E6DB

This a dataset is compiled from the e621 database. It currently provides indexes for normalizing tags and embeddings for finding similar tags.

Utilities

query_tags.py

A small command-line utility that finds related tags and generates 2D plots illustrating tag relationships through local PCA projection. It utilizes collaborative filtering embeddings, computed with alternating least squares.

When only tags are provided as arguments, it displays the top-k most similar tags (where k is set with -k). By using -o plot.png or -o -, it saves or displays a 2D plot showing the local projection of the query and related tags.

Tag categories are represented with the e621 color scheme. Results can be filtered based on one or more categories using the -c flag once or multiple times. The -f flag sets a post count threshold.

Filtering occurs in the following sequence:

• Tags used fewer than twice are excluded from the dataset; tags with a post count lower than the -f threshold are also discarded.
• If a category filter is specified, only matching tags are retained.
• For each query tag, the most -k similar neighboring tags are selected.
• The per-query neighbors are printed, and if no plot is being generated, filtering halts at this point.
• Similarity scores are aggregated across queries, and the -n tags closest to all queries are chosen for the PCA.
• Only the highest -N scoring tags are displayed in the plot.

e6db.utils

This module contains utilities for loading the provided data and use it to normalize tag sets.

• load_tags and load_implications loads the tags indexes and implications,
• TagNormalizer allows to adapt the spellings of tag and alias for working with various datasets. It can normalize spellings by converting tag strings to numerical ids and back,
• TagSetNormalizer uses the above class along tag implications to normalize tag sets and strip implied tags.

See this example notebook that cleans T2I datasets in the sd-script format.

Additionally, the e6db.utils.numpy and e6db.utils.torch modules provides functions to construct post-tag interaction matrices in sparse matrix format. For this you'll need to generate the posts.parquet file from CSVs.

importdb

Reads the CSVs from e621 db export

python -m e6db.importdb ./data reads tags, aliases, implications and posts CSV files. The following operations are performed:

• Tags used at least twice are assigned numerical ids based on their rank.
• Computes the transitive closure of implications,
• For each post, split the tags into direct and implied tag.
• Write parquets files for tags and posts (~500MB) and convert tags indexes using simple formats described in the next section.

The CSV files must be decompressed beforehand.

Dataset content

This dataset currently focus on tags alone using simple file formats that are easily parsed without additional dependencies.

• tag2idx.json.gz: a dictionary mapping tag strings and aliases to numerical id (tag rank),
• tags.txt.gz: list of tags sorted by rank, can be indexed by the ids given by tag2idx.json.gz,
• tags_categories.bin.gz: a raw array of bytes representing tag categories in the same order than tags.txt.gz,
• implications.json.gz: maps tags id to implied tag ids (including transitive implications),
• implications_rej.json.gz: maps tag strings to a list of implied numerical ids. Keys in implications_rej are tags that have a very little usage (less than 2 posts) and don't have numerical ids associated with them.
• implicit_tag_factors.safetensors: Tag embedding computed by alternating least squares.

No post data is currently included, since this wouldn't add any useful information compared to what's inside the CSVs. If you want the post parquet files with normalized tags, you can download the CSVs and run the e6db.importdb script yourself.

I plan to compile more post data in the future, such as aesthetic predictions, adjusted favcounts, etc. Utilities will then be added to assists with the selection of a subset of posts for specific ML tasks.