| --- |
| title: Continual learning results (%) on data separated by region groups |
| license: mit |
| task_categories: |
| - image-classification |
| tags: |
| - continual-learning |
| - resnet |
| - vit |
| - vit-hybrid |
| - swin-transformer |
| --- |
| |
| # Continual Learning Results (Region Groups) |
|
|
| This repository contains the weights and experimental results of continual learning fine-tuning using several mainstream vision backbones. The methods evaluated include FineTune (Cross Entropy), EWC, VCL, and FIP. |
|
|
| ## 📂 Directory Structure |
|
|
| ```text |
| . |
| ├── FineTune_Cross_Entropy/ |
| │ ├── model/ |
| │ │ ├── config.json # Model configuration and hyperparameters |
| │ │ └── pytorch_model.bin # Model weights file |
| │ └── result/ |
| │ ├── misclassified_images.csv # Records of misclassified images |
| │ ├── test_evaluation.csv # Evaluation metrics and test results |
| │ ├── training_time.csv # Logged training duration per group (hours) |
| │ └── training_metrics_set1.csv ~ training_metrics_set7.csv # Training dynamics and output logs |
| ├── EWC/ |
| │ ├── model/ |
| │ │ ├── config.json |
| │ │ └── pytorch_model.bin |
| │ └── result/ |
| │ ├── misclassified_images.csv |
| │ ├── test_evaluation.csv |
| │ ├── training_time.csv |
| │ └── training_metrics_set1.csv ~ training_metrics_set7.csv |
| ├── VCL/ |
| │ ├── model/ |
| │ │ ├── config.json |
| │ │ └── pytorch_model.bin |
| │ └── result/ |
| │ ├── misclassified_images.csv |
| │ ├── test_evaluation.csv |
| │ ├── training_time.csv |
| │ └── training_metrics_set1.csv ~ training_metrics_set7.csv |
| └── FIP/ |
| ├── model/ |
| │ ├── config.json |
| │ └── pytorch_model.bin |
| └── result/ |
| ├── misclassified_images.csv |
| ├── test_evaluation.csv |
| ├── training_time.csv |
| └── training_metrics_set1.csv ~ training_metrics_set7.csv |
| ``` |
|
|
| ## 🚀 Usage |
|
|
| You can download the model weights for a specific continual learning method using the `huggingface_hub` SDK: |
|
|
| ```python |
| from huggingface_hub import hf_hub_download |
| |
| # Example: Download Swin Transformer model weights for the VCL method |
| model_weight_path = hf_hub_download( |
| repo_id="Dataset-and-Model/continual-learning-region-groups", |
| filename="Swin Transformer/VCL/model/pytorch_model.bin", |
| repo_type="model" |
| ) |
| |
| print(f"Weights successfully downloaded to: {model_weight_path}") |
| ``` |
|
|
| ## 🏗️ Base Models |
|
|
| The fine-tuning and continual learning experiments in this study are built upon the following pre-trained models provided by Hugging Face: |
| * **ResNet-34** (e.g., `microsoft/resnet-34`) |
| * **Vision Transformer (ViT)** (e.g., `google/vit-large-patch32-384`) |
| * **ViT Hybrid** (e.g., `google/vit-hybrid-base-bit-384`) |
| * **SWIN Transformer** (e.g., `microsoft/swin-large-patch4-window12-384`) |
|
|
| ## 📜 License |
|
|
| These fine-tuned model weights are derived from the open-source architectures mentioned above and are released under the **MIT License**. You are free to use, modify, and distribute these weights in both academic research and commercial applications, provided that proper attribution to the original authors is maintained. |
|
|
