Skip to content

Filtering empty text examples #1222

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

Merged
artemru merged 6 commits into from
Jul 3, 2025
Merged

Filtering empty text examples #1222

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
69cb058
filter empty text examples asr dataloading
Jul 2, 2025
8fe00ad
comment
Jul 2, 2025
1e69b88
flake8
Jul 2, 2025
074593f
early exit
Jul 2, 2025
e7ff7dc
isort
Jul 2, 2025
6810288
rm UNK with option
Jul 3, 2025
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
29 changes: 25 additions & 4 deletions src/fairseq2/datasets/asr.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -160,6 +160,28 @@ def create_reader(
self._name, split, builder.and_return(), gang, options
)

@staticmethod
def add_tokenization_pipeline(
builder: DataPipelineBuilder,
tokenizer: TextTokenizer,

@mattsetz mattsetz Jul 2, 2025

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

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

we can add another arg here remove_unk to optionally remove all UNKs tokens from transcript. we should default this to False but it can be something that we experiment with.

@mattsetz mattsetz Jul 2, 2025

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

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

also, maybe we can log the number of UNKs (or % of num_unks/total_tokens within a sample/batch) in weights and biases? its useful for us to know if our dataset contains too many UNKs.

@artemru artemru Jul 3, 2025

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

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

also, maybe we can log the number of UNKs (or % of num_unks/total_tokens within a sample/batch) in weights and biases? its useful for us to know if our dataset contains too many UNKs.

it's possible (requires a bit to propagate some metric bags to data reader) but for now we can just do in a adhoc analysis

@artemru artemru Jul 3, 2025

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

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

we can add another arg here remove_unk to optionally remove all UNKs tokens from transcript. we should default this to False but it can be something that we experiment with.

done !

) -> DataPipelineBuilder:
# Tokenize target text.
text_encoder = tokenizer.create_encoder()

# to avoid to tokenize empty text, we filter out them out first
builder = builder.filter(lambda x: bool(len(x["text"]) > 0))

builder.map(text_encoder, selector="text")

unk_idx = tokenizer.vocab_info.unk_idx

def empty_text(example: Dict[str, Any]) -> bool:
return bool((example["text"] != unk_idx).sum().item() > 0)
Comment thread
artemru marked this conversation as resolved.

builder = builder.filter(empty_text)

return builder

@staticmethod
def build_asr_main_pipeline(
builder: DataPipelineBuilder,
Expand All @@ -170,6 +192,9 @@ def build_asr_main_pipeline(
max_audio_len: int,
) -> DataPipelineBuilder:

# Tokenize target text.
builder = GenericAsrDataset.add_tokenization_pipeline(builder, tokenizer)

# Bucketize examples by audio length.
builder = GenericSpeechDataset.add_bucketing_pipeline(
builder,
Expand All @@ -188,10 +213,6 @@ def build_asr_main_pipeline(
builder, options, GenericSpeechDataset.rename_feature
)

# Tokenize target text.
text_encoder = tokenizer.create_encoder()
builder.map(text_encoder, selector="[*].text", num_parallel_calls=options.npc)

# Collate bucketed examples into a batch.
text_collate_opts = CollateOptionsOverride(
"text", pad_value=tokenizer.vocab_info.pad_idx
Expand Down
6 changes: 2 additions & 4 deletions src/fairseq2/datasets/asr_parquet.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -125,6 +125,8 @@ def build_asr_main_pipeline(
builder = builder.shuffle(options.example_shuffle_window, seed=options.seed)
options.seed += 1

builder = GenericAsrDataset.add_tokenization_pipeline(builder, tokenizer)

builder = GenericSpeechDataset.add_bucketing_pipeline(
builder,
options,
Expand All @@ -150,10 +152,6 @@ def build_parquet_audio_text_reading(
builder, options, GenericSpeechDataset.rename_feature
)

# Tokenize target text.
text_encoder = tokenizer.create_encoder()
builder.map(text_encoder, selector="[*].text", num_parallel_calls=options.npc)

# Collate bucketed examples into a batch.
text_collate_opts = CollateOptionsOverride(
"text", pad_value=tokenizer.vocab_info.pad_idx
Expand Down
Loading