Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Fixing the vocab size of the trained Unigram model #952

Merged
merged 6 commits into from
Mar 18, 2022
Merged

Fixing the vocab size of the trained Unigram model #952

Show file tree
Hide file tree
Changes from 5 commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
c75eafc
Fixing the vocab size of the trained Unigram model
kaisugi Mar 16, 2022
fb8955c
add test for the vocab size of the trained Unigram model
kaisugi Mar 17, 2022
1015b6c
Revert "add test for the vocab size of the trained Unigram model"
kaisugi Mar 17, 2022
b28377f
Fixing the vocab size of the trained Unigram model
kaisugi Mar 17, 2022
7004239
format codes
kaisugi Mar 18, 2022
f06220a
get the position of vocab-size calculation out of loop
kaisugi Mar 18, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
22 changes: 22 additions & 0 deletions bindings/python/tests/bindings/test_trainers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -238,6 +238,28 @@ def test_train_with_special_tokens(self):
"[SEP]",
]

tokenizer = Tokenizer(models.Unigram())
trainer = trainers.UnigramTrainer(
show_progress=False,
special_tokens=["[PAD]", "[SEP]", "[CLS]"],
unk_token="[UNK]",
vocab_size=100,
)
tokenizer.train([filename], trainer=trainer)

assert tokenizer.get_vocab_size() == 100

tokenizer = Tokenizer(models.Unigram())
trainer = trainers.UnigramTrainer(
show_progress=False,
special_tokens=["[PAD]", "[SEP]", "[CLS]", "[UNK]"],
unk_token="[UNK]",
vocab_size=100,
)
tokenizer.train([filename], trainer=trainer)

assert tokenizer.get_vocab_size() == 100

def test_cannot_train_different_model(self):
tokenizer = Tokenizer(models.BPE())
trainer = trainers.UnigramTrainer(show_progress=False)
Expand Down
32 changes: 20 additions & 12 deletions tokenizers/src/models/unigram/trainer.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -126,19 +126,7 @@ impl UnigramTrainer {
min_score_penalty += min_score_penalty_delta;
}
}
for (token, score) in model.iter() {
if inserted.contains::<str>(token) {
continue;
}
inserted.insert(token.to_string());
pieces.push((token.to_string(), if score.is_nan() { 0.0 } else { *score }));
if pieces.len() == self.vocab_size as usize {
break;
}
}
pieces.sort_by(|(_, a), (_, b)| b.partial_cmp(a).unwrap());

// Insert the necessary tokens
let (unk_id, need_add_unk) = if let Some(ref unk) = self.unk_token {
let unk_id = self.special_tokens.iter().enumerate().find_map(|(i, t)| {
if t.content == *unk {
Expand All @@ -154,6 +142,26 @@ impl UnigramTrainer {
} else {
(None, false)
};

for (token, score) in model.iter() {
if inserted.contains::<str>(token) {
continue;
}
inserted.insert(token.to_string());
pieces.push((token.to_string(), if score.is_nan() { 0.0 } else { *score }));

let vocab_size_without_special_tokens = if need_add_unk {
self.vocab_size as usize - self.special_tokens.len() - 1
} else {
self.vocab_size as usize - self.special_tokens.len()
};
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Sorry I should have seen that earlier, do you mind extracting this from the loop (it can be defined on top.
This will likely be optimized away, but it still help readability too.

I can take care of it if you want.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I should have found it out...

if pieces.len() == vocab_size_without_special_tokens {
break;
}
}
pieces.sort_by(|(_, a), (_, b)| b.partial_cmp(a).unwrap());

// Insert the necessary tokens
let mut special_tokens = self
.special_tokens
.iter()
Expand Down