---
tags:
- image-classification
- timm
- transformers
- animetimm
- dghs-imgutils
library_name: timm
license: gpl-3.0
datasets:
- animetimm/danbooru-wdtagger-v4-w640-ws-full
base_model:
- timm/mobilenetv4_conv_aa_large.e230_r448_in12k_ft_in1k
---

# Anime Tagger mobilenetv4_conv_aa_large.dbv4-full

## Model Details

- **Model Type:** Multilabel Image classification / feature backbone
- **Model Stats:**
  - Params: 47.3M
  - FLOPs / MACs: 19.2G / 9.6G
  - Image size: train = 448 x 448, test = 448 x 448
- **Dataset:** [animetimm/danbooru-wdtagger-v4-w640-ws-full](https://huggingface.co/datasets/animetimm/danbooru-wdtagger-v4-w640-ws-full)
  - Tags Count: 12476
    - General (#0) Tags Count: 9225
    - Character (#4) Tags Count: 3247
    - Rating (#9) Tags Count: 4

## Results

|     #      |    Macro@0.40 (F1/MCC/P/R)    |    Micro@0.40 (F1/MCC/P/R)    |  Macro@Best (F1/P/R)  |
|:----------:|:-----------------------------:|:-----------------------------:|:---------------------:|
| Validation | 0.457 / 0.473 / 0.586 / 0.407 | 0.641 / 0.642 / 0.693 / 0.596 |          ---          |
|    Test    | 0.458 / 0.473 / 0.586 / 0.408 | 0.641 / 0.642 / 0.694 / 0.596 | 0.511 / 0.537 / 0.513 |

* `Macro/Micro@0.40` means the metrics on the threshold 0.40.
* `Macro@Best` means the mean metrics on the tag-level thresholds on each tags, which should have the best F1 scores.

## Thresholds

|  Category  |   Name    |  Alpha  |  Threshold  |  Micro@Thr (F1/P/R)   |  Macro@0.40 (F1/P/R)  |  Macro@Best (F1/P/R)  |
|:----------:|:---------:|:-------:|:-----------:|:---------------------:|:---------------------:|:---------------------:|
|     0      |  general  |    1    |    0.33     | 0.631 / 0.639 / 0.624 | 0.323 / 0.475 / 0.272 | 0.388 / 0.404 / 0.405 |
|     4      | character |    1    |    0.42     | 0.873 / 0.921 / 0.829 | 0.840 / 0.901 / 0.793 | 0.860 / 0.913 / 0.819 |
|     9      |  rating   |    1    |    0.38     | 0.807 / 0.759 / 0.862 | 0.813 / 0.788 / 0.841 | 0.814 / 0.783 / 0.851 |

* `Micro@Thr` means the metrics on the category-level suggested thresholds, which are listed in the table above.
* `Macro@0.40` means the metrics on the threshold 0.40.
* `Macro@Best` means the metrics on the tag-level thresholds on each tags, which should have the best F1 scores.

For tag-level thresholds, you can find them in [selected_tags.csv](https://huggingface.co/animetimm/mobilenetv4_conv_aa_large.dbv4-full/resolve/main/selected_tags.csv).

## How to Use

We provided a sample image for our code samples, you can find it [here](https://huggingface.co/animetimm/mobilenetv4_conv_aa_large.dbv4-full/blob/main/sample.webp).

### Use TIMM And Torch

Install [dghs-imgutils](https://github.com/deepghs/imgutils), [timm](https://github.com/huggingface/pytorch-image-models) and other necessary requirements with the following command

```shell
pip install 'dghs-imgutils>=0.17.0' torch huggingface_hub timm pillow pandas
```

After that you can load this model with timm library, and use it for train, validation and test, with the following code

```python
import json

import pandas as pd
import torch
from huggingface_hub import hf_hub_download
from imgutils.data import load_image
from imgutils.preprocess import create_torchvision_transforms
from timm import create_model

repo_id = 'animetimm/mobilenetv4_conv_aa_large.dbv4-full'
model = create_model(f'hf-hub:{repo_id}', pretrained=True)
model.eval()

with open(hf_hub_download(repo_id=repo_id, repo_type='model', filename='preprocess.json'), 'r') as f:
    preprocessor = create_torchvision_transforms(json.load(f)['test'])
# Compose(
#     PadToSize(size=(512, 512), interpolation=bilinear, background_color=white)
#     Resize(size=448, interpolation=bicubic, max_size=None, antialias=True)
#     CenterCrop(size=[448, 448])
#     MaybeToTensor()
#     Normalize(mean=tensor([0.4850, 0.4560, 0.4060]), std=tensor([0.2290, 0.2240, 0.2250]))
# )

image = load_image('https://huggingface.co/animetimm/mobilenetv4_conv_aa_large.dbv4-full/resolve/main/sample.webp')
input_ = preprocessor(image).unsqueeze(0)
# input_, shape: torch.Size([1, 3, 448, 448]), dtype: torch.float32
with torch.no_grad():
    output = model(input_)
    prediction = torch.sigmoid(output)[0]
# output, shape: torch.Size([1, 12476]), dtype: torch.float32
# prediction, shape: torch.Size([12476]), dtype: torch.float32

df_tags = pd.read_csv(
    hf_hub_download(repo_id=repo_id, repo_type='model', filename='selected_tags.csv'),
    keep_default_na=False
)
tags = df_tags['name']
mask = prediction.numpy() >= df_tags['best_threshold']
print(dict(zip(tags[mask].tolist(), prediction[mask].tolist())))
# {'sensitive': 0.6181432604789734,
#  '1girl': 0.9969968795776367,
#  'solo': 0.9696205258369446,
#  'looking_at_viewer': 0.8432332873344421,
#  'blush': 0.7917149662971497,
#  'smile': 0.9405843615531921,
#  'short_hair': 0.6273495554924011,
#  'shirt': 0.5353975892066956,
#  'long_sleeves': 0.7138653993606567,
#  'brown_hair': 0.8164870738983154,
#  'holding': 0.6878705024719238,
#  'dress': 0.6111152172088623,
#  'closed_mouth': 0.5007601976394653,
#  'white_shirt': 0.34434816241264343,
#  'purple_eyes': 0.7064062356948853,
#  'flower': 0.9301103949546814,
#  'braid': 0.869755208492279,
#  'sidelocks': 0.23272593319416046,
#  'outdoors': 0.4784768223762512,
#  'hand_up': 0.17624469101428986,
#  'blunt_bangs': 0.3426509499549866,
#  'head_tilt': 0.11214721202850342,
#  'sunlight': 0.15041665732860565,
#  'plant': 0.22287964820861816,
#  'light_smile': 0.08559338748455048,
#  'blue_flower': 0.8238141536712646,
#  'backlighting': 0.17485418915748596,
#  'crown_braid': 0.6755908131599426}
```
### Use ONNX Model For Inference

Install [dghs-imgutils](https://github.com/deepghs/imgutils) with the following command

```shell
pip install 'dghs-imgutils>=0.17.0'
```

Use `multilabel_timm_predict` function with the following code

```python
from imgutils.generic import multilabel_timm_predict

general, character, rating = multilabel_timm_predict(
    'https://huggingface.co/animetimm/mobilenetv4_conv_aa_large.dbv4-full/resolve/main/sample.webp',
    repo_id='animetimm/mobilenetv4_conv_aa_large.dbv4-full',
    fmt=('general', 'character', 'rating'),
)

print(general)
# {'1girl': 0.9969969987869263,
#  'solo': 0.969620406627655,
#  'smile': 0.940584659576416,
#  'flower': 0.9301101565361023,
#  'braid': 0.8697538375854492,
#  'looking_at_viewer': 0.8432332277297974,
#  'blue_flower': 0.8238140344619751,
#  'brown_hair': 0.816490650177002,
#  'blush': 0.7917153835296631,
#  'long_sleeves': 0.7138651609420776,
#  'purple_eyes': 0.7064056396484375,
#  'holding': 0.6878722906112671,
#  'crown_braid': 0.6755940318107605,
#  'short_hair': 0.6273516416549683,
#  'dress': 0.6111209392547607,
#  'shirt': 0.5354008078575134,
#  'closed_mouth': 0.5007631182670593,
#  'outdoors': 0.47849011421203613,
#  'white_shirt': 0.34435153007507324,
#  'blunt_bangs': 0.34265023469924927,
#  'sidelocks': 0.2327292263507843,
#  'plant': 0.22287851572036743,
#  'hand_up': 0.17624491453170776,
#  'backlighting': 0.17485374212265015,
#  'sunlight': 0.15041813254356384,
#  'head_tilt': 0.11214578151702881,
#  'light_smile': 0.08559340238571167}
print(character)
# {}
print(rating)
# {'sensitive': 0.6181411743164062}
```

For further information, see [documentation of function multilabel_timm_predict](https://dghs-imgutils.deepghs.org/main/api_doc/generic/multilabel_timm.html#multilabel-timm-predict).