Fixing Trainer with u8 instead of chars. (#452)

* Fixing Trainer with u8 instead of chars.

Now check both optimized and unoptimized encodings schemes for Unigram.

* Small fixes.

* Fixing makefile.

bindings/python/tests/bindings/test_trainers.py

@@ -7,10 +7,9 @@ from ..utils import data_dir, train_files
 
 
 class TestUnigram:
-    @pytest.mark.slow
     def test_train(self, train_files):
         tokenizer = SentencePieceUnigramTokenizer()
-        tokenizer.train(train_files["big"], show_progress=False)
+        tokenizer.train(train_files["small"], show_progress=False)
 
         filename = "tests/data/unigram_trained.json"
         tokenizer.save(filename)

bindings/python/tests/utils.py

@@ -61,8 +61,17 @@ def openai_files(data_dir):
 
 @pytest.fixture(scope="session")
 def train_files(data_dir):
+    big = download("https://norvig.com/big.txt")
+    small = os.path.join(DATA_PATH, "small.txt")
+    with open(small, "w") as f:
+        with open(big, "r") as g:
+            for i, line in enumerate(g):
+                f.write(line)
+                if i > 100:
+                    break
     return {
-        "big": download("https://norvig.com/big.txt"),
+        "small": small,
+        "big": big,
     }
 
 

tokenizers/Makefile

@@ -4,8 +4,8 @@ TESTS_DIR = tests
 
 dir_guard=@mkdir -p $(@D)
 
-SHARED_RESOURCES = $(DATA_DIR)/gpt2-vocab.json $(DATA_DIR)/gpt2-merges.txt $(DATA_DIR)/bert-base-uncased-vocab.txt
+SHARED_RESOURCES = $(DATA_DIR)/gpt2-vocab.json $(DATA_DIR)/gpt2-merges.txt $(DATA_DIR)/bert-base-uncased-vocab.txt $(DATA_DIR)/big.txt $(DATA_DIR)/small.txt
-BENCHMARK_RESOURCES = $(SHARED_RESOURCES) $(DATA_DIR)/big.txt $(DATA_DIR)/small.txt
+BENCHMARK_RESOURCES = $(SHARED_RESOURCES)
 TESTS_RESOURCES = $(SHARED_RESOURCES) $(DATA_DIR)/unigram.json $(DATA_DIR)/unigram_wagahaiwa_nekodearu.txt
 
 .PHONY : build

tokenizers/src/models/unigram/lattice.rs

@@ -137,7 +137,7 @@ fn log_sum_exp(x: f64, y: f64, init_mode: bool) -> f64 {
 
 impl<'a> Lattice<'a> {
     pub fn from(sentence: &'a str, unk_id: usize, bos_id: usize, eos_id: usize) -> Lattice<'a> {
-        let len = sentence.chars().count();
+        let len = sentence.bytes().count();
         let k_reserved_node_size = 16;
         // We are adding 2 tokens, bos and eos
         let mut nodes: Vec<NodeRef> = Vec::with_capacity(k_reserved_node_size);
@@ -177,7 +177,8 @@ impl<'a> Lattice<'a> {
 
     pub fn viterbi(&mut self) -> Vec<NodeRef> {
         let len = self.len;
-        for pos in 0..=len {
+        let mut pos = 0;
+        while pos <= len {
             if self.begin_nodes[pos].is_empty() {
                 return vec![];
             }
@@ -201,6 +202,11 @@ impl<'a> Lattice<'a> {
                     None => return vec![],
                 }
             }
+            if let Some(c) = self.sentence[pos..].chars().next() {
+                pos += c.len_utf8();
+            } else {
+                break;
+            }
         }
 
         let mut results: Vec<NodeRef> = vec![];
@@ -220,11 +226,7 @@ impl<'a> Lattice<'a> {
     }
 
     pub fn piece(&self, node: &Node) -> String {
-        self.sentence
+        self.sentence[node.pos..node.pos + node.length].to_owned()
-            .chars()
-            .skip(node.pos)
-            .take(node.length)
-            .collect()
     }
 
     pub fn tokens(&mut self) -> Vec<String> {
@@ -306,14 +308,11 @@ impl<'a> Lattice<'a> {
     pub fn len(&self) -> usize {
         self.len
     }
+
     pub fn is_empty(&self) -> bool {
         self.len == 0
     }
 
-    pub fn utf8_len(&self) -> usize {
-        self.sentence.len()
-    }
-
     pub fn bos_node(&self) -> NodeRef {
         Rc::clone(&self.end_nodes[0][0])
     }
@@ -443,18 +442,14 @@ mod tests {
         let lattice = Lattice::from("", 0, 1, 2);
 
         assert_eq!(lattice.len(), 0);
-        assert_eq!(lattice.utf8_len(), 0);
-        // EXPECT_EQ(0, lattice.utf8_size());
 
         let lattice = Lattice::from("", 0, 1, 2);
         assert_eq!(lattice.len(), 0);
-        assert_eq!(lattice.utf8_len(), 0);
         assert_eq!(lattice.sentence(), "");
         assert_eq!(lattice.surface(0), "");
 
         let lattice = Lattice::from("test", 0, 1, 2);
         assert_eq!(lattice.len(), 4);
-        assert_eq!(lattice.utf8_len(), 4);
         assert_eq!(lattice.sentence(), "test");
         assert_eq!(lattice.surface(0), "test");
         assert_eq!(lattice.surface(1), "est");
@@ -476,8 +471,7 @@ mod tests {
         );
 
         let lattice = Lattice::from("テストab", 0, 1, 2);
-        assert_eq!(lattice.len(), 5);
+        assert_eq!(lattice.len(), 11);
-        assert_eq!(lattice.utf8_len(), 11);
         assert_eq!(lattice.sentence(), "テストab");
         assert_eq!(lattice.surface(0), "テストab");
         assert_eq!(lattice.surface(1), "ストab");
@@ -492,11 +486,11 @@ mod tests {
 
         lattice.insert(0, 1, 0.0, 3);
         lattice.insert(1, 1, 0.0, 4);
-        lattice.insert(2, 1, 0.0, 5);
+        lattice.insert(2, 3, 0.0, 5);
-        lattice.insert(3, 1, 0.0, 6);
+        lattice.insert(5, 3, 0.0, 6);
         lattice.insert(0, 2, 0.0, 7);
-        lattice.insert(1, 2, 0.0, 8);
+        lattice.insert(1, 4, 0.0, 8);
-        lattice.insert(2, 2, 0.0, 9);
+        lattice.insert(2, 6, 0.0, 9);
         // 0 & 1 are bos and eos
         let node0 = lattice.nodes[2].borrow();
         let node1 = lattice.nodes[3].borrow();
@@ -517,18 +511,18 @@ mod tests {
         assert_eq!(node0.pos, 0);
         assert_eq!(node1.pos, 1);
         assert_eq!(node2.pos, 2);
-        assert_eq!(node3.pos, 3);
+        assert_eq!(node3.pos, 5);
         assert_eq!(node4.pos, 0);
         assert_eq!(node5.pos, 1);
         assert_eq!(node6.pos, 2);
 
         assert_eq!(node0.length, 1);
         assert_eq!(node1.length, 1);
-        assert_eq!(node2.length, 1);
+        assert_eq!(node2.length, 3);
-        assert_eq!(node3.length, 1);
+        assert_eq!(node3.length, 3);
         assert_eq!(node4.length, 2);
-        assert_eq!(node5.length, 2);
+        assert_eq!(node5.length, 4);
-        assert_eq!(node6.length, 2);
+        assert_eq!(node6.length, 6);
 
         assert_eq!(lattice.bos_node().borrow().id, 1);
         assert_eq!(lattice.eos_node().borrow().id, 2);
@@ -543,14 +537,14 @@ mod tests {
         assert_eq!(lattice.begin_nodes[0].len(), 2);
         assert_eq!(lattice.begin_nodes[1].len(), 2);
         assert_eq!(lattice.begin_nodes[2].len(), 2);
-        assert_eq!(lattice.begin_nodes[3].len(), 1);
+        assert_eq!(lattice.begin_nodes[5].len(), 1);
-        assert_eq!(lattice.begin_nodes[4].len(), 1);
+        assert_eq!(lattice.begin_nodes[8].len(), 1);
 
         assert_eq!(lattice.end_nodes[0].len(), 1);
         assert_eq!(lattice.end_nodes[1].len(), 1);
         assert_eq!(lattice.end_nodes[2].len(), 2);
-        assert_eq!(lattice.end_nodes[3].len(), 2);
+        assert_eq!(lattice.end_nodes[5].len(), 2);
-        assert_eq!(lattice.end_nodes[4].len(), 2);
+        assert_eq!(lattice.end_nodes[8].len(), 2);
 
         assert_eq!(lattice.begin_nodes[0][0].borrow().id, node0.id);
         assert_eq!(lattice.begin_nodes[0][1].borrow().id, node4.id);
@@ -558,10 +552,10 @@ mod tests {
         assert_eq!(lattice.begin_nodes[1][1].borrow().id, node5.id);
         assert_eq!(lattice.begin_nodes[2][0].borrow().id, node2.id);
         assert_eq!(lattice.begin_nodes[2][1].borrow().id, node6.id);
-        assert_eq!(lattice.begin_nodes[3][0].borrow().id, node3.id);
+        assert_eq!(lattice.begin_nodes[5][0].borrow().id, node3.id);
         assert_eq!(
             lattice.eos_node().borrow().id,
-            lattice.begin_nodes[4][0].borrow().id
+            lattice.begin_nodes[8][0].borrow().id
         );
 
         assert_eq!(
@@ -571,10 +565,10 @@ mod tests {
         assert_eq!(node0.id, lattice.end_nodes[1][0].borrow().id);
         assert_eq!(node1.id, lattice.end_nodes[2][0].borrow().id);
         assert_eq!(node4.id, lattice.end_nodes[2][1].borrow().id);
-        assert_eq!(node2.id, lattice.end_nodes[3][0].borrow().id);
+        assert_eq!(node2.id, lattice.end_nodes[5][0].borrow().id);
-        assert_eq!(node5.id, lattice.end_nodes[3][1].borrow().id);
+        assert_eq!(node5.id, lattice.end_nodes[5][1].borrow().id);
-        assert_eq!(node3.id, lattice.end_nodes[4][0].borrow().id);
+        assert_eq!(node3.id, lattice.end_nodes[8][0].borrow().id);
-        assert_eq!(node6.id, lattice.end_nodes[4][1].borrow().id);
+        assert_eq!(node6.id, lattice.end_nodes[8][1].borrow().id);
     }
 
     #[test]

tokenizers/src/models/unigram/model.rs

@@ -24,6 +24,7 @@ pub struct Unigram {
     pub(super) eos_id: usize,
 
     fuse_unk: bool,
+    is_optimized: bool,
 }
 impl PartialEq for Unigram {
     fn eq(&self, other: &Self) -> bool {
@@ -46,6 +47,7 @@ impl Clone for Unigram {
             bos_id: self.bos_id,
             eos_id: self.eos_id,
             fuse_unk: self.fuse_unk,
+            is_optimized: self.is_optimized,
         }
     }
 }
@@ -122,6 +124,7 @@ impl Unigram {
         }
         let trie = builder.build();
         let fuse_unk = true;
+        let is_optimized = true;
 
         Ok(Unigram {
             vocab,
@@ -133,6 +136,7 @@ impl Unigram {
             unk_id,
             fuse_unk,
             cache: Cache::default(),
+            is_optimized,
         })
     }
 
@@ -142,6 +146,11 @@ impl Unigram {
         self.cache = self.cache.fresh();
     }
 
+    #[cfg(test)]
+    pub(super) fn set_optimized(&mut self, is_optimized: bool) {
+        self.is_optimized = is_optimized;
+    }
+
     pub(super) fn len(&self) -> usize {
         self.vocab.len()
     }
@@ -179,7 +188,7 @@ impl Unigram {
             }
 
             if !has_single_node {
-                lattice.insert(begin_pos, 1, unk_score, self.unk_id);
+                lattice.insert(begin_pos, mblen, unk_score, self.unk_id);
             }
             begin_pos += mblen
         }
@@ -212,7 +221,11 @@ impl Unigram {
         if let Some(result) = self.cache.get(sentence) {
             result.to_vec()
         } else {
-            let result = self.encode_optimized(sentence);
+            let result = if self.is_optimized {
+                self.encode_optimized(sentence)
+            } else {
+                self.encode_unoptimized(sentence)
+            };
             self.cache.set(sentence.to_owned(), result.clone());
             result
         }
@@ -327,7 +340,6 @@ impl Unigram {
         results
     }
 
-    #[allow(dead_code)]
     fn encode_unoptimized(&self, sentence: &str) -> Vec<String> {
         let mut lattice = Lattice::from(sentence, self.unk_id, self.bos_id, self.eos_id);
         self.populate_nodes(&mut lattice);
@@ -541,27 +553,34 @@ mod tests {
         ];
 
         let mut model = Unigram::from(sentencepieces, 0).unwrap();
-        assert_eq!(model.encode("abc"), vec!["abc"]);
-        assert_eq!(model.encode("AB"), vec!["AB"]);
 
-        model.set_fuse_unk(false);
+        for is_optimized in &[true, false] {
-        assert_eq!(model.encode("AB"), vec!["A", "B"]);
+            model.set_optimized(*is_optimized);
-        model.set_fuse_unk(true);
+            println!("IsOptimized {:?}", is_optimized);
+            assert_eq!(model.encode("abc"), vec!["abc"]);
+            assert_eq!(model.encode("AB"), vec!["AB"]);
 
-        assert_eq!(model.encode("abcd"), vec!["ab", "cd"]);
+            model.set_fuse_unk(false);
-        assert_eq!(model.encode("abcc"), vec!["abc", "c"]);
+            assert_eq!(model.encode("AB"), vec!["A", "B"]);
-        assert_eq!(
+            model.set_fuse_unk(true);
-            model.encode("xabcabaabcdd"),
+            assert_eq!(model.encode("AB"), vec!["AB"]);
-            vec!["x", "abc", "ab", "a", "ab", "cd", "d"]
-        );
-        model.set_fuse_unk(false);
-        assert_eq!(model.encode("xyz東京"), vec!["x", "y", "z", "東", "京"]);
-        model.set_fuse_unk(true);
 
-        // User encoded in original version
+            assert_eq!(model.encode("abcd"), vec!["ab", "cd"]);
-        assert_eq!(model.encode("ABC"), vec!["ABC"]);
+            assert_eq!(model.encode("abcc"), vec!["abc", "c"]);
-        assert_eq!(model.encode("abABCcd"), vec!["ab", "ABC", "cd"]);
+            assert_eq!(
-        assert_eq!(model.encode("ababcdabcdcd"), vec!["ab", "abcdabcd", "cd"]);
+                model.encode("xabcabaabcdd"),
-        assert_eq!(model.encode("abqrcd"), vec!["ab", "q", "r", "cd"]);
+                vec!["x", "abc", "ab", "a", "ab", "cd", "d"]
+            );
+            model.set_fuse_unk(false);
+            assert_eq!(model.encode("xyz東京"), vec!["x", "y", "z", "東", "京"]);
+            model.set_fuse_unk(true);
+            assert_eq!(model.encode("xyz東京"), vec!["xyz東京"]);
+
+            // User encoded in original version
+            assert_eq!(model.encode("ABC"), vec!["ABC"]);
+            assert_eq!(model.encode("abABCcd"), vec!["ab", "ABC", "cd"]);
+            assert_eq!(model.encode("ababcdabcdcd"), vec!["ab", "abcdabcd", "cd"]);
+            assert_eq!(model.encode("abqrcd"), vec!["ab", "q", "r", "cd"]);
+        }
     }
 }

tokenizers/tests/unigram.rs

@@ -1,14 +1,13 @@
 #[cfg(not(debug_assertions))]
 use assert_approx_eq::assert_approx_eq;
-#[cfg(not(debug_assertions))]
 use std::collections::HashMap;
-#[cfg(not(debug_assertions))]
 use std::fs::read_to_string;
 use std::path::Path;
 #[cfg(not(debug_assertions))]
 use tokenizers::models::unigram::Lattice;
 use tokenizers::models::unigram::Unigram;
-use tokenizers::tokenizer::Model;
+use tokenizers::models::unigram::UnigramTrainer;
+use tokenizers::tokenizer::{Model, Trainer};
 
 #[test]
 fn test_unigram_from_file() {
@@ -39,6 +38,26 @@ fn test_unigram_from_file() {
     );
 }
 
+#[test]
+fn test_train_unigram_from_file() {
+    let content = read_to_string("data/small.txt").unwrap();
+    let mut word_counts = HashMap::new();
+    content.split_whitespace().for_each(|word| {
+        // This is important for the test of char vs u8
+        let word = format!("▁{}", word.to_string());
+        *word_counts.entry(word).or_insert(0) += 1;
+    });
+
+    // println!("Words counts {:?}", word_counts);
+
+    let trainer = UnigramTrainer::builder()
+        .show_progress(false)
+        .build()
+        .unwrap();
+    let (model, _) = trainer.train(word_counts).unwrap();
+    assert_eq!(model.get_vocab_size(), 719);
+}
+
 #[cfg(not(debug_assertions))]
 #[test]
 fn test_sample() {