# LeViT

**Category:** Hybrid · **Code:**
[`timm/models/levit.py`](https://github.com/huggingface/pytorch-image-models/blob/main/timm/models/levit.py) ·
**Architectures:** 20 · **Pretrained weights:** 10 (10 with ImageNet-1k results)

LeViT is a hybrid network for fast inference, a convolutional stem followed by attention stages that shrink the
resolution with attention based downsampling and use attention biases for position.

## Papers

* [LeViT: a Vision Transformer in ConvNet's Clothing for Faster Inference](https://arxiv.org/abs/2104.01136)

## 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('levit_256.fb_dist_in1k', pretrained=True)
>>> data_config = timm.data.resolve_model_data_config(model)
```

## Model API support

| Feature | `Levit` | `LevitDistilled` |
|---|---|---|
| [`features_only=True` feature maps](../model_api#intermediate-features) | ✓ (strides 16, 32, 64) | ✓ (strides 16, 32, 64) |
| [`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 (`Levit` with `levit_256d` at 224,
`LevitDistilled` with `levit_128s` 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 | Class | Params (M) | GMACs | Input size | Pretrained weights |
|---|---|---:|---:|---:|---:|
| `levit_128` | `LevitDistilled` | 9.19 | 0.41 | 224 | 1 |
| `levit_128s` | `LevitDistilled` | 7.76 | 0.30 | 224 | 1 |
| `levit_192` | `LevitDistilled` | 10.92 | 0.66 | 224 | 1 |
| `levit_256` | `LevitDistilled` | 18.86 | 1.13 | 224 | 1 |
| `levit_256d` | `Levit` | 26.16 | 1.39 | 224 | 0 |
| `levit_384` | `LevitDistilled` | 39.07 | 2.35 | 224 | 1 |
| `levit_384_s8` | `LevitDistilled` | 39.06 | 9.95 | 224 | 0 |
| `levit_512` | `Levit` | 95.08 | 5.62 | 224 | 0 |
| `levit_512_s8` | `Levit` | 73.97 | 21.77 | 224 | 0 |
| `levit_512d` | `Levit` | 92.39 | 5.84 | 224 | 0 |
| `levit_conv_128` | `LevitDistilled` | 9.19 | 0.41 | 224 | 1 |
| `levit_conv_128s` | `LevitDistilled` | 7.76 | 0.30 | 224 | 1 |
| `levit_conv_192` | `LevitDistilled` | 10.92 | 0.66 | 224 | 1 |
| `levit_conv_256` | `LevitDistilled` | 18.86 | 1.13 | 224 | 1 |
| `levit_conv_256d` | `Levit` | 26.16 | 1.39 | 224 | 0 |
| `levit_conv_384` | `LevitDistilled` | 39.07 | 2.35 | 224 | 1 |
| `levit_conv_384_s8` | `LevitDistilled` | 39.06 | 9.95 | 224 | 0 |
| `levit_conv_512` | `Levit` | 95.08 | 5.62 | 224 | 0 |
| `levit_conv_512_s8` | `Levit` | 73.97 | 21.77 | 224 | 0 |
| `levit_conv_512d` | `Levit` | 92.39 | 5.84 | 224 | 0 |

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 |
|---|---:|---:|---:|---:|---|---|
| `levit_128.fb_dist_in1k` | 224 | 78.486 | 224 | 1000 | apache-2.0 | [timm](https://huggingface.co/timm/levit_128.fb_dist_in1k) |
| `levit_128s.fb_dist_in1k` | 224 | 76.520 | 224 | 1000 | apache-2.0 | [timm](https://huggingface.co/timm/levit_128s.fb_dist_in1k) |
| `levit_192.fb_dist_in1k` | 224 | 79.858 | 224 | 1000 | apache-2.0 | [timm](https://huggingface.co/timm/levit_192.fb_dist_in1k) |
| `levit_256.fb_dist_in1k` | 224 | 81.506 | 224 | 1000 | apache-2.0 | [timm](https://huggingface.co/timm/levit_256.fb_dist_in1k) |
| `levit_384.fb_dist_in1k` | 224 | 82.600 | 224 | 1000 | apache-2.0 | [timm](https://huggingface.co/timm/levit_384.fb_dist_in1k) |
| `levit_conv_128.fb_dist_in1k` | 224 | 78.488 | 224 | 1000 | apache-2.0 | [timm](https://huggingface.co/timm/levit_conv_128.fb_dist_in1k) |
| `levit_conv_128s.fb_dist_in1k` | 224 | 76.534 | 224 | 1000 | apache-2.0 | [timm](https://huggingface.co/timm/levit_conv_128s.fb_dist_in1k) |
| `levit_conv_192.fb_dist_in1k` | 224 | 79.860 | 224 | 1000 | apache-2.0 | [timm](https://huggingface.co/timm/levit_conv_192.fb_dist_in1k) |
| `levit_conv_256.fb_dist_in1k` | 224 | 81.506 | 224 | 1000 | apache-2.0 | [timm](https://huggingface.co/timm/levit_conv_256.fb_dist_in1k) |
| `levit_conv_384.fb_dist_in1k` | 224 | 82.600 | 224 | 1000 | apache-2.0 | [timm](https://huggingface.co/timm/levit_conv_384.fb_dist_in1k) |

## Citation

```bibtex
@article{graham2021levit,
  title={LeViT: a Vision Transformer in ConvNet's Clothing for Faster Inference},
  author={Benjamin Graham and Alaaeldin El-Nouby and Hugo Touvron and Pierre Stock and Armand Joulin and Herv'e J'egou and Matthijs Douze},
  journal={arXiv preprint arXiv:2104.01136},
  year={2021}
}
```