Vision Transformer (ViT)

ViT: An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

Abstract: While the Transformer architecture has become the de-facto standard for natural language processing tasks, its applications to computer vision remain limited. In vision, attention is either applied in conjunction with convolutional networks, or used to replace certain components of convolutional networks while keeping their overall structure in place. We show that this reliance on CNNs is not necessary and a pure transformer applied directly to sequences of image patches can perform very well on image classification tasks. When pre-trained on large amounts of data and transferred to multiple mid-sized or small image recognition benchmarks (ImageNet, CIFAR-100, VTAB, etc. ), Vision Transformer (ViT) attains excellent results compared to state-of-the-art convolutional networks while requiring substantially fewer computational resources to train.

Tasks: Image Classification, Fine-Grained Image Classification, Document Image Classification

Datasets: CIFAR-10, ImageNet, CIFAR-100

Conference: ICLR 2021

Licence: Apache-2.0

Authors: Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby

https://paperswithcode.com/paper/an-image-is-worth-16x16-words-transformers-1

Kornia-ViT

We provide the operator VisionTransformer that is meant to be used across tasks. One can use the ViT in Kornia as follows:

img = torch.rand(1, 3, 224, 224)
vit = VisionTransformer(image_size=224, patch_size=16)
out = vit(img)

Usage

kornia-vit does not include any classification head. For this reason, we provide an ClassificationHead which can be easily combined with a nn.Sequential in order to easily build a custom image classification pipeline.

import torch.nn as nn
import kornia.contrib as K

classifier = nn.Sequential(
    K.VisionTransformer(image_size=224, patch_size=16),
    K.ClassificationHead(num_classes=1000)
)

img = torch.rand(1, 3, 224, 224)
out = classifier(img)     # BxN
scores = out.argmax(-1)   # B

In addition to create simple image classification, our API is flexible enough to design your pipelines e.g to solve problems for multi-task, object detection, segmentation, etc. We show an example of a multi-task class with two different classification heads:

class MultiTaskTransfornmer(nn.Module):
    def __init__(self) -> None:
        super().__init__()
        self.transformer = K.VisionTransformer(
            image_size=224, patch_size=16)
        self.head1 = K.ClassificationHead(num_classes=10)
        self.head2 = K.ClassificationHead(num_classes=50)

    def forward(self, x: torch.Tensor):
        out = self.transformer(x)
        return {
            "head1": self.head1(out),
            "head2": self.head2(out),
        }

Tip

More heads, examples and a training API soon !!