Implementation:Datajuicer Data juicer ImageTaggingMapper

Knowledge Sources	Datajuicer_Data_juicer
Domains	Data_Processing, Mapping
Last Updated	2026-02-14 16:00 GMT

Overview

Concrete tool for generating descriptive tags for images using the RAM Plus model provided by Data-Juicer.

Description

ImageTaggingMapper is a mapper operator that generates image tags using the RAM (Recognize Anything Model) Plus model with a Swin Large backbone (ram_plus_swin_large_14m). It processes images at 384x384 resolution, runs the model's generate_tag method to produce pipe-delimited tag strings, splits them into individual words, sorts by frequency using a Counter, and stores the resulting tag array in sample metadata under the configured field name (default: image_tags). Skips reprocessing if tags already exist. Requires CUDA acceleration and approximately 9GB of GPU memory.

Usage

Use when you need automated semantic annotation of images without manual labeling, enabling content-based filtering, search, and organization of image datasets.

Code Reference

Source Location

• Repository: Datajuicer_Data_juicer
• File: data_juicer/ops/mapper/image_tagging_mapper.py

Signature

@OPERATORS.register_module("image_tagging_mapper")
class ImageTaggingMapper(Mapper):
    def __init__(self,
                 tag_field_name: str = MetaKeys.image_tags,
                 *args, **kwargs):

Import

from data_juicer.ops.mapper.image_tagging_mapper import ImageTaggingMapper

I/O Contract

Inputs

Name	Type	Required	Description
tag_field_name	str	No	Field name to store the tags in metadata, defaults to MetaKeys.image_tags

Outputs

Name	Type	Description
samples	Dict	Transformed samples with frequency-sorted image tags stored in meta field

Usage Examples

process:
  - image_tagging_mapper:
      tag_field_name: "image_tags"

Related Pages

• Environment:Datajuicer_Data_juicer_Python_Runtime_Environment