---
library_name: timm
pipeline_tag: image-classification
tags:
- timm
- bamboo
- arxiv:2203.07845
---

# vit_base_patch16_224.bamboo

A ViT-B/16 pretrained on Bamboo-CLS, converted from the
[authors' release](https://github.com/ZhangYuanhan-AI/Bamboo) to native timm weights. No fine-tuning or weight rescaling was performed.
The checkpoint wrapper and any `module.` prefix were removed. The original 115,217-class classifier is retained.

## Model details

- Paper: [Bamboo](https://arxiv.org/abs/2203.07845)
- Parameters: 174,400,529
- Input: RGB, 224 x 224
- Classifier outputs: 115,217
- Feature dimension: 768
- Training data: Bamboo-CLS

## Preprocessing

Uses the authors' [benchmark configuration](https://github.com/ZhangYuanhan-AI/Bamboo/blob/f6337a76b3e48a1c5eff2bbd5b9dd044e42f0654/Bamboo-Benchmark/configs/models_cfg/ViTB16_bamboo_v0_1.yaml): bicubic resize of the shorter side
to 224, then a 224 x 224 center crop (`crop_pct=1.0`).

- Mean: `(0.48145466, 0.4578275, 0.40821073)`
- Standard deviation: `(0.26862954, 0.26130258, 0.27577711)`

The ViT benchmark uses CLIP normalization; the ResNet benchmark uses ImageNet normalization.
The authors' separate ViT classification demo uses a different recipe. This export follows the benchmark.

## Usage

```python
import timm
import torch
from PIL import Image

model = timm.create_model("hf-hub:timm/vit_base_patch16_224.bamboo", pretrained=True).eval()
transform = timm.data.create_transform(**timm.data.resolve_model_data_config(model), is_training=False)
image = Image.open("image.jpg").convert("RGB")
with torch.inference_mode():
    output = model(transform(image).unsqueeze(0))

probabilities, indices = output.softmax(dim=-1).topk(5)
labels = model.pretrained_cfg["label_names"]
descriptions = model.pretrained_cfg["label_descriptions"]
for index, probability in zip(indices[0].tolist(), probabilities[0].tolist()):
    label_id = labels[index]
    print(label_id, descriptions[label_id], probability)
```

## Labels

`config.json` includes `label_names` (classifier index to stable Bamboo concept ID) and
`label_descriptions` (concept ID to English names/synonyms). Indices are ordered numerically
using the authors' `trainid2nid.json`, not the ordering of the broader Bamboo taxonomy.
The original `trainid2name.json` and `trainid2nid.json` are included alongside the weights.

Names are not unique: distinct concepts can share an English name. Preserve concept IDs
when interpreting predictions. Blank source names use their concept ID as a fallback.
Fallback rows: 5831 (`Q24760423`).

Pass `num_classes=0` to extract pooled 768-dimensional features, or set `num_classes`
to replace the pretrained classifier for fine-tuning.

## Source and licensing

Source URLs and SHA256 hashes are recorded in `conversion.json`.
The authors' [demo metadata](https://huggingface.co/spaces/CVPR/Bamboo_ViT-B16_demo/blob/2bbf27dfb670d5e1b1d3f1e0b97596f95e4c62a0/README.md) declares CC-BY-4.0;
the [dataset access instructions](https://github.com/ZhangYuanhan-AI/Bamboo/tree/f6337a76b3e48a1c5eff2bbd5b9dd044e42f0654#downloads) state academic,
non-commercial use. These are upstream statements, not a new license grant for the converted weights.

## Citation

```bibtex
@misc{zhang2022bamboo,
  title={Bamboo: Building Mega-Scale Vision Dataset Continually with Human-Machine Synergy},
  author={Yuanhan Zhang and Qinghong Sun and Yichun Zhou and Zexin He and Zhenfei Yin and
           Kun Wang and Lu Sheng and Yu Qiao and Jing Shao and Ziwei Liu},
  year={2022},
  eprint={2203.07845},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}
```