wesjos
/

fairface-monkeycake-3task-224

@@ -1,97 +1,96 @@
----
 language:
-  - zh
   - en
-license: cc-by-4.0
-library_name: pytorch
-pipeline_tag: image-classification
-tags:
-  - image-classification
-  - fairface
-  - face-attribute
-  - pytorch
-  - timm
-datasets:
-  - HuggingFaceM4/FairFace
-base_model: timm/eva02_large_patch14_224.mim_in22k
-metrics:
-  - accuracy
-  - f1
----
-# FairFace-MonkeyCake-3Task-224
-## 概述
-基于 **EVA02-L/14 (MIM-22K)** 的人脸属性分类模型，在 [HuggingFaceM4/FairFace](https://huggingface.co/datasets/HuggingFaceM4/FairFace) 数据集上训练。
-输出种族 / 性别 / 年龄三项属性。
-> 教师模型，精度最高但 303M、344 img/s，适合离线评测，不适合实时部署。
-## 模型详情
-| 项 | 值 |
-|---|---|
-| 骨干 | EVA02-L/14 (MIM-22K) ([timm/eva02_large_patch14_224.mim_in22k](https://huggingface.co/timm/eva02_large_patch14_224.mim_in22k)) |
-| 参数量 | 303,286,225 |
-| 输入 | RGB, 224×224 |
-| 归一化 | mean=(0.4815, 0.4578, 0.4082), std=(0.2686, 0.2613, 0.2758) |
-| 输出 | race 7 + gender 2 + age 8 阈值（有序） |
-| 测试集 | 10,931 张（FairFace 官方 val 去重后冻结） |
-## 训练数据
-[HuggingFaceM4/FairFace](https://huggingface.co/datasets/HuggingFaceM4/FairFace)（Karkkainen & Joo, WACV 2021），原始 97,698 张。本项目按 SHA-256 内容去重后：
-train 77,926 / selection 8,659 / test 10,931，三对交集为 0，官方 train 的图 0 张进入测试集。
-## 评测结果
-| 任务 | 准确率 | macro-F1 |
-|---|---:|---:|
-| 种族（7 类） | 77.02% | 76.83% |
-| 性别（2 类） | 96.87% | 96.86% |
-| 年龄（9 档有序） | 63.93% | 61.29% |
-三项平均准确率 **79.27%**。年龄 MAE **0.3838 档**，±1 档命中 **97.94%**。
-## 逐类表现
-![per-class recall](assets/per_class_recall.png)
-![confusion matrix](assets/confusion_matrix.png)
-## 使用方式
-```python
-import inference
-model = inference.load("model.pt")
-print(inference.predict(model, "face.jpg"))
-```
-依赖：`torch`, `timm`, `torchvision`, `pillow`。
-## 限制与风险
-- **研究用途**：本模型用于人脸属性研究与公平性审计。**不得用于监控、身份核验、招聘、信贷等高风险自动决策。**
-- **测试集复用**：该冻结测试集在本项目中累计使用 8 次，指标为描述性，不再具备独立确认意义。
-- **单 seed**：训练只跑了一个随机种子，小于 0.3pp 的差异视为噪声。
-- **已知群体差异**：种族边界（尤其 Latino / Middle Eastern / Indian / White 之间）本身模糊；
-  年龄是最弱的一项；部分性别×种族子群错误率显著偏高。
-- **标签来源**：标签来自 FairFace 数据集标注，不代表模型可推断个体真实身份或自我认同。
-## 数据与许可
-- 数据集：[HuggingFaceM4/FairFace](https://huggingface.co/datasets/HuggingFaceM4/FairFace)，CC BY 4.0（图像来自 YFCC100M，原始许可各异）。
-- 模型权重：基于上述数据训练的衍生作品，随本卡以 CC BY 4.0 提供。
-- 预训练骨干模型：[timm/eva02_large_patch14_224.mim_in22k](https://huggingface.co/timm/eva02_large_patch14_224.mim_in22k)，原始权重发布页：EVA02-L/14 (MIM-22K)（原始许可见其发布页）。
-## 引用
-```bibtex
-@inproceedings{karkkainen2021fairface,
-  title={FairFace: Face Attribute Dataset for Balanced Race, Gender, and Age for Bias Measurement and Mitigation},
-  author={Karkkainen, Kimmo and Joo, Jungseock},
-  booktitle={WACV},
-  year={2021}
-}
-```

+---
 language:
   - en
+license: cc-by-4.0
+library_name: pytorch
+pipeline_tag: image-classification
+tags:
+  - image-classification
+  - fairface
+  - face-attribute
+  - pytorch
+  - timm
+datasets:
+  - HuggingFaceM4/FairFace
+base_model: timm/eva02_large_patch14_224.mim_in22k
+metrics:
+  - accuracy
+  - f1
+---
+# FairFace-MonkeyCake-3Task-224
+## Overview
+A face attribute classification model based on **EVA02-L/14 (MIM-22K)**, trained on the [HuggingFaceM4/FairFace](https://huggingface.co/datasets/HuggingFaceM4/FairFace) dataset.
+Predicts three attributes: race, gender, and age.
+> Teacher model with highest overall accuracy; 303M parameters and 344 img/s. Best suited for offline benchmarking rather than real-time edge deployment.
+## Model Details
+| Field | Value |
+|---|---|
+| Backbone | EVA02-L/14 (MIM-22K) ([timm/eva02_large_patch14_224.mim_in22k](https://huggingface.co/timm/eva02_large_patch14_224.mim_in22k)) |
+| Parameters | 303,286,225 |
+| Input | RGB, 224×224 |
+| Normalization | mean=(0.4815, 0.4578, 0.4082), std=(0.2686, 0.2613, 0.2758) |
+| Output | race (7 classes) + gender (2 classes) + age (8 ordinal thresholds / 9 brackets) |
+| Test Set | 10,931 images (frozen deduplicated FairFace official val set) |
+## Training Data
+[HuggingFaceM4/FairFace](https://huggingface.co/datasets/HuggingFaceM4/FairFace)（Karkkainen & Joo, WACV 2021），原始 97,698 张。本项目按 SHA-256 内容去重后：
+train 77,926 / selection 8,659 / test 10,931，三对交集为 0，官方 train 的图 0 张进入测试集。
+## Evaluation Results
+| Task | Accuracy | Macro-F1 |
+|---|---:|---:|
+| Race (7 classes) | 77.02% | 76.83% |
+| Gender (2 classes) | 96.87% | 96.86% |
+| Age (9 ordinal brackets) | 63.93% | 61.29% |
+3-Task average accuracy: **79.27%**. Age MAE: **0.3838 brackets**, ±1 bracket accuracy: **97.94%**.
+## Per-Class Performance
+![per-class recall](assets/per_class_recall.png)
+![confusion matrix](assets/confusion_matrix.png)
+## Usage
+```python
+import inference
+model = inference.load("model.pt")
+print(inference.predict(model, "face.jpg"))
+```
+Dependencies: `torch`, `timm`, `torchvision`, `pillow`.
+## Limitations & Ethical Considerations
+- **Research Use Only**: This model is developed strictly for face attribute research, bias measurement, and fairness auditing. **It must NOT be used for surveillance, identity verification, recruitment, credit evaluation, or any high-risk automated decision-making.**
+- **Test Set Reuse**: This frozen test set was evaluated multiple times across project iterations; reported metrics are descriptive.
+- **Single Random Seed**: Training was conducted under a single seed; differences smaller than 0.3 pp should be treated as noise.
+- **已知群体差异**：种族边界（尤其 Latino / Middle Eastern / Indian / White 之间）本身模糊；
+  年龄是最弱的一项；部分性别×种族子群错误率显著偏高。
+- **Label Attribution**: Labels reflect the categorical annotations of the FairFace dataset and do not represent individual identity, ethnicity, or personal self-identification.
+## Data & Licensing
+- **Dataset**: [HuggingFaceM4/FairFace](https://huggingface.co/datasets/HuggingFaceM4/FairFace), CC BY 4.0 (images sourced from YFCC100M under respective licenses).
+- **Model Checkpoint**: Derivative work trained on FairFace, released under CC BY 4.0.
+- **Pretrained Backbone**: [timm/eva02_large_patch14_224.mim_in22k](https://huggingface.co/timm/eva02_large_patch14_224.mim_in22k) (see upstream release for original license).EVA02-L/14 (MIM-22K)（原始许可见其发布页）。
+## Citation
+```bibtex
+@inproceedings{karkkainen2021fairface,
+  title={FairFace: Face Attribute Dataset for Balanced Race, Gender, and Age for Bias Measurement and Mitigation},
+  author={Karkkainen, Kimmo and Joo, Jungseock},
+  booktitle={WACV},
+  year={2021}
+}
+```