You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
279 lines
7.5 KiB
279 lines
7.5 KiB
3 years ago
|
# Vision Transformer (ViT)
|
||
3 years ago
|
|
||
|
|
The **Vision Transformer** is a model for image classification that employs a Transformer-like architecture over patches of the image. This includes the use of [Multi-Head Attention](https://paperswithcode.com/method/multi-head-attention), [Scaled Dot-Product Attention](https://paperswithcode.com/method/scaled) and other architectural features seen in the [Transformer](https://paperswithcode.com/method/transformer) architecture traditionally used for NLP.
|
||
|
|
|
||
|
|
{% include 'code_snippets.md' %}
|
||
|
|
|
||
|
|
## How do I train this model?
|
||
|
|
|
||
|
|
You can follow the [timm recipe scripts](https://rwightman.github.io/pytorch-image-models/scripts/) for training a new model afresh.
|
||
|
|
|
||
|
|
## Citation
|
||
|
|
|
||
|
|
```BibTeX
|
||
|
|
@misc{dosovitskiy2020image,
|
||
|
|
title={An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale},
|
||
|
|
author={Alexey Dosovitskiy and Lucas Beyer and Alexander Kolesnikov and Dirk Weissenborn and Xiaohua Zhai and Thomas Unterthiner and Mostafa Dehghani and Matthias Minderer and Georg Heigold and Sylvain Gelly and Jakob Uszkoreit and Neil Houlsby},
|
||
|
|
year={2020},
|
||
|
|
eprint={2010.11929},
|
||
|
|
archivePrefix={arXiv},
|
||
|
|
primaryClass={cs.CV}
|
||
|
|
}
|
||
|
|
```
|
||
|
|
|
||
|
|
<!--
|
||
|
|
Models:
|
||
|
|
- Name: vit_large_patch16_384
|
||
|
|
Metadata:
|
||
|
|
FLOPs: 174702764032
|
||
|
|
Batch Size: 512
|
||
|
|
Training Data:
|
||
|
|
- ImageNet
|
||
|
|
- JFT-300M
|
||
|
|
Training Techniques:
|
||
|
|
- Cosine Annealing
|
||
|
|
- Gradient Clipping
|
||
|
|
- SGD with Momentum
|
||
|
|
Training Resources: TPUv3
|
||
|
|
Architecture:
|
||
|
|
- Attention Dropout
|
||
|
|
- Convolution
|
||
|
|
- Dense Connections
|
||
|
|
- Dropout
|
||
|
|
- GELU
|
||
|
|
- Layer Normalization
|
||
|
|
- Multi-Head Attention
|
||
|
|
- Scaled Dot-Product Attention
|
||
|
|
- Tanh Activation
|
||
|
|
File Size: 1218907013
|
||
|
|
Tasks:
|
||
|
|
- Image Classification
|
||
|
|
Training Time: ''
|
||
|
|
ID: vit_large_patch16_384
|
||
|
|
Crop Pct: '1.0'
|
||
|
|
Momentum: 0.9
|
||
|
|
Image Size: '384'
|
||
|
|
Weight Decay: 0.0
|
||
|
|
Interpolation: bicubic
|
||
|
|
Code: https://github.com/rwightman/pytorch-image-models/blob/5f9aff395c224492e9e44248b15f44b5cc095d9c/timm/models/vision_transformer.py#L561
|
||
|
|
Config: ''
|
||
|
|
In Collection: Vision Transformer
|
||
|
|
- Name: vit_base_patch16_224
|
||
|
|
Metadata:
|
||
|
|
FLOPs: 67394605056
|
||
|
|
Epochs: 90
|
||
|
|
Batch Size: 4096
|
||
|
|
Training Data:
|
||
|
|
- ImageNet
|
||
|
|
- JFT-300M
|
||
|
|
Training Techniques:
|
||
|
|
- Cosine Annealing
|
||
|
|
- Gradient Clipping
|
||
|
|
- SGD with Momentum
|
||
|
|
Training Resources: TPUv3
|
||
|
|
Architecture:
|
||
|
|
- Attention Dropout
|
||
|
|
- Convolution
|
||
|
|
- Dense Connections
|
||
|
|
- Dropout
|
||
|
|
- GELU
|
||
|
|
- Layer Normalization
|
||
|
|
- Multi-Head Attention
|
||
|
|
- Scaled Dot-Product Attention
|
||
|
|
- Tanh Activation
|
||
|
|
File Size: 346292833
|
||
|
|
Tasks:
|
||
|
|
- Image Classification
|
||
|
|
Training Time: ''
|
||
|
|
ID: vit_base_patch16_224
|
||
|
|
LR: 0.0008
|
||
|
|
Dropout: 0.0
|
||
|
|
Crop Pct: '0.9'
|
||
|
|
Image Size: '224'
|
||
|
|
Warmup Steps: 10000
|
||
|
|
Weight Decay: 0.03
|
||
|
|
Interpolation: bicubic
|
||
|
|
Code: https://github.com/rwightman/pytorch-image-models/blob/5f9aff395c224492e9e44248b15f44b5cc095d9c/timm/models/vision_transformer.py#L503
|
||
|
|
Config: ''
|
||
|
|
In Collection: Vision Transformer
|
||
|
|
- Name: vit_base_patch16_384
|
||
|
|
Metadata:
|
||
|
|
FLOPs: 49348245504
|
||
|
|
Batch Size: 512
|
||
|
|
Training Data:
|
||
|
|
- ImageNet
|
||
|
|
- JFT-300M
|
||
|
|
Training Techniques:
|
||
|
|
- Cosine Annealing
|
||
|
|
- Gradient Clipping
|
||
|
|
- SGD with Momentum
|
||
|
|
Training Resources: TPUv3
|
||
|
|
Architecture:
|
||
|
|
- Attention Dropout
|
||
|
|
- Convolution
|
||
|
|
- Dense Connections
|
||
|
|
- Dropout
|
||
|
|
- GELU
|
||
|
|
- Layer Normalization
|
||
|
|
- Multi-Head Attention
|
||
|
|
- Scaled Dot-Product Attention
|
||
|
|
- Tanh Activation
|
||
|
|
File Size: 347460194
|
||
|
|
Tasks:
|
||
|
|
- Image Classification
|
||
|
|
Training Time: ''
|
||
|
|
ID: vit_base_patch16_384
|
||
|
|
Crop Pct: '1.0'
|
||
|
|
Momentum: 0.9
|
||
|
|
Image Size: '384'
|
||
|
|
Weight Decay: 0.0
|
||
|
|
Interpolation: bicubic
|
||
|
|
Code: https://github.com/rwightman/pytorch-image-models/blob/5f9aff395c224492e9e44248b15f44b5cc095d9c/timm/models/vision_transformer.py#L522
|
||
|
|
Config: ''
|
||
|
|
In Collection: Vision Transformer
|
||
|
|
- Name: vit_large_patch16_224
|
||
|
|
Metadata:
|
||
|
|
FLOPs: 119294746624
|
||
|
|
Batch Size: 512
|
||
|
|
Training Data:
|
||
|
|
- ImageNet
|
||
|
|
- JFT-300M
|
||
|
|
Training Techniques:
|
||
|
|
- Cosine Annealing
|
||
|
|
- Gradient Clipping
|
||
|
|
- SGD with Momentum
|
||
|
|
Training Resources: TPUv3
|
||
|
|
Architecture:
|
||
|
|
- Attention Dropout
|
||
|
|
- Convolution
|
||
|
|
- Dense Connections
|
||
|
|
- Dropout
|
||
|
|
- GELU
|
||
|
|
- Layer Normalization
|
||
|
|
- Multi-Head Attention
|
||
|
|
- Scaled Dot-Product Attention
|
||
|
|
- Tanh Activation
|
||
|
|
File Size: 1217350532
|
||
|
|
Tasks:
|
||
|
|
- Image Classification
|
||
|
|
Training Time: ''
|
||
|
|
ID: vit_large_patch16_224
|
||
|
|
Crop Pct: '0.9'
|
||
|
|
Momentum: 0.9
|
||
|
|
Image Size: '224'
|
||
|
|
Weight Decay: 0.0
|
||
|
|
Interpolation: bicubic
|
||
|
|
Code: https://github.com/rwightman/pytorch-image-models/blob/5f9aff395c224492e9e44248b15f44b5cc095d9c/timm/models/vision_transformer.py#L542
|
||
|
|
Config: ''
|
||
|
|
In Collection: Vision Transformer
|
||
|
|
- Name: vit_base_patch32_384
|
||
|
|
Metadata:
|
||
|
|
FLOPs: 12656142336
|
||
|
|
Batch Size: 512
|
||
|
|
Training Data:
|
||
|
|
- ImageNet
|
||
|
|
- JFT-300M
|
||
|
|
Training Techniques:
|
||
|
|
- Cosine Annealing
|
||
|
|
- Gradient Clipping
|
||
|
|
- SGD with Momentum
|
||
|
|
Training Resources: TPUv3
|
||
|
|
Architecture:
|
||
|
|
- Attention Dropout
|
||
|
|
- Convolution
|
||
|
|
- Dense Connections
|
||
|
|
- Dropout
|
||
|
|
- GELU
|
||
|
|
- Layer Normalization
|
||
|
|
- Multi-Head Attention
|
||
|
|
- Scaled Dot-Product Attention
|
||
|
|
- Tanh Activation
|
||
|
|
File Size: 353210979
|
||
|
|
Tasks:
|
||
|
|
- Image Classification
|
||
|
|
Training Time: ''
|
||
|
|
ID: vit_base_patch32_384
|
||
|
|
Crop Pct: '1.0'
|
||
|
|
Momentum: 0.9
|
||
|
|
Image Size: '384'
|
||
|
|
Weight Decay: 0.0
|
||
|
|
Interpolation: bicubic
|
||
|
|
Code: https://github.com/rwightman/pytorch-image-models/blob/5f9aff395c224492e9e44248b15f44b5cc095d9c/timm/models/vision_transformer.py#L532
|
||
|
|
Config: ''
|
||
|
|
In Collection: Vision Transformer
|
||
|
|
- Name: vit_base_resnet50_384
|
||
|
|
Metadata:
|
||
|
|
FLOPs: 49461491712
|
||
|
|
Batch Size: 512
|
||
|
|
Training Data:
|
||
|
|
- ImageNet
|
||
|
|
- JFT-300M
|
||
|
|
Training Techniques:
|
||
|
|
- Cosine Annealing
|
||
|
|
- Gradient Clipping
|
||
|
|
- SGD with Momentum
|
||
|
|
Training Resources: TPUv3
|
||
|
|
Architecture:
|
||
|
|
- Attention Dropout
|
||
|
|
- Convolution
|
||
|
|
- Dense Connections
|
||
|
|
- Dropout
|
||
|
|
- GELU
|
||
|
|
- Layer Normalization
|
||
|
|
- Multi-Head Attention
|
||
|
|
- Scaled Dot-Product Attention
|
||
|
|
- Tanh Activation
|
||
|
|
File Size: 395854632
|
||
|
|
Tasks:
|
||
|
|
- Image Classification
|
||
|
|
Training Time: ''
|
||
|
|
ID: vit_base_resnet50_384
|
||
|
|
Crop Pct: '1.0'
|
||
|
|
Momentum: 0.9
|
||
|
|
Image Size: '384'
|
||
|
|
Weight Decay: 0.0
|
||
|
|
Interpolation: bicubic
|
||
|
|
Code: https://github.com/rwightman/pytorch-image-models/blob/5f9aff395c224492e9e44248b15f44b5cc095d9c/timm/models/vision_transformer.py#L653
|
||
|
|
Config: ''
|
||
|
|
In Collection: Vision Transformer
|
||
|
|
- Name: vit_small_patch16_224
|
||
|
|
Metadata:
|
||
|
|
FLOPs: 28236450816
|
||
|
|
Training Data:
|
||
|
|
- ImageNet
|
||
|
|
- JFT-300M
|
||
|
|
Training Techniques:
|
||
|
|
- Cosine Annealing
|
||
|
|
- Gradient Clipping
|
||
|
|
- SGD with Momentum
|
||
|
|
Training Resources: TPUv3
|
||
|
|
Architecture:
|
||
|
|
- Attention Dropout
|
||
|
|
- Convolution
|
||
|
|
- Dense Connections
|
||
|
|
- Dropout
|
||
|
|
- GELU
|
||
|
|
- Layer Normalization
|
||
|
|
- Multi-Head Attention
|
||
|
|
- Scaled Dot-Product Attention
|
||
|
|
- Tanh Activation
|
||
|
|
File Size: 195031454
|
||
|
|
Tasks:
|
||
|
|
- Image Classification
|
||
|
|
Training Time: ''
|
||
|
|
ID: vit_small_patch16_224
|
||
|
|
Crop Pct: '0.9'
|
||
|
|
Image Size: '224'
|
||
|
|
Interpolation: bicubic
|
||
|
|
Code: https://github.com/rwightman/pytorch-image-models/blob/5f9aff395c224492e9e44248b15f44b5cc095d9c/timm/models/vision_transformer.py#L490
|
||
|
|
Config: ''
|
||
|
|
In Collection: Vision Transformer
|
||
|
|
Collections:
|
||
|
|
- Name: Vision Transformer
|
||
|
|
Paper:
|
||
|
|
title: 'An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale'
|
||
|
3 years ago
|
url: https://paperswithcode.com//paper/an-image-is-worth-16x16-words-transformers-1
|
||
|
3 years ago
|
type: model-index
|
||
|
|
Type: model-index
|
||
|
|
-->
|