Preacher-26/touhou-embeddings-dataset

Touhou Project Dataset (Images + WD-ConvNeXt Features)

[ English ] | 中文

Dataset Description (En)

This dataset contains images of characters from the Touhou Project series, paired with pre-computed feature embeddings and logistic regression classifiers. It is intended for multi-label classification research, linear probing experiments, and character recognition tasks.

Key Features

• Embeddings are extracted from the penultimate layer of SmilingWolf/wd-convnext-tagger-v3.
• Includes .npy files for training/validation splits, allowing for immediate model training without heavy image preprocessing.
• Includes joblib files containing trained scikit-learn One-vs-Rest logistic regression models for the included characters.

Statistics

• Total Samples: 10,786
• Split:
  • Train: 8,687
  • Validation: 2,099
• Classes (Characters): 134 unique tags

Distribution Imbalance

The dataset is unbalanced. Popular characters have significantly more samples than niche characters.

• Top 3: Hakurei Reimu (1,022), Kirisame Marisa (833), Cirno (696).
• Bottom 3: Motoori Kosuzu (224), Ruukoto (102), Satsuki Rin (100).

Directory Structure

.
├── images/                # Source images (various formats)
├── embeddings_backbone/   # Pre-computed features
│   ├── X_train.npy        # Training features
│   ├── y_train.npy        # Training labels (multi-hot encoded)
│   ├── X_val.npy          # Validation features
│   ├── y_val.npy          # Validation labels
│   ├── multi_label_binarizer.joblib      # Sklearn LabelBinarizer object
│   └── touhou_classifier_br_list.joblib  # List of trained LogisticRegression models
└── labels.csv             # Metadata (id, filename, split, raw_tags)

Notes

• Missing Data: Due to network issues during collection, the character list is not exhaustive. Some Touhou characters are not present in this version.
• The dataset includes 100 samples of Satsuki Rin, representing the smallest class.

🚀 Quick Start: Inference with Pre-computed Embeddings

You can use the pre-computed validation embeddings (X_val.npy) and the trained classifier list (touhou_classifier_br_list.joblib) to evaluate the model without downloading images or loading the heavy backbone model.

Requirements: pip install numpy scikit-learn joblib huggingface_hub

import joblib
import numpy as np
from huggingface_hub import hf_hub_download

# 1. Download necessary artifacts (or use local paths if cloned)
REPO_ID = "Preacher-26/touhou-embeddings-dataset"  # Replace with your actual repo ID
SUBFOLDER = "embeddings_backbone"

print("Loading artifacts...")
# Load Classifiers (List of 134 LogisticRegression models)
clf_path = hf_hub_download(
    repo_id=REPO_ID, subfolder=SUBFOLDER, filename="touhou_classifier_br_list.joblib"
)
classifiers = joblib.load(clf_path)

# Load Label Binarizer (Maps indices to Character Names)
mlb_path = hf_hub_download(
    repo_id=REPO_ID, subfolder=SUBFOLDER, filename="multi_label_binarizer.joblib"
)
mlb = joblib.load(mlb_path)

# Load a sample batch of features (Validation Set)
x_val_path = hf_hub_download(repo_id=REPO_ID, subfolder=SUBFOLDER, filename="X_val.npy")
X_val = np.load(x_val_path)

# 2. Perform Inference on a random sample
# Pick a random sample index from the validation set
sample_idx = np.random.randint(0, len(X_val))
sample_embedding = X_val[sample_idx].reshape(1, -1)  # Shape: (1, 1024)

print(f"\nRunning inference on sample index: {sample_idx}")

# The model is a list of One-vs-Rest classifiers. We must iterate through them.
probs = []
for clf in classifiers:
    # Handle DummyClassifier (used for classes with 0 training samples)
    if hasattr(clf, "predict_proba"):
        prob = clf.predict_proba(sample_embedding)[0, 1]
    else:
        prob = 0.0
    probs.append(prob)

probs = np.array(probs)

# 3. Decode and Print Results
# Using threshold 0.2 (optimized for F1-Score as per report)
THRESHOLD = 0.2
active_indices = np.where(probs >= THRESHOLD)[0]

print(f"--- Predictions (Threshold: {THRESHOLD}) ---")
if len(active_indices) == 0:
    print("No characters detected above threshold.")
else:
    for idx in active_indices:
        tag_name = mlb.classes_[idx]
        confidence = probs[idx]
        print(f"Character: {tag_name:<25} | Confidence: {confidence:.4f}")

# (Optional) Show Top-3 Raw Probabilities
top3_indices = np.argsort(probs)[-3:][::-1]
print("\n--- Top 3 Raw Probabilities ---")
for idx in top3_indices:
    print(f"{mlb.classes_[idx]:<25}: {probs[idx]:.4f}")

📚 Citation

If you use this dataset in your research, please cite it as follows:

@misc{touhou-embeddings-dataset,
    author = {Preacher-26},
    title = {touhou-embeddings-dataset},
    year = {2025},

---

数据集说明 (Zh)

本数据集包含《东方 Project》系列角色的图片，以及基于 WD-ConvNeXt Tagger 提取的特征向量和预训练分类器。该数据集主要用于多标签分类研究、Linear Probing 实验及角色识别任务。

主要特性

• 嵌入向量 (Embeddings) 提取自 SmilingWolf/wd-convnext-tagger-v3 模型的倒数第二层。
• 包含已处理好的 .npy 格式训练/验证集张量，无需进行繁重的图像预处理即可直接用于下游模型训练。
• 包含已训练好的 scikit-learn One-vs-Rest 逻辑回归模型 (.joblib)。

统计信息

• 样本总数: 10,786 张
• 数据集划分:
  • 训练集 (Train): 8,687
  • 验证集 (Val): 2,099
• 类别数 (角色): 134 个独立标签

数据分布

数据分布存在不均衡现象，热门角色样本量远多于长尾角色。

• 头部角色 (Top 3): 博丽灵梦 (1,022), 雾雨魔理沙 (833), 琪露诺 (696)。
• 尾部角色 (Bottom 3): 本居小铃 (224), Ruukoto (102), 冴月麟 (100)。

目录结构说明

• images/: 原始图片文件。
• embeddings_backbone/:
  • X_train.npy / X_val.npy: 提取出的特征向量。
  • y_train.npy / y_val.npy: 经过 Multi-hot 编码的标签。
  • touhou_classifier_br_list.joblib: 包含 100+个二分类逻辑回归模型的列表。
• labels.csv: 元数据表，包含 ID、文件名、划分情况及原始标签。

注意事项

1. 数据缺失: 受限于采集时的网络状况，本数据集并未覆盖东方 Project 的所有角色。
2. 即使是样本最少的类别（如冴月麟），也保留了约 100 张样本，具备一定的 Few-shot 学习价值。