| license: apache-2.0 | |
| library_name: timm | |
| # WD SwinV2 Tagger v3 | |
| Supports ratings, characters and general tags. | |
| Trained using https://github.com/SmilingWolf/JAX-CV. | |
| TPUs used for training kindly provided by the [TRC program](https://sites.research.google/trc/about/). | |
| ## Dataset | |
| Last image id: 7220105 | |
| Trained on Danbooru images with IDs modulo 0000-0899. | |
| Validated on images with IDs modulo 0950-0999. | |
| Images with less than 10 general tags were filtered out. | |
| Tags with less than 600 images were filtered out. | |
| ## Validation results | |
| `P=R: threshold = 0.2521, F1 = 0.4411` | |
| ## What's new | |
| Model v1.1/Dataset v3: | |
| Amended the JAX model config file: add image size. | |
| No change to the trained weights. | |
| Model v1.0/Dataset v3: | |
| More training images, more and up-to-date tags (up to 2024-02-28). | |
| Now `timm` compatible! Load it up and give it a spin using the canonical one-liner! | |
| ONNX model is compatible with code developed for the v2 series of models. | |
| The batch dimension of the ONNX model is not fixed to 1 anymore. Now you can go crazy with batch inference. | |
| Switched to Macro-F1 to measure model performance since it gives me a better gauge of overall training progress. | |
| ## Final words | |
| Subject to change and updates. | |
| Downstream users are encouraged to use tagged releases rather than relying on the head of the repo. | |