Description #

Video safety detector – multi-source input, unified frame sampling, and batch inference

Key features #

• Based on a pre-trained image classification model, performs NSFW probability detection on frames sampled from the video and outputs the final score according to the aggregation strategy.
• Supports multiple video input sources:
• URL path (video_url)
• Base64 encoding (video_base64)
• Binary stream (video_binary)
• Supports multiple sampling methods (via VideoFrameSampler):
• by_count_uniform / by_interval_time / by_interval_frames / by_fps / by_timestamps
• Batch inference:
• The inference batch size can be controlled via batch_size to improve throughput performance.

Application scenarios #

• Content safety review: Assess NSFW risk for user-uploaded videos.
• Pre-filtering in production pipelines: Screen video safety before subsequent processing.

Caution #

• Only outputs numerical NSFW confidence (0~1); frame images are not saved.

Operator parameters #

Input #

Output #

A structured result array, where each element is the NSFW confidence score for a video; returns None if sampling/inference fails or no valid frames are available

Parameters #

If a parameter does not have a default value, it is required

Examples #

The following code demonstrates how to use Daft (for distributed scenarios) to run the operator for video safety detection.

from __future__ import annotations

import os

import daft
from daft import col
from daft.las.functions.udf import las_udf
from daft.las.functions.video.video_nsfw_detect import VideoNsfwDetect

if __name__ == "__main__":
    if os.getenv("DAFT_RUNNER", "native") == "ray":
        import logging

        import ray

        def configure_logging():
            logging.basicConfig(
                level=logging.INFO,
                format="%(asctime)s - %(name)s - %(levelname)s - %(message)s",
                datefmt="%Y-%m-%d %H:%M:%S.%s".format(),
            )
            logging.getLogger("tracing.span").setLevel(logging.WARNING)
            logging.getLogger("daft_io.stats").setLevel(logging.WARNING)
            logging.getLogger("DaftStatisticsManager").setLevel(logging.WARNING)
            logging.getLogger("DaftFlotillaScheduler").setLevel(logging.WARNING)
            logging.getLogger("DaftFlotillaDispatcher").setLevel(logging.WARNING)

        ray.init(dashboard_host="0.0.0.0", runtime_env={"worker_process_setup_hook": configure_logging})
        daft.set_runner_ray()

    daft.set_execution_config(actor_udf_ready_timeout=600)
    daft.set_execution_config(min_cpu_per_task=0)

    tos_dir_url = os.getenv("TOS_DIR_URL", "las-cn-beijing-public-online.tos-cn-beijing.volces.com")
    samples = {
        "video_path": [
            f"https://{tos_dir_url}/public/shared_video_dataset/eating_56.mp4"
        ]
    }

    video_src_type = "video_url"
    model_path = os.getenv("MODEL_PATH", "/opt/las/models")
    model_name = "Falconsai/nsfw_image_detection"
    num_gpus = 0

    ds = daft.from_pydict(samples)
    ds = ds.with_column(
        "nsfw",
        las_udf(
            VideoNsfwDetect,
            construct_args={
                "model_path": model_path,
                "video_src_type": video_src_type,
                "sample_mode": "by_count_uniform",
                "count_k": 3,
            },
            num_gpus=num_gpus,
            batch_size=1,
            concurrency=1,
        )(col("video_path")),
    )

    ds.show()

    # ╭────────────────────────────────┬──────────╮
    # │ video_path                     ┆ nsfw     │
    # │ ---                            ┆ ---      │
    # │ Utf8                           ┆ Float64  │
    # ╞════════════════════════════════╪══════════╡
    # │ https://las-cn-beijing-public… ┆ 0.000684 │
    # ╰────────────────────────────────┴──────────╯
、、、