MetaFormer Baselines for Vision

This is a PyTorch implementation of several MetaFormer baslines including IdentityFormer, RandFormer, ConvFormer and CAFormer proposed by our paper "MetaFormer Baselines for Vision".

Figure 1: Performance of MetaFormer baselines and other state-of-the-art models on ImageNet-1K at 224x224 resolution. The architectures of our proposed models are shown in Figure 2. (a) IdentityFormer/RandFormer achieve over 80%/81% accuracy, indicating MetaFormer has solid lower bound of performance and works well on arbitrary token mixers. The accuracy of well-trained ResNet-50 is from "ResNet strikes back". (b) Without novel token mixers, pure CNN-based ConvFormer outperforms ConvNeXt, while CAFormer sets a new record of 85.5% accuracy on ImageNet-1K at 224x224 resolution under normal supervised training without external data or distillation.

Figure 2: (a-d) Overall frameworks of IdentityFormer, RandFormer, ConvFormer and CAFormer. Similar to ResNet, the models adopt hierarchical architecture of 4 stages, and stage $i$ has $L_i$ blocks with feature dimension $D_i$. Each downsampling module is implemented by a layer of convolution. The first downsampling has kernel size of 7 and stride of 4, while the last three ones have kernel size of 3 and stride of 2. (e-h) Architectures of IdentityFormer, RandFormer, ConvFormer and Transformer blocks, which have token mixer of identity mapping, global random mixing, separable depthwise convolutions, or vanilla self-attention, respectively.

News

Models of MetaFormer baselines are now integrated in timm by Fredo Guan and Ross Wightman. Many thanks!

Requirements

torch>=1.7.0; torchvision>=0.8.0; pyyaml; timm (pip install timm==0.6.11)

Data preparation: ImageNet with the following folder structure, you can extract ImageNet by this script.

│imagenet/
├──train/
│  ├── n01440764
│  │   ├── n01440764_10026.JPEG
│  │   ├── n01440764_10027.JPEG
│  │   ├── ......
│  ├── ......
├──val/
│  ├── n01440764
│  │   ├── ILSVRC2012_val_00000293.JPEG
│  │   ├── ILSVRC2012_val_00002138.JPEG
│  │   ├── ......
│  ├── ......

MetaFormer baselines

Models with common token mixers trained on ImageNet-1K

😲 😲 * To the best of our knowledge, the model sets a new record on ImageNet-1K with the accuracy of 85.5% at 224x224 resolution under normal supervised setting (without external data or distillation).

Models with common token mixers pretrained on ImageNet-21K and finetuned on ImgeNet-1K

Models with common token mixers pretrained on ImageNet-21K

Models with basic token mixers trained on ImageNet-1K

The underlined numbers mean the numbers of parameters that are frozen after random initialization.

The checkpoints can also be found in Baidu Disk.

Usage

We also provide a Colab notebook which run the steps to perform inference with MetaFormer baselines:

Validation

To evaluate our CAFormer-S18 models, run:

MODEL=caformer_s18
python3 validate.py /path/to/imagenet  --model $MODEL -b 128 \
  --checkpoint /path/to/checkpoint 

Train

We use batch size of 4096 by default and we show how to train models with 8 GPUs. For multi-node training, adjust --grad-accum-steps according to your situations.

DATA_PATH=/path/to/imagenet
CODE_PATH=/path/to/code/metaformer # modify code path here


ALL_BATCH_SIZE=4096
NUM_GPU=8
GRAD_ACCUM_STEPS=4 # Adjust according to your GPU numbers and memory size.
let BATCH_SIZE=ALL_BATCH_SIZE/NUM_GPU/GRAD_ACCUM_STEPS


cd $CODE_PATH && sh distributed_train.sh $NUM_GPU $DATA_PATH \
--model convformer_s18 --opt adamw --lr 4e-3 --warmup-epochs 20 \
-b $BATCH_SIZE --grad-accum-steps $GRAD_ACCUM_STEPS \
--drop-path 0.2 --head-dropout 0.0

Training (fine-tuning) scripts of other models are shown in scripts.

Acknowledgment

Weihao Yu would like to thank TRC program and GCP research credits for the support of partial computational resources. Our implementation is based on the wonderful pytorch-image-models codebase.

Bibtex

@article{yu2022metaformer_baselines,
  title={Metaformer baselines for vision},
  author={Yu, Weihao and Si, Chenyang and Zhou, Pan and Luo, Mi and Zhou, Yichen and Feng, Jiashi and Yan, Shuicheng and Wang, Xinchao},
  journal={arXiv preprint arXiv:2210.13452},
  year={2022}
}