nemo_curator.stages.interleaved.filter.clip_score_filter

Module Contents

Classes

Name	Description
`InterleavedCLIPScoreFilterStage`	Filter interleaved image rows by CLIP image-text relevance score.

Functions

Name	Description
`_indices_and_decoded_images_from_rows`	Decode image bytes per row; clear keep_mask entries where decode fails.
`_sample_texts_list_from_df`	Return list of text_content from all text rows for the given sample_id (non-empty).

Data

DEFAULT_CLIP_MIN_SCORE

API

class nemo_curator.stages.interleaved.filter.clip_score_filter.InterleavedCLIPScoreFilterStage(
    name: str = 'interleaved_clip_score_fil...,
    drop_invalid_rows: bool = True,
    model_dir: str | None = None,
    min_score: float = DEFAULT_CLIP_MIN_SCORE,
    resources: nemo_curator.stages.resources.Resources = (lambda: Resources(gpu_memo...
)

Dataclass

Bases: BaseInterleavedFilterStage

Filter interleaved image rows by CLIP image-text relevance score.

For each image row, all text rows with the same sample_id form (image, text) pairs. CLIP similarity is computed for each pair. An image is kept only if at least one pair has score >= min_score; otherwise it is dropped.

min_score

float = DEFAULT_CLIP_MIN_SCORE

model_dir

str | None = None

name

str = 'interleaved_clip_score_filter'

resources

Resources

nemo_curator.stages.interleaved.filter.clip_score_filter.InterleavedCLIPScoreFilterStage.content_keep_mask(
    task: nemo_curator.tasks.InterleavedBatch,
    df: pandas.DataFrame
) -> pandas.Series

nemo_curator.stages.interleaved.filter.clip_score_filter.InterleavedCLIPScoreFilterStage.setup(
    worker_metadata: nemo_curator.backends.base.WorkerMetadata | None = None
) -> None

nemo_curator.stages.interleaved.filter.clip_score_filter.InterleavedCLIPScoreFilterStage.setup_on_node(
    node_info: nemo_curator.backends.base.NodeInfo,
    worker_metadata: nemo_curator.backends.base.WorkerMetadata
) -> None

Download the weights for the CLIP model on the node.

nemo_curator.stages.interleaved.filter.clip_score_filter._indices_and_decoded_images_from_rows(
    rows: list[tuple[int, bytes]],
    keep_mask: pandas.Series
) -> tuple[list[int], list[numpy.ndarray]]

Decode image bytes per row; clear keep_mask entries where decode fails.

nemo_curator.stages.interleaved.filter.clip_score_filter._sample_texts_list_from_df(
    df: pandas.DataFrame,
    sample_id: str
) -> list[str]

Return list of text_content from all text rows for the given sample_id (non-empty).

nemo_curator.stages.interleaved.filter.clip_score_filter.DEFAULT_CLIP_MIN_SCORE: float = 0.15