Review — FastViT: A Fast Hybrid Vision Transformer using Structural Reparameterization

FastViT, Better Accuracy-Latency Trade-Off

Sik-Ho Tsang
5 min readSep 30


FastViT, Better Accuracy-Latency Trade-Off

A Fast Hybrid Vision Transformer using Structural Reparameterization
, by Apple
2023 ICCV (Sik-Ho Tsang @ Medium)

Image Classification
1989 … 2023
[Vision Permutator (ViP)] [ConvMixer] [CrossFormer++]
==== My Other Paper Readings Are Also Over Here ====

  • A mixing operator, RepMixer, a building block of FastViT, is proposed that uses structural reparameterization to lower the memory access cost by removing skip-connections in the network.
  • Train-time overparametrization and large kernel convolutions are further applied to boost accuracy and empirically show that these choices have minimal effect on latency.


  1. FastViT
  2. Results

1. FastViT

FastViT Overview
  • FastViT applies different architectural choices onto PoolFormer to improve the model: RepMixer, Factorized Dense Convolution, Linear Train-Time Reparameterization, and Large Kernel Convolution.
Contribution of Each Component.

Each Component Contribute to FastViT.

1.1. Reparameterizing Skip Connections

1.1.1. RepMixer

  • Convolutional mixing was first introduced in ConvMixer:
  • where σ is a non-linear activation function and BN is Batch Normalization and DWConv is depthwise convolutional layer.

The operations are reaaranged and the non-linear activation function is removed as shown below:

At inference time, it can be reparameterized to a single depthwise convolutional layer as shown below:

1.1.2. Positional Encodings

Latency Comparisons

Conditional positional encodings, in Twins and CPVT, that is dynamically generated and conditioned on the local neighborhood of the input tokens. These encodings are generated as a result of a depth-wise convolution operator and are added to the patch embeddings.

There is lack of non-linearities in this group of operations, hence this block can be reparameterized.

  • It is found that at 384×384 using RepMixer will lower the latency by 25.1% and at larger resolutions like 1024×1024, latency is lowered significantly by 43.9%.

1.2. Factorized Dense Convolution

All dense k×k convolutions are replaced with its factorized version, i.e. k×k depthwise followed by 1×1 pointwise convolutions.

1.3. Linear Train-Time Reparameterization

Linear Train-Time Reparameterization

Linear train-time overparameterization as in MobileOne is applied. MobileOne-style overparameterization is used in stem, patch embedding, and projection layers which help in boosting performance.

  • From Table 3, this train-time overparameterization improves Top-1 accuracy on ImageNet by 0.6% on FastViT-SA12 model. On a smaller FastViT-S12 variant, Top-1 accuracy improves by 0.9% as in Table 1.

1.4. Large Kernel Convolutions

  • A computationally efficient approach to improve the receptive field of early stages that do not use self-attention is by incorporating depthwise large kernel convolutions.

Depthwise large kernel convolutions are introduced in FFN and patch embedding layers.

  • From Table 4, it is noted that variants using depthwise large kernel convolutions can be highly competitive to variants using self-attention layers while incurring a modest increase in latency.
  • Overall, as in Table 1, large kernel convolutions provide 0.9% improvement in Top-1 accuracy on FastViT-S12.

1.5. FastViT Variants

FastViT Variants

FastViTs of different scales are designed as above.

2. Results

SOTA Comparisons

2.1. Image Classification

Table 5: FastViT improve over LITv2 [42] on both parameter count and FLOPs.

  • At Top-1 accuracy of 84.9%, FastViT-MA36 is 49.3% smaller and consumes 55.4% less FLOPs than LITv2-B. FastViT-S12 is 26.3% faster than MobileOne-S4 on iPhone 12 Pro and 26.9% faster on GPU.
  • At Top-1 accuracy of 83.9%, FastViT-MA36 is 1.9× faster than an optimized ConvNeXt-B model on iPhone 12 Pro and 2.0× faster on GPU.
  • At Top-1 accuracy of 84.9%, FastViT-MA36 is just as fast as NFNet-F1 on GPU while being 66.7% smaller and using 50.1% less FLOPs and 42.8% faster on mobile device.

2.2. Knowledge Distillation

Table 6: With Distillation, FastViT-SA24 attains similar performance as EfficientFormer-L7 while having 3.8× less parameters, 2.7× less FLOPs and 2.7× lower latency.

2.3. Robustness Evaluation

Table 7: FastViT is highly competitive to RVT and ConvNeXt, in fact FastViT-M36 has better clean accuracy, better robustness to corruptions and similar outof- distribution robustness as ConvNeXt-S which has 6.1M more parameters and has 10% more FLOPs than our model.

2.4. 3D Hand Mesh Estimation

3D Hand Mesh Estimation
  • The 3D hand mesh estimation backbones usually belong to ResNet or MobileNet family of architectures with the exception of METRO and MeshGraphormer which use HRNets.
  • With FastViT uses as backbone, authors also replace the complex mesh regression head with a simple regression module.

Table 8: Amongst real-time methods, FastViT outperforms other methods on all joint and vertex error related metrics while being 1.9× faster than MobileHand [16] and 2.8× faster than recent state-of-art MobRecon.

Qualitative Results

2.5. Semantic Segmentation

Table 9: FastViT-MA36 model obtains 5.2% higher mIoU than PoolFormer-M36 which has higher FLOPs, parameter count and latency on both desktop GPU and mobile device.

2.6. Object Detection and Instance Segmentation

Table 10: FastViT-MA36 model has similar performance as CMT-S, while being 2.4× and 4.3× faster on desktop GPU and mobile device respectively.



Sik-Ho Tsang

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