File size: 1,764 Bytes
056e4c5 296b77d 056e4c5 7ece298 790b0e9 7ece298 790b0e9 5c6bcf3 790b0e9 5c6bcf3 7ece298 790b0e9 7ece298 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 |
---
license: apache-2.0
library_name: timm
---
# WD ViT 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
`v2.0: P=R: threshold = 0.2614, F1 = 0.4402`
`v1.0: P=R: threshold = 0.2547, F1 = 0.4278`
## What's new
Model v2.0/Dataset v3:
Trained for a few more epochs.
Used tag frequency-based loss scaling to combat class imbalance.
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.
# Runtime deps
ONNX model requires `onnxruntime >= 1.17.0`
# Inference code examples
For timm: https://github.com/neggles/wdv3-timm
For ONNX: https://huggingface.co/spaces/SmilingWolf/wd-tagger
For JAX: https://github.com/SmilingWolf/wdv3-jax
## Final words
Subject to change and updates.
Downstream users are encouraged to use tagged releases rather than relying on the head of the repo.
|