skywolf46/kaunsera-at-home

カウンセラー@Home _{Kaunserā (Japanese): A person whose profession is counseling, or a counselor.}

[ 한국어 ] [ English ]

The Kaunsera@Home project is a lightweight analysis model project based on EfficientNet-B2(v1.1) / MobileNetV2(v1.0), intended for embedding purposes for emotion analysis of anime-style images.

Training Information

Kaunsera@Home is a specialized model for emotion analysis, trained on 23,948 manually labeled images.

More than 98% valid image datasets were used, and accuracy was improved by adjusting only the face region using deepghs's anime_face_detection.

• During the scheduled training of a total of 100 epochs, early stopping was triggered to prevent overfitting, halting the training at 18 epochs.
• The best checkpoint weights at epoch 6 were finally adopted.

How to Use

Kaunsera@Home is designed to receive and process tensor inputs using onnx-runtime. For face region detection and cropping, the use of deepghs's anime_face_detection preprocessing pipeline is recommended.

[ONNX Model Input Specifications]

• Shape: [1, 3, 224, 224] (NCHW format)
• Color Space: RGB (0~1.0 scale)
• Normalization (ImageNet Standard):
  • Mean: [0.485, 0.456, 0.406]
  • Std (Standard Deviation): [0.229, 0.224, 0.225]

Data Labeling

Since Kaunsera@Home was premised on the use of a userscript targeting Korean users, the classification criteria were also conducted in Korean.

The following emotions can be derived through the Argmax (index of the highest value) of the output tensor array.

Index	Korean Value	English Value
0	놀람	surprised
1	무표정	neutral
2	슬픔	sad
3	웃음	happy
4	화남	angry

Hyperparameters

Kaunsera@Home was trained applying AMP.

Parameter	Value	Description
Base Model	MobileNetV2(v1.0)	Trained based on Google/MobileNetV2
Base Model	EfficientNet-B2(v1.1)	Trained based on Google/EfficientNet-B2
Optimizer	AdamW	-
Learning Rate	1e-4 (0.0001)	-
Batch Size	256	-
Epochs	6(18)/100	Stopped at 18 out of 100 scheduled epochs due to continuous decrease in accuracy, adopted 6th epoch checkpoint

Training Results

• v1.1 Final Accuracy: 65.89%

• v1.1 Final Loss: 0.6862

• v1.0 Final Accuracy: 63.09%

• v1.0 Final Loss: 0.6787

Miscellaneous

Kaunsera@Home is part of the Crystallized Chasm Project.

The Crystallized Chasm project is maintained by the IGX team.

License

Kaunsera@Home has adopted a dual license to contribute to the ecosystem and prevent commercial abuse.

The weights (.onnx) and related code of this model fundamentally follow the Creative Commons Attribution-NonCommercial 4.0 (CC BY-NC 4.0) license. Anyone can use it freely and free of charge for personal research, learning, and hobbyist open-source projects.

For Individuals and Indie Developers (Commercial Use Fully Allowed)

Because Crystallized Chasm is a non-profit project, a non-commercial license was applied. However, we deeply understand that the open-source ecosystem should be provided with proper compensation and opportunities for revenue generation.

Therefore, if you meet the conditions below, you are exceptionally permitted to use the weights of this project commercially and freely engage in profit-generating activities. (We do not enforce attribution (compliance with CC BY 4.0), but we would appreciate it if you did.)

• Solo/individual developers and students
• Small indie teams / doujin circles with 10 or fewer members

Enterprise Commercial License (Large Corporations and For-Profit Entities)

If enterprises, institutions, or large for-profit entities that exceed the scale of the indie developer exception clause above wish to introduce this model into their intranet or integrate it into their commercial services (SaaS, apps, etc.) to generate revenue, we would appreciate it if you look for another project.

• Legal action may be taken if unauthorized introduction into corporate services is detected.
• We have no intention of generating corporate capital with this model; however, we are very reluctant to provide "ready-to-use refined weights," rather than reusable code, to giant corporations free of charge. We ask for your understanding in this regard.
• If you absolutely must use it despite this, please visit the IGX team's support channel. Provision may only be possible after some condition negotiations and the consent process of all team members.