DeepDanbooru

Introduction to DeepDanbooru

DeepDanbooru is an innovative system designed to estimate tags for anime-style girl images. Developed using Python 3.11, this tool allows users to categorize and analyze images efficiently. For those interested in seeing it in action, a live demo is available at the DeepDanbooru Web.

Requirements

DeepDanbooru relies on a few Python packages to function effectively. To get started, the following dependencies must be installed:

• Click (version 8.1.7 or newer)
• numpy (version 1.26.4 or newer)
• requests (version 2.32.3 or newer)
• scikit-image (version 0.24.0 or newer)
• six (version 1.16.0 or newer)
• tensorflow (version 2.17.0 or newer)
• tensorflow-io (version 0.31.0 or newer)

These can be quickly installed using the requirements.txt file by running the command:

> pip install -r requirements.txt

Alternatively, DeepDanbooru can be installed with pip. Note that TensorFlow is not included by default. To include it, use the following commands:

> # default installation
> pip install .
> # with tensorflow package
> pip install .[tensorflow]

Usage

DeepDanbooru requires some preparatory steps before it can be used for training and evaluation:

1. Prepare a Dataset: If you don't have an existing dataset, use DanbooruDownloader to download one from Danbooru.

2. Create a Training Project Folder:

   > deepdanbooru create-project [your_project_folder]
   

3. Prepare a Tag List: To obtain the latest tags from the Danbooru server, the following command can be used (account and API key required):

   > deepdanbooru download-tags [your_project_folder] --username [your_danbooru_account] --api-key [your_danbooru_api_key]
   

4. Optional Dataset Filtering: Convert optional tags (such as rating and score) into system tags if desired:

   > deepdanbooru make-training-database [your_dataset_sqlite_path] [your_filtered_sqlite_path]
   

5. Modify the project.json File: Update the database_path setting to point to the actual SQLite file path.

6. Start Training:

   > deepdanbooru train-project [your_project_folder]
   

7. Evaluate the System:

   > deepdanbooru evaluate [image_file_path or folder]... --project-path [your_project_folder] --allow-folder
   

Dataset Structure

The input dataset for DeepDanbooru follows a specific folder structure. The core component is an SQLite database file, which should include a particular table structure as detailed below:

• Main Folder: Contains an images subfolder and the SQLite file (*.sqlite).
• Images Folder: Images are stored in subfolders named after the first two characters of their filename.

Example structure:

MyDataset/
├── images/
│   ├── 00/
│   │   ├── 00000000000000000000000000000000.jpg
│   │   ├── ...
│   ├── 01/
│   │   ├── 01000000000000000000000000000000.jpg
│   │   ├── ...
│   └── ff/
│       ├── ff000000000000000000000000000000.jpg
│       ├── ...
└── my-dataset.sqlite

The database file should have the following table structure:

posts
├── id (INTEGER)
├── md5 (TEXT)
├── file_ext (TEXT)
├── tag_string (TEXT)
└── tag_count_general (INTEGER)

Each image filename must match the format [md5].[file_ext]. The tag_string is a space-separated list of tags, such as 1girl ahoge long_hair. The tag_count_general setting is used to filter images based on their tag count for training purposes.

Project Structure

Each training project in DeepDanbooru is organized into a folder that contains essential configuration files, such as project.json and tags.txt.

• project.json: Describes general settings and paths for the project.
• tags.txt: Lists all tags that the system will use for estimation. This file should be a simple list of tags, each on a new line.

Example content of tags.txt:

1girl
ahoge
...

With this setup, DeepDanbooru provides a robust platform for creating and testing anime-style image tag estimation models.