# ViTamin

**Category:** Hybrid · **Code:**
[`timm/models/vitamin.py`](https://github.com/huggingface/pytorch-image-models/blob/main/timm/models/vitamin.py) ·
**Architectures:** 13 · **Pretrained weights:** 15 (0 with ImageNet-1k results)

ViTamin is a vision model designed for vision-language models, with convolutional (MBConv) stages followed by
transformer stages. The weights are CLIP image towers.

## Papers

* [ViTamin: Designing Scalable Vision Models in the Vision-Language Era](https://arxiv.org/abs/2404.02132)

## Usage

See the [Quickstart](../quickstart) for loading pretrained weights, image preprocessing and fine-tuning, and [Feature
Extraction](../feature_extraction) for using the models as backbones.

```py
>>> import timm
>>> model = timm.create_model('vitamin_base_224.datacomp1b_clip', pretrained=True)
>>> data_config = timm.data.resolve_model_data_config(model)
```

## Model API support

| Feature | `VisionTransformer` |
|---|---|
| [`features_only=True` feature maps](../model_api#intermediate-features) | ✓ (strides 16, 16, 16) |
| [`forward_intermediates()`](../model_api#intermediate-features) | ✓ |
| [`prune_intermediate_layers()`](../model_api#intermediate-features) | ✓ |
| [Any input size, no changes needed](../model_api#input-size) | ✗ |
| [`set_input_size()`](../model_api#input-size) | ✓ |
| [`dynamic_img_size=True`](../model_api#input-size) | ✓ |
| [`set_grad_checkpointing()`](../model_api#training-hooks) | ✓ |
| [Layer-wise LR decay (`group_matcher()`)](../model_api#training-hooks) | ✓ |
| [`reset_classifier()`](../model_api#classifier) | ✓ |

✓ supported, ✓ (square) works for square sizes only, ✗ not supported, – not needed (the model already accepts any input
size). Checked by running each feature on a randomly initialized model (`VisionTransformer` with `vitamin_small_224` at
224), other architectures of the same class are expected to behave the same. See the [Model API](../model_api) page for
details.

## Architectures

| Architecture | Params (M) | GMACs | Input size | Pretrained weights |
|---|---:|---:|---:|---:|
| `vitamin_base_224` | 87.72 | 22.68 | 224 | 2 |
| `vitamin_large2_224` | 333.58 | 75.05 | 224 | 1 |
| `vitamin_large2_256` | 333.64 | 99.00 | 256 | 1 |
| `vitamin_large2_336` | 333.83 | 175.72 | 336 | 1 |
| `vitamin_large2_384` | 333.97 | 234.44 | 384 | 1 |
| `vitamin_large_224` | 333.32 | 75.05 | 224 | 1 |
| `vitamin_large_256` | 333.38 | 99.00 | 256 | 1 |
| `vitamin_large_336` | 333.57 | 175.72 | 336 | 1 |
| `vitamin_large_384` | 333.71 | 234.44 | 384 | 1 |
| `vitamin_small_224` | 22.17 | 5.92 | 224 | 2 |
| `vitamin_xlarge_256` | 436.06 | 130.13 | 256 | 1 |
| `vitamin_xlarge_336` | 436.06 | 230.18 | 336 | 1 |
| `vitamin_xlarge_384` | 436.06 | 306.38 | 384 | 1 |

GMACs and parameter counts are for the default input size and classifier of each architecture, from the [benchmark
results](https://github.com/huggingface/pytorch-image-models/tree/main/results).

## Pretrained weights

Top-1 is ImageNet-1k validation accuracy from
[`results-imagenet.csv`](https://github.com/huggingface/pytorch-image-models/blob/main/results/results-imagenet.csv) at
the listed eval size, - if the weight has no ImageNet-1k result (e.g. pretraining or image-text weights). Classes is the
size of the pretrained classifier, none for weights without one.

| Weight | Input size | Top-1 (%) | Eval size | Classes | License | Hub |
|---|---:|---:|---:|---:|---|---|
| `vitamin_base_224.datacomp1b_clip` | 224 | - | - | 768 | mit | [jienengchen/ViTamin-B](https://huggingface.co/jienengchen/ViTamin-B) |
| `vitamin_base_224.datacomp1b_clip_ltt` | 224 | - | - | 768 | mit | [jienengchen/ViTamin-B-LTT](https://huggingface.co/jienengchen/ViTamin-B-LTT) |
| `vitamin_large2_224.datacomp1b_clip` | 224 | - | - | 1024 | mit | [jienengchen/ViTamin-L2-224px](https://huggingface.co/jienengchen/ViTamin-L2-224px) |
| `vitamin_large2_256.datacomp1b_clip` | 256 | - | - | 1024 | mit | [jienengchen/ViTamin-L2-256px](https://huggingface.co/jienengchen/ViTamin-L2-256px) |
| `vitamin_large2_336.datacomp1b_clip` | 336 | - | - | 1024 | mit | [jienengchen/ViTamin-L2-336px](https://huggingface.co/jienengchen/ViTamin-L2-336px) |
| `vitamin_large2_384.datacomp1b_clip` | 384 | - | - | 1024 | mit | [jienengchen/ViTamin-L2-384px](https://huggingface.co/jienengchen/ViTamin-L2-384px) |
| `vitamin_large_224.datacomp1b_clip` | 224 | - | - | 768 | mit | [jienengchen/ViTamin-L-224px](https://huggingface.co/jienengchen/ViTamin-L-224px) |
| `vitamin_large_256.datacomp1b_clip` | 256 | - | - | 768 | mit | [jienengchen/ViTamin-L-256px](https://huggingface.co/jienengchen/ViTamin-L-256px) |
| `vitamin_large_336.datacomp1b_clip` | 336 | - | - | 768 | mit | [jienengchen/ViTamin-L-336px](https://huggingface.co/jienengchen/ViTamin-L-336px) |
| `vitamin_large_384.datacomp1b_clip` | 384 | - | - | 768 | mit | [jienengchen/ViTamin-L-384px](https://huggingface.co/jienengchen/ViTamin-L-384px) |
| `vitamin_small_224.datacomp1b_clip` | 224 | - | - | 384 | mit | [jienengchen/ViTamin-S](https://huggingface.co/jienengchen/ViTamin-S) |
| `vitamin_small_224.datacomp1b_clip_ltt` | 224 | - | - | 768 | mit | [jienengchen/ViTamin-S-LTT](https://huggingface.co/jienengchen/ViTamin-S-LTT) |
| `vitamin_xlarge_256.datacomp1b_clip` | 256 | - | - | 1152 | mit | [jienengchen/ViTamin-XL-256px](https://huggingface.co/jienengchen/ViTamin-XL-256px) |
| `vitamin_xlarge_336.datacomp1b_clip` | 336 | - | - | 1152 | mit | [jienengchen/ViTamin-XL-336px](https://huggingface.co/jienengchen/ViTamin-XL-336px) |
| `vitamin_xlarge_384.datacomp1b_clip` | 384 | - | - | 1152 | mit | [jienengchen/ViTamin-XL-384px](https://huggingface.co/jienengchen/ViTamin-XL-384px) |

## Citation

```bibtex
@inproceedings{chen2024vitamin,
  title={ViTamin: Designing Scalable Vision Models in the Vision-language Era},
  author={Chen, Jieneng and Yu, Qihang and Shen, Xiaohui and Yuille, Alan and Chen, Liang-Chieh},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2024}
}
```