#nlp #ngrams #bare-metal #combinator #computer-vision #features #composable

bin+lib creature_feature

Composable n-gram combinators that are ergonomic and bare-metal fast

7 releases

0.1.7 Jul 12, 2023
0.1.6 Jul 6, 2023
0.1.4 May 3, 2023
0.1.2 Dec 21, 2022
0.1.1 Dec 24, 2021

#76 in Machine learning

MPL-2.0 license

145KB
2K SLoC

CREATURE FEATUR(ization)

A crate for polymorphic ML & NLP featurization that leverages zero-cost abstraction. It provides composable n-gram combinators that are ergonomic and bare-metal fast. Although created with NLP in mind, it's very general and can be applied in a plethera of domains such as computer vision.

There are many n-gram crates, but the majority force heap allocation or lock you into a concrete type that doesn’t fit your use-case or performance needs. In most benchmarks, creature_feature is anywhere between 4x - 60x faster.

Image

See a live demo

Here is a live demo of creature_feature using WASM

A Swiss Army Knife of Featurization

use creature_feature::traits::Ftzr;
use creature_feature::ftzrs::{bigram, bislice, for_each, whole};
use creature_feature::HashedAs;
use creature_feature::convert::Bag;
use std::collections::{HashMap, HashSet, BTreeMap, LinkedList, BTreeSet, BinaryHeap, VecDeque};

let int_data = &[1, 2, 3, 4, 5];
let str_data = "one fish two fish red fish blue fish";

// notice how the left-hand side remains almost unchanged.

// we're using 'bislice' right now (which is a 2-gram of referenced data), 'ftzrs::bigram' would yield owned data instead of references

let ref_feats: Vec<&str>                  = bislice().featurize(str_data);
let ref_feats: LinkedList<&[u8]>          = bislice().featurize(str_data);
let ref_bag:   Bag<HashMap<&[usize], u8>> = bislice().featurize(int_data);
let ref_trigram_bag:   Bag<BTreeMap<&str, i16>>   = for_each(whole()).featurize(str_data.split_ascii_whitespace());
let hashed_trigrams: BTreeSet<HashedAs<u64>> = trislice().featurize(int_data);

The above five will have the following values, respectively.

["on", "ne", "e ", " f", "fi", "is", "sh", "h ", " t", "tw", "wo", "o ", " f", "fi", "is", "sh", "h ", " r", "re", "ed", "d ", " f", "fi", "is", "sh", "h ", " b", "bl", "lu", "ue", "e ", " f", "fi", "is", "sh"]

[[111, 110], [110, 101], [101, 32], [32, 102], [102, 105], [105, 115], [115, 104], [104, 32], [32, 116], [116, 119], [119, 111], [111, 32], [32, 102], [102, 105], [105, 115], [115, 104], [104, 32], [32, 114], [114, 101], [101, 100], [100, 32], [32, 102], [102, 105], [105, 115], [115, 104], [104, 32], [32, 98], [98, 108], [108, 117], [117, 101], [101, 32], [32, 102], [102, 105], [105, 115], [115, 104]]

Bag({[2, 3, 4]: 1, [3, 4, 5]: 1, [1, 2, 3]: 1})

Bag({"blue": 1, "fish": 4, "one": 1, "red": 1, "two": 1})

{HashedAs(3939941806544028562), HashedAs(7191405660579021101), HashedAs(16403185381100005216)}

Here are more examples of what's possible:

 // let's now switch to 'bigram'
let owned_feats: BTreeSet<[u8; 2]>        = bigram().featurize(str_data);
let owned_feats: Vec<String>              = bigram().featurize(str_data);
let owned_feats: HashSet<Vec<usize>>      = bigram().featurize(int_data);
let owned_bag:   Bag<HashMap<Vec<usize>, u16>>      = bigram().featurize(int_data);

let hashed_feats: BinaryHeap<HashedAs<u32>> = bislice().featurize(str_data);
let hashed_feats: VecDeque<HashedAs<u64>>   =  bigram().featurize(int_data);

let sentence = str_data.split_ascii_whitespace();
let bag_of_words: Bag<HashMap<String, u128>> = for_each(bigram()) .featurize(sentence.clone());
let bag_of_words: Bag<HashMap<&str, u8>>     = for_each(bislice()).featurize(sentence.clone());

 // and many, MANY more posibilities

We can even produce multiple outputs while still only featurizing the input once

let (set, list): (BTreeSet<HashedAs<u64>>, Vec<&str>) = bislice().featurize_x2(str_data);

creature_feature provides three general flavors of featurizers:

  1. NGram<const N: usize> provides n-grams over copied data and produces owned data or multiple [T;N]. Examples include ftzrs::n_gram, ftzrs::bigram and ftzrs::trigram.

  2. SliceGram provides n-grams over referenced data and produces owned data or multiple &[T]. Examples include ftzrs::n_slice, ftzrs::bislice and ftzrs::trislice.

  3. Combinators that compose one or more featurizers and return a new featurizer with different behavior. Examples include ftzrs::for_each, ftzrs::gap_gram, featurizers! and ftzrs::bookends.

WHY POLYMORPHISM == PERFORMANCE

Here is a small quiz to show why polymorphic featurization and FAST featurization go hand-in-hand.

Here are four different ways to featurize a string that are basically equivalent. But, which one of the four is fastest? By how much?

let sentence = "It is a truth universally acknowledged that Jane Austin must be used in nlp examples";

let one:   Vec<String> = trigram().featurize(sentence);
let two:   Vec<[u8;3]> = trigram().featurize(sentence);
let three: Vec<&str>   = trislice().featurize(sentence); // same performance as &[u8]
let four:  Vec<HashedAs<u64>> = trislice().featurize(sentence); // could have used trigram

Trigrams of String, HashedAs<u64>, &str and [u8; 3] each have their place depending on your use-case. But there can be roughly two orders of magnitude of difference in performance between the fastest and the slowest. If you choose the wrong one for your needs (or use a less polymorphic crate), you're losing out on speed!

What type should I use to represent my features?

  • use Collection<[T; N]> via ftzrs::n_gram if both T and N are small. This is most of the time.

  • use Collection<&[T]> (or Collection<&str>) via ftzrs::n_slice if [T; N] would be larger (in bytes) than (usize, usize). This is more common if N is large or you're using char instead of u8. This is also depends on the lifetime of the original data vs the lifetime of the features produced.

  • HashedAs<u64> has the opposite time complexity as &[T], linear creation and O(1) equality. If you're ok with one-in-a-millionish hash collisions, this can be a great compromise.

  • Never use Collection<String> or Collection<Vec<T>> in a performance critical section.

Where does creature_feature fit in with other tokenizers?

creature_feature is very flexible, and traits::Ftzr/traits::IterFtzr can be easily implemented with a newtype for whatever other tokenizer/featurizer you please. Anything could be featurized: images, documents, time-series data, etc.

Example: Featurizing books

Consider a custom struct to represent a book

struct Book {
   author: String,
   genre: Genre,
   sub_genre: SubGenre,
   year: u16,
}

#[derive(Hash)]
enum Genre {
   Fiction,
   NonFiction,
   Religion,
}

#[derive(Hash)]
enum SubGenre {
   Romance,
   History,
   DataScience,
}

impl Book {
   fn decade(&self) -> u8 {
       unimplemented!()
   }
}

We can easily make a custom featurizer for Book by visitation with traits::Ftzr.

use creature_feature::ftzrs::whole;
use creature_feature::traits::*;
use creature_feature::HashedAs;

struct BookFtzr;

impl<'a> Ftzr<&'a Book> for BookFtzr {
   type TokenGroup = HashedAs<u64>;
   fn push_tokens<Push: FnMut(Self::TokenGroup)>(&self, book: &'a Book, push: &mut Push) {
       whole().push_tokens_from(&book.author, push);
       push(FeatureFrom::from(&book.genre));
       push(FeatureFrom::from(&book.sub_genre));
       push(FeatureFrom::from(book.year));
       push(FeatureFrom::from(book.decade()));
   }
}

Now we could easily implement a similarity metric for Book via Vec<HashedAs<u64>>, like cosine or jaccard.

Usage notes

  • No bounds checking is performed. This is the responsibility of the user.
  • To handle unicode, convert to Vec<char>

YOU CAN HELP

I'm actually an English teacher, not a dev. So any PRs, observations or feedback is very welcome. I've done my best to document everything well, but if you have any questions feel free to reach out. Your help with any of the small number of current issues would be VERY much welcome :)

Dependencies

~110–425KB