Review — All Tokens Matter: Token Labeling for Training Better Vision Transformers

MixToken or LV-ViT, Token Labeling for Image Classification

Sik-Ho Tsang
5 min readNov 14, 2022


Proposed LV-ViT Using MixToken

All Tokens Matter: Token Labeling for Training Better Vision Transformers
MixToken, LV-ViT
, by National University of Singapore, Nankai University, Peking University, and ByteDance
2021 NeurIPS, Over 50 Citations (Sik-Ho Tsang @ Medium)
Image Classification, Vision Transformer, ViT

  • Conventional ViTs compute the classification loss on an additional trainable class token, other tokens are not utilized:
  • MixToken takes advantage of all the image patch tokens to compute the training loss in a dense manner.
  • Specifically, token labeling reformulates the image classification problem into multiple token-level recognition problems and assigns each patch token with an individual location-specific supervision generated by a machine annotator.


  1. MixToken or LV-ViT
  2. MixToken with CutMix
  3. Ablation Study
  4. Experimental Results

1. MixToken

1.1. Conventional ViT

  • Given an image I, denote the output of the last transformer block as [Xcls, X1 , …, XN], where N is the total number of patch tokens, and Xcls and X1, .., XN correspond to the class token and the patch tokens, respectively. The classification loss for image I is:
  • where H(.,.) is the softmax cross-entropy loss and ycls is the class label.

The above classification problem only adopts an image-level label as supervision whereas it neglects the rich information embedded in each image patch.

1.2. Token Labeling

NFNet-F6 as Machine Annotator
  • The dense score map can be generated by a pretrained model in advance, e.g.: NFNet-F6.
Pipeline of training ViT with token labeling

MixToken/LV-ViT leverages a dense score map for each training image and use the cross-entropy loss between each output patch token and the corresponding aligned label in the dense score map as an auxiliary loss at the training phase.

  • Thus, the token labeling objective is:
  • The total loss function is:
  • where β=0.5.

Unlike Knowledge Distillation (KD), token labeling is a cheap operation. During training, MixToken only needs to crop the score map and perform interpolation to make it aligned with the cropped image.

2. MixToken with CutMix

Left: CutMix is operated on the input images, Right: our proposed MixToken is operated on the token
  • CutMix is directly on the raw images, let say I1 and I2, some of the resulting patches may contain content from two images, leading to mixed regions within a small patch.
  • (Please feel free to visit CutMix.)
  • MixToken, which can be viewed as a modified version of CutMix operating on the tokens after patch embedding.
  • MixToken mixes the corresponding token labels by using a binary mask M:
  • where ⊙ is element-wise multiplication.
  • The label for the class token can be written as:

3. Ablation Study

Performance of the proposed LV-ViT with different model sizes
  • A slight architecture modification to ViT is that the patch embedding module is replaced with a 4-layer convolution to better tokenize the input image and integrate local information.

LV-ViT-T with only 8.5M parameters can already achieve a top-1 accuracy of 79.1% on ImageNet.

Left: Ablation on the proposed MixToken and token labeling augmentations, Right: Ablation on different widely-used data augmentations.
  • Left: When training with the original ImageNet labels, using MixToken is 0.1% higher than using CutMix. When using the ReLabel supervision, we can also see an advantage of 0.2% over the CutMix baseline. Combining with token labeling, the performance can be further raised to 83.3%.

Right: When all the four augmentation methods are used, a top-1 accuracy of 83.1% is achieved. Interestingly, when the mixup augmentation is removed, the performance can be improved to 83.3%.

Comparison of token labeling (TL), knowledge distillation (KD) based method and ReLabel method based on utilized tokens

For both online and offline cases, using token-level supervision can improve the overall performance with only negligible additional training cost.

Performance of the proposed token labeling objective on three different ViTs: DeiT (Left), T2T-ViT (Middle), and LV-ViT (Right).

For different variants and scales of the models, the improvement is also consistent.

Performance of the proposed token labeling objective on representative CNN-based (ResNeSt) and MLP-based (MLP-Mixer) models

The proposed method has a consistent improvement on all different models.

4. Experimental Results

4.1 ImageNet & ImageNet-ReaL

Top-1 accuracy comparison with other methods on ImageNet and ImageNet-ReaL.

The proposed model consistently outperforms other CNN-based and Transformer-based counterparts.

4.2. ADE20K

Transfer performance of the proposed LV-ViT in semantic segmentation.
Comparison with previous work on ADE20K validation set.

LV-ViT-L + UPerNet achieves the best result on ADE20K with only ImageNet-1K as training data in pretraining.

MixToken is used in such as Pale Transformer (2022 AAAI) to further boost the performance, in which I will review later on.


[2021 NeurIPS] [MixToken]
All Tokens Matter: Token Labeling for Training Better Vision Transformers

1.1. Image Classification

19892021 [MixToken] 2022 [ConvNeXt] [PVTv2] [ViT-G] [AS-MLP] [ResTv2]

My Other Previous Paper Readings



Sik-Ho Tsang

PhD, Researcher. I share what I learn. :) Linktree: for Twitter, LinkedIn, etc.