unwrap_or(&normalized);
+                }
+                current_offset = (*start, *end);
+                let (s, e) = n_source
+                    .convert_offsets(Range::Normalized(*start..*end))
+                    .map_or((*start, *end), |range| (range.start, range.end));
+                *start = s;
+                *end = e;
+            });
+
+        Ok(output)
     }
 
     /// Encode all the sentences in parallel, using multiple threads
@@ -471,7 +530,7 @@ impl Tokenizer {
                     match file.read_line(&mut buf)? {
                         0 => break,
                         b => {
-                            let mut normalized = self.normalize(&buf)?;
+                            let mut normalized = self.do_normalize(&buf)?;
                             let pre_tokenized = self.pre_tokenize(&mut normalized)?;
                             trainer.process_tokens(
                                 &mut words,
@@ -522,7 +581,7 @@ impl Tokenizer {
     }
 
     /// Normalization logic, go through all normalizers
-    fn normalize(&self, sequence: &str) -> Result<NormalizedString> {
+    fn do_normalize(&self, sequence: &str) -> Result<NormalizedString> {
         let mut normalized = NormalizedString::from(sequence);
 
         if let Some(normalizer) = &self.normalizer {

tokenizers/src/tokenizer/normalizer.rs

@@ -1,21 +1,63 @@
-use super::Result;
 use std::cmp::Ordering;
+use std::ops::{Bound, RangeBounds};
 use unicode_normalization_alignments::UnicodeNormalization;
 
-/// Takes care of pre-processing strings.
+/// Represents a Range usable by the NormalizedString to index its content.
-pub trait Normalizer {
+/// A Range can use indices relative to either the `Original` or the `Normalized` string
-    fn normalize(&self, normalized: &mut NormalizedString) -> Result<()>;
+#[derive(Debug, Clone, Copy)]
+pub enum Range<T: RangeBounds<usize>> {
+    Original(T),
+    Normalized(T),
 }
 
-/// A normalized string takes care of keeping both versions of a `String`, and
+impl<T> Range<T>
-/// provides necessary alignments to retrieve ranges of both strings.
+where
+    T: RangeBounds<usize>,
+{
+    /// Unwrap the underlying range
+    fn unwrap(self) -> T {
+        match self {
+            Range::Original(r) => r,
+            Range::Normalized(r) => r,
+        }
+    }
+
+    /// Converts the current Range to a `std::ops::Range<usize>`. This requires the `max_len`
+    /// of the represented string (in chars, not bytes) in order to cover the case where the
+    /// original provided range was unbounded
+    fn into_full_range(self, max_len: usize) -> std::ops::Range<usize> {
+        let range = self.unwrap();
+
+        let start = match range.start_bound() {
+            Bound::Unbounded => 0,
+            Bound::Included(i) => *i,
+            Bound::Excluded(i) => *i + 1,
+        };
+        let end = match range.end_bound() {
+            Bound::Unbounded => max_len,
+            Bound::Included(i) => *i + 1,
+            Bound::Excluded(i) => *i,
+        };
+
+        start..end
+    }
+}
+
+/// A `NormalizedString` takes care of processing an "original" string to modify it and obtain a
+/// "normalized" string. It keeps both version of the string, alignments information between both
+/// and provides an interface to retrieve ranges of each string, using offsets from any of them.
+///
+/// It is possible to retrieve a part of the original string, by indexing it with offsets from the
+/// normalized one, and the other way around too. It is also possible to convert offsets from one
+/// referential to the other one easily.
 #[derive(Default, Debug, Clone)]
 pub struct NormalizedString {
+    /// The original version of the string, before any modification
     original: String,
+    /// The normalized version of the string, after all modifications
     normalized: String,
-    /// Mapping from normalized string to original one
+    /// Mapping from normalized string to original one: (start, end) for each character of the
-    /// (pos, changes) where pos is the position in the modified string, and changes an isize
+    /// normalized string
-    /// representing the number of insertions or deletions
     alignments: Vec<(usize, usize)>,
 }
 
@@ -26,6 +68,7 @@ impl std::cmp::PartialEq for NormalizedString {
 }
 
 impl NormalizedString {
+    /// Create a NormalizedString from the given str
     pub fn from(s: &str) -> Self {
         NormalizedString {
             original: s.to_owned(),
@@ -44,14 +87,35 @@ impl NormalizedString {
         &self.original
     }
 
-    /// Return the range of the original string corresponding to the received range on the
+    /// Convert the given offsets range from one referential to the other one:
-    /// normalized string. Returns None if out of bounds
+    /// `Original => Normalized` or `Normalized => Original`
-    pub fn get_original_offsets(
+    pub fn convert_offsets<T: RangeBounds<usize>>(
         &self,
-        range: std::ops::Range<usize>,
+        range: Range<T>,
     ) -> Option<std::ops::Range<usize>> {
+        match range {
+            Range::Original(_) => {
+                let (mut start, mut end) = (0, 0);
+                let r = range.into_full_range(self.alignments.last().map_or(0, |(_, e)| *e));
+                println!("{:?}\t{:?}", r, self.alignments);
                 self.alignments
-            .get(range)
+                    .iter()
+                    .enumerate()
+                    .take_while(|(_, alignment)| r.end >= alignment.1)
+                    .for_each(|(i, alignment)| {
+                        println!("{:?}", alignment);
+                        if alignment.0 <= r.start {
+                            start = i;
+                        }
+                        if alignment.1 <= r.end {
+                            end = i + 1;
+                        }
+                    });
+                Some(start..end)
+            }
+            Range::Normalized(_) => self
+                .alignments
+                .get(range.into_full_range(self.alignments.len()))
                 .map(|alignments| {
                     if alignments.is_empty() {
                         None
@@ -61,43 +125,30 @@ impl NormalizedString {
                         Some(start..end)
                     }
                 })
-            .flatten()
+                .flatten(),
-    }
-
-    fn get_range_of(&self, s: &str, range: std::ops::Range<usize>) -> Option<String> {
-        let len = s.chars().count();
-        if range.start >= len || range.end > len {
-            None
-        } else {
-            Some(
-                s.chars()
-                    .enumerate()
-                    .skip(range.start)
-                    .map(|(i, c)| {
-                        if i >= range.start && i < range.end {
-                            Some(c)
-                        } else {
-                            None
-                        }
-                    })
-                    .fuse()
-                    .filter(|c| c.is_some())
-                    .map(|c| c.unwrap())
-                    .collect::<String>(),
-            )
         }
     }
 
     /// Return a range of the normalized string (indexing on char not bytes)
-    pub fn get_range(&self, range: std::ops::Range<usize>) -> Option<String> {
+    pub fn get_range<T: RangeBounds<usize>>(&self, range: Range<T>) -> Option<&str> {
-        self.get_range_of(&self.normalized, range)
+        match range {
+            Range::Original(_) => self
+                .convert_offsets(range)
+                .map(|r| get_range_of(&self.normalized, r))
+                .flatten(),
+            Range::Normalized(r) => get_range_of(&self.normalized, r),
+        }
     }
 
-    /// Return a range of the original string, using a range from the normalized string
+    /// Return a range of the original string (indexing on char not bytes)
-    pub fn get_range_original(&self, range: std::ops::Range<usize>) -> Option<String> {
+    pub fn get_range_original<T: RangeBounds<usize>>(&self, range: Range<T>) -> Option<&str> {
-        self.get_original_offsets(range)
+        match range {
-            .map(|range| self.get_range_of(&self.original, range))
+            Range::Original(r) => get_range_of(&self.original, r),
-            .flatten()
+            Range::Normalized(_) => self
+                .convert_offsets(range)
+                .map(|r| get_range_of(&self.original, r))
+                .flatten(),
+        }
     }
 
     /// Applies transformations to the current normalized version, updating the current
@@ -115,19 +166,10 @@ impl NormalizedString {
     /// them has a `change` of `1`, but more doesn't make any sense.
     /// We treat any value above `1` as `1`.
     pub fn transform<I: Iterator<Item = (char, isize)>>(&mut self, dest: I, initial_offset: usize) {
-        let mut offset = 0;
+        let mut offset = -(initial_offset as isize);
-        let mut remaining_offset = initial_offset;
         let (ch, alignments): (Vec<_>, Vec<_>) = dest
             .enumerate()
             .map(|(index, (c, changes))| {
-                let changes = if remaining_offset != 0 {
-                    let c = changes - remaining_offset as isize;
-                    remaining_offset = 0;
-                    c
-                } else {
-                    changes
-                };
-
                 let uof = if offset < 0 {
                     -offset as usize
                 } else {
@@ -149,24 +191,10 @@ impl NormalizedString {
                     }
                     // No changes required here
                     Ordering::Equal => self.alignments.get(idx).copied(),
-                    // Some characters where removed, so we merge our range with the one from the
+                    // Some characters where removed, nothing to change in alignments
-                    // removed characters as the new alignment
                     Ordering::Less => {
-                        let uch = -changes as usize;
                         offset += changes;
-                        self.alignments.get(idx..=idx + uch).map(|alignments| {
+                        self.alignments.get(idx).copied()
-                            let min = alignments
-                                .iter()
-                                .map(|(start, end)| usize::min(*start, *end))
-                                .min()
-                                .unwrap();
-                            let max = alignments
-                                .iter()
-                                .map(|(start, end)| usize::max(*start, *end))
-                                .max()
-                                .unwrap();
-                            (min, max)
-                        })
                     }
                 };
 
@@ -421,6 +449,37 @@ impl NormalizedString {
     }
 }
 
+/// Returns a range of the given string slice, by indexing chars instead of bytes
+pub fn get_range_of<T: RangeBounds<usize>>(s: &str, range: T) -> Option<&str> {
+    let len = s.chars().count();
+    let start = match range.start_bound() {
+        Bound::Unbounded => 0,
+        Bound::Included(i) => *i,
+        Bound::Excluded(i) => *i + 1,
+    };
+    let end = match range.end_bound() {
+        Bound::Unbounded => len,
+        Bound::Included(i) => *i + 1,
+        Bound::Excluded(i) => *i,
+    };
+
+    if start >= len || end > len || start >= end {
+        None
+    } else {
+        let start_b = s
+            .char_indices()
+            .map(|(i, _)| i)
+            .nth(start as usize)
+            .unwrap_or(0);
+        let end_b = s
+            .char_indices()
+            .map(|(i, _)| i)
+            .nth(end as usize)
+            .unwrap_or_else(|| s.len());
+        Some(&s[start_b..end_b])
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -462,7 +521,7 @@ mod tests {
         n.filter(|c| *c != 'n');
         assert_eq!(
             &n.alignments,
-            &[(0, 1), (1, 2), (2, 3), (3, 4), (4, 6), (6, 7)]
+            &[(0, 1), (1, 2), (2, 3), (3, 4), (4, 5), (6, 7)]
         );
     }
 
@@ -472,18 +531,43 @@ mod tests {
         n.nfd().filter(|c| !c.is_mark_nonspacing() && *c != 'n');
         assert_eq!(
             &n.alignments,
-            &[(0, 1), (1, 2), (2, 3), (3, 4), (4, 6), (6, 7)]
+            &[(0, 1), (1, 2), (2, 3), (3, 4), (4, 5), (6, 7)]
         );
     }
 
+    #[test]
+    fn range_conversion() {
+        let mut n = NormalizedString::from("    __Hello__   ");
+        n.filter(|c| !c.is_whitespace()).lowercase();
+        let hello_n = n.convert_offsets(Range::Original(6..11));
+        assert_eq!(hello_n, Some(2..7));
+        assert_eq!(
+            n.get_range(Range::Normalized(hello_n.clone().unwrap())),
+            Some("hello")
+        );
+        assert_eq!(
+            n.get_range_original(Range::Normalized(hello_n.unwrap())),
+            Some("Hello")
+        );
+        assert_eq!(n.get_range(Range::Original(6..11)), Some("hello"));
+        assert_eq!(n.get_range_original(Range::Original(6..11)), Some("Hello"));
+    }
+
     #[test]
     fn original_range() {
         let mut n = NormalizedString::from("Hello_______ World!");
         n.filter(|c| *c != '_').lowercase();
-        let world_n = n.get_range(6..11).unwrap();
+        let world_n = n.get_range(Range::Normalized(6..11)).unwrap();
-        let world_o = n.get_range_original(6..11).unwrap();
+        let world_o = n.get_range_original(Range::Normalized(6..11)).unwrap();
         assert_eq!(world_n, "world");
         assert_eq!(world_o, "World");
+        let original_range = Range::Original(n.convert_offsets(Range::Normalized(6..11)).unwrap());
+        assert_eq!(n.get_range(original_range.clone()).unwrap(), "world");
+        assert_eq!(
+            n.get_range_original(original_range.clone()).unwrap(),
+            "World"
+        );
+        assert_eq!(original_range.into_full_range(n.len_original()), 13..18);
     }
 
     #[test]
@@ -505,8 +589,8 @@ mod tests {
 
         assert_eq!(&n.normalized, " Hello ");
         assert_eq!(
-            n.get_range_original(0..n.normalized.len()),
+            n.get_range_original(Range::Normalized(1..n.normalized.len() - 1)),
-            Some("Hello".into())
+            Some("Hello")
         );
     }
 
@@ -514,10 +598,13 @@ mod tests {
     fn remove_at_beginning() {
         let mut n = NormalizedString::from("     Hello");
         n.filter(|c| !c.is_whitespace());
-        assert_eq!(n.get_range_original(1.."Hello".len()), Some("ello".into()));
         assert_eq!(
-            n.get_range_original(0..n.normalized.len()),
+            n.get_range_original(Range::Normalized(1.."Hello".len())),
-            Some("     Hello".into())
+            Some("ello")
+        );
+        assert_eq!(
+            n.get_range_original(Range::Normalized(0..n.normalized.len())),
+            Some("Hello")
         );
     }
 
@@ -525,10 +612,10 @@ mod tests {
     fn remove_at_end() {
         let mut n = NormalizedString::from("Hello    ");
         n.filter(|c| !c.is_whitespace());
-        assert_eq!(n.get_range_original(0..4), Some("Hell".into()));
+        assert_eq!(n.get_range_original(Range::Normalized(0..4)), Some("Hell"));
         assert_eq!(
-            n.get_range_original(0..n.normalized.len()),
+            n.get_range_original(Range::Normalized(0..n.normalized.len())),
-            Some("Hello    ".into())
+            Some("Hello")
         );
     }
 
@@ -539,10 +626,13 @@ mod tests {
         assert_eq!(&n.normalized, "Hello");
 
         assert_eq!(
-            n.get_range_original(0.."Hello".len()),
+            n.get_range_original(Range::Normalized(0.."Hello".len())),
-            Some("  Hello  ".into())
+            Some("Hello")
+        );
+        assert_eq!(
+            n.get_range_original(Range::Normalized(1.."Hell".len())),
+            Some("ell")
         );
-        assert_eq!(n.get_range_original(1.."Hell".len()), Some("ell".into()));
     }
 
     #[test]
@@ -551,8 +641,8 @@ mod tests {
         n.lstrip();
         assert_eq!(&n.normalized, "This is an example  ");
         assert_eq!(
-            n.get_range_original(0..n.normalized.len()),
+            n.get_range_original(Range::Normalized(0..n.normalized.len())),
-            Some("  This is an example  ".into())
+            Some("This is an example  ")
         );
     }
 
@@ -562,8 +652,8 @@ mod tests {
         n.rstrip();
         assert_eq!(&n.normalized, "  This is an example");
         assert_eq!(
-            n.get_range_original(0..n.normalized.len()),
+            n.get_range_original(Range::Normalized(0..n.normalized.len())),
-            Some("  This is an example  ".into())
+            Some("  This is an example")
         );
     }
 
@@ -573,8 +663,8 @@ mod tests {
         n.strip();
         assert_eq!(&n.normalized, "This is an example");
         assert_eq!(
-            n.get_range_original(0..n.normalized.len()),
+            n.get_range_original(Range::Normalized(0..n.normalized.len())),
-            Some("  This is an example  ".into())
+            Some("This is an example")
         );
     }
 
@@ -597,7 +687,7 @@ mod tests {
                 (4, 5)
             ]
         );
-        assert_eq!(n.get_original_offsets(0..4), Some(0..0));
+        assert_eq!(n.convert_offsets(Range::Normalized(0..4)), Some(0..0));
     }
 
     #[test]
@@ -619,6 +709,16 @@ mod tests {
                 (3, 3)
             ]
         );
-        assert_eq!(n.get_original_offsets(3.." there".len()), Some(3..3));
+        assert_eq!(
+            n.convert_offsets(Range::Normalized(3.." there".len())),
+            Some(3..3)
+        );
+    }
+
+    #[test]
+    fn get_range() {
+        let s = String::from("Hello my name is John 👋");
+        assert_eq!(get_range_of(&s, ..), Some(&s[..]));
+        assert_eq!(get_range_of(&s, 17..), Some("John 👋"));
     }
 }

tokenizers/src/utils/iter.rs

@@ -0,0 +1,58 @@
+//! This comes from the Rust libcore and is duplicated here because it is not exported
+//! (cf https://github.com/rust-lang/rust/blob/25091ed9b7739e12466fb2490baa1e8a2815121c/src/libcore/iter/adapters/mod.rs#L2664)
+//! We are now using the version from https://stackoverflow.com/questions/44544323/how-to-unzip-a-sequence-of-resulta-b-e-to-a-veca-vecb-and-stop-on-f
+//! because the one from the libcore seems to cause overflowing stacks in some cases
+
+pub struct ResultShunt<I, E> {
+    iter: I,
+    error: Option<E>,
+}
+
+impl<I, T, E> ResultShunt<I, E>
+where
+    I: Iterator<Item = Result<T, E>>,
+{
+    /// Process the given iterator as if it yielded a `T` instead of a
+    /// `Result<T, _>`. Any errors will stop the inner iterator and
+    /// the overall result will be an error.
+    pub fn process<F, U>(iter: I, mut f: F) -> Result<U, E>
+    where
+        F: FnMut(&mut Self) -> U,
+    {
+        let mut shunt = ResultShunt::new(iter);
+        let value = f(shunt.by_ref());
+        shunt.reconstruct(value)
+    }
+
+    fn new(iter: I) -> Self {
+        ResultShunt { iter, error: None }
+    }
+
+    /// Consume the adapter and rebuild a `Result` value. This should
+    /// *always* be called, otherwise any potential error would be
+    /// lost.
+    fn reconstruct<U>(self, val: U) -> Result<U, E> {
+        match self.error {
+            None => Ok(val),
+            Some(e) => Err(e),
+        }
+    }
+}
+
+impl<I, T, E> Iterator for ResultShunt<I, E>
+where
+    I: Iterator<Item = Result<T, E>>,
+{
+    type Item = T;
+
+    fn next(&mut self) -> Option<Self::Item> {
+        match self.iter.next() {
+            Some(Ok(v)) => Some(v),
+            Some(Err(e)) => {
+                self.error = Some(e);
+                None
+            }
+            None => None,
+        }
+    }
+}

tokenizers/src/utils/mod.rs

@@ -0,0 +1,3 @@
+pub mod iter;
+pub mod padding;
+pub mod truncation;

tokenizers/src/utils/padding.rs

@@ -0,0 +1,63 @@
+use crate::tokenizer::{Encoding, Result};
+use rayon::prelude::*;
+
+/// The various possible padding directions.
+#[derive(Debug, Clone, Copy)]
+pub enum PaddingDirection {
+    Left,
+    Right,
+}
+
+impl std::convert::AsRef<str> for PaddingDirection {
+    fn as_ref(&self) -> &str {
+        match self {
+            PaddingDirection::Left => "left",
+            PaddingDirection::Right => "right",
+        }
+    }
+}
+
+#[derive(Debug, Clone)]
+pub struct PaddingParams {
+    pub strategy: PaddingStrategy,
+    pub direction: PaddingDirection,
+    pub pad_id: u32,
+    pub pad_type_id: u32,
+    pub pad_token: String,
+}
+
+#[derive(Debug, Clone)]
+pub enum PaddingStrategy {
+    BatchLongest,
+    Fixed(usize),
+}
+
+pub fn pad_encodings(
+    mut encodings: Vec<Encoding>,
+    params: &PaddingParams,
+) -> Result<Vec<Encoding>> {
+    if encodings.is_empty() {
+        return Ok(encodings);
+    }
+
+    let pad_length = match params.strategy {
+        PaddingStrategy::Fixed(size) => size,
+        PaddingStrategy::BatchLongest => encodings
+            .par_iter()
+            .map(|e| e.get_ids().len())
+            .max()
+            .unwrap(),
+    };
+
+    encodings.par_iter_mut().for_each(|encoding| {
+        encoding.pad(
+            pad_length,
+            params.pad_id,
+            params.pad_type_id,
+            &params.pad_token,
+            params.direction,
+        )
+    });
+
+    Ok(encodings)
+}

tokenizers/src/utils/truncation.rs

@@ -1,5 +1,4 @@
-use crate::tokenizer::{Encoding, PaddingDirection, Result};
+use crate::tokenizer::{Encoding, Result};
-use rayon::prelude::*;
 
 #[derive(Debug, Clone)]
 pub struct TruncationParams {
@@ -8,21 +7,6 @@ pub struct TruncationParams {
     pub stride: usize,
 }
 
-#[derive(Debug, Clone)]
-pub struct PaddingParams {
-    pub strategy: PaddingStrategy,
-    pub direction: PaddingDirection,
-    pub pad_id: u32,
-    pub pad_type_id: u32,
-    pub pad_token: String,
-}
-
-#[derive(Debug, Clone)]
-pub enum PaddingStrategy {
-    BatchLongest,
-    Fixed(usize),
-}
-
 #[derive(Debug)]
 pub enum Error {
     SecondSequenceNotProvided,
@@ -118,33 +102,3 @@ pub fn truncate_encodings(
 
     Ok((encoding, pair_encoding))
 }
-
-pub fn pad_encodings(
-    mut encodings: Vec<Encoding>,
-    params: &PaddingParams,
-) -> Result<Vec<Encoding>> {
-    if encodings.is_empty() {
-        return Ok(encodings);
-    }
-
-    let pad_length = match params.strategy {
-        PaddingStrategy::Fixed(size) => size,
-        PaddingStrategy::BatchLongest => encodings
-            .par_iter()
-            .map(|e| e.get_ids().len())
-            .max()
-            .unwrap(),
-    };
-
-    encodings.par_iter_mut().for_each(|encoding| {
-        encoding.pad(
-            pad_length,
-            params.pad_id,
-            params.pad_type_id,
-            &params.pad_token,
-            params.direction,
-        )
-    });
-
-    Ok(encodings)
-}

tokenizers/tests/offsets.rs

@@ -0,0 +1,154 @@
+use tokenizers::models::bpe::BPE;
+use tokenizers::pre_tokenizers::byte_level::ByteLevel;
+use tokenizers::tokenizer::{get_range_of, EncodeInput, Tokenizer};
+
+fn get_byte_level(add_prefix_space: bool, trim_offsets: bool) -> Tokenizer {
+    let mut tokenizer = Tokenizer::new(Box::new(
+        BPE::from_files("data/gpt2-vocab.json", "data/gpt2-merges.txt")
+            .build()
+            .expect("Files not found, run `make test` to download these files"),
+    ));
+    tokenizer.with_pre_tokenizer(Box::new(
+        ByteLevel::default().add_prefix_space(add_prefix_space),
+    ));
+    tokenizer.with_decoder(Box::new(ByteLevel::default()));
+    tokenizer.with_post_processor(Box::new(ByteLevel::default().trim_offsets(trim_offsets)));
+
+    tokenizer
+}
+
+#[inline]
+fn offset_as_range(offset: (usize, usize)) -> std::ops::Range<usize> {
+    offset.0..offset.1
+}
+
+#[test]
+fn byte_level_basic() {
+    // Without trimming offsets
+    let tokenizer = get_byte_level(true, false);
+
+    let input = String::from("Hello there, how are you?");
+    let output = tokenizer
+        .encode(EncodeInput::Single(input.clone()), false)
+        .unwrap();
+
+    let offsets = output.get_offsets();
+    assert_eq!(
+        get_range_of(&input, offset_as_range(offsets[0])),
+        Some("Hello")
+    );
+    assert_eq!(
+        get_range_of(&input, offset_as_range(offsets[1])),
+        Some(" there")
+    );
+    assert_eq!(get_range_of(&input, offset_as_range(offsets[2])), Some(","));
+    assert_eq!(
+        get_range_of(&input, offset_as_range(offsets[3])),
+        Some(" how")
+    );
+    assert_eq!(
+        get_range_of(&input, offset_as_range(offsets[4])),
+        Some(" are")
+    );
+    assert_eq!(
+        get_range_of(&input, offset_as_range(offsets[5])),
+        Some(" you")
+    );
+    assert_eq!(get_range_of(&input, offset_as_range(offsets[6])), Some("?"));
+
+    // And when trimming offsets:
+    let tokenizer = get_byte_level(true, true);
+
+    let input = String::from("Hello there, how are you?");
+    let output = tokenizer
+        .encode(EncodeInput::Single(input.clone()), false)
+        .unwrap();
+
+    let offsets = output.get_offsets();
+    assert_eq!(
+        get_range_of(&input, offset_as_range(offsets[0])),
+        Some("Hello")
+    );
+    assert_eq!(
+        get_range_of(&input, offset_as_range(offsets[1])),
+        Some("there")
+    );
+    assert_eq!(get_range_of(&input, offset_as_range(offsets[2])), Some(","));
+    assert_eq!(
+        get_range_of(&input, offset_as_range(offsets[3])),
+        Some("how")
+    );
+    assert_eq!(
+        get_range_of(&input, offset_as_range(offsets[4])),
+        Some("are")
+    );
+    assert_eq!(
+        get_range_of(&input, offset_as_range(offsets[5])),
+        Some("you")
+    );
+    assert_eq!(get_range_of(&input, offset_as_range(offsets[6])), Some("?"));
+}
+
+#[test]
+fn byte_level_unicode() {
+    let tokenizer = get_byte_level(true, false);
+
+    let input = String::from("i⭢j");
+    let output = tokenizer
+        .encode(EncodeInput::Single(input.clone()), false)
+        .unwrap();
+
+    let offsets = output.get_offsets();
+    assert_eq!(get_range_of(&input, offset_as_range(offsets[1])), Some("⭢"));
+    assert_eq!(get_range_of(&input, offset_as_range(offsets[2])), Some("⭢"));
+    assert_eq!(get_range_of(&input, offset_as_range(offsets[3])), Some("⭢"));
+}
+
+#[test]
+fn byte_level_double_sequence() {
+    let input_a = String::from("My name is Anthony");
+    let input_b = String::from("What is my name?");
+
+    // Without trimming offsets
+    let tokenizer = get_byte_level(true, false);
+    let output = tokenizer
+        .encode(EncodeInput::Dual(input_a.clone(), input_b.clone()), false)
+        .unwrap();
+
+    let offsets = output.get_offsets();
+    assert_eq!(
+        offsets,
+        &[
+            (0, 2),
+            (2, 7),
+            (7, 10),
+            (10, 18),
+            (0, 4),
+            (4, 7),
+            (7, 10),
+            (10, 15),
+            (15, 16)
+        ]
+    );
+
+    // When trimming offsets
+    let tokenizer = get_byte_level(true, true);
+    let output = tokenizer
+        .encode(EncodeInput::Dual(input_a, input_b), false)
+        .unwrap();
+    let offsets = output.get_offsets();
+    assert_eq!(
+        offsets,
+        &[
+            (0, 2),
+            (3, 7),
+            (8, 10),
+            (11, 18),
+            (0, 4),
+            (5, 7),
+            (8, 10),
+            (11, 15),
+            (15, 16)
+        ]
+    );
+}