Skip to content

Latest commit

 

History

History
237 lines (174 loc) · 9.12 KB

README.md

File metadata and controls

237 lines (174 loc) · 9.12 KB
Raw

MS-MLP: Mixing and Shifting in Vision Transformers

This is the official implementation of our MS-MLP -- "Mixing and Shifting: Exploiting Global and Local Dependencies in Vision MLPs", by Huangjie Zheng, Pengcheng He, Weizhu Chen and Mingyuan Zhou.

multi-scale-regional-mixing-teaser

The proposed mixing and shifting operation exploit both long-range and short-range dependencies without self-attention. In a MLP-based archtecture, Mix-Shift-MLP (MS-MLP) makes the size of the local receptive field used for mixing increase with respect to the amount of relative distance achived by the spatial shifting. This directly contributes to the interactions between neighbor and distant tokens.

Model Overview

msblock-teaser

In each MS-block, we first split the feature map into several groups along the channel dimension, with the first group regarded as the source of query tokens. In the other groups, as the centers of the attended regions (marked with yellow stars) become more and more distant, we gradually increase the mixing spatial range from 1 x 1 to 7 x 7. After the mixing operation, we shift the split channel groups to align their mixed center tokens with the query and then continue the channel-wise mixing with a channel MLP.

Image Classification on ImageNet-1K

Network Resolution Top-1 (%) Params FLOPs Throughput (image/s) model
MS-MLP-Tiny 224x224 82.1 28M 4.9G 792.0 download
MS-MLP-Small 224x224 83.4 50M 9.0G 483.8 download
MS-MLP-Base 224x224 83.8 88M 16.1G 366.5 download

Getting Started

Install

  • Clone this repo:
git clone https://github.com/JegZheng/MS-MLP
cd MS-MLP
  • Create a conda virtual environment and activate it:
conda create -n msmlp python=3.7 -y
conda activate msmlp
conda install pytorch==1.7.1 torchvision==0.8.2 cudatoolkit=10.1 -c pytorch
  • Install timm==0.3.2:
pip install timm==0.3.2
  • Install cupy-cuda101 (optional, if you choose to use kernel speed up used in AS-MLP):
pip install cupy-cuda101
  • Install Apex:
git clone https://github.com/NVIDIA/apex
cd apex
pip install -v --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./
  • Install other requirements:
pip install opencv-python==4.4.0.46 termcolor==1.1.0 yacs==0.1.8

Data preparation

We use standard ImageNet dataset, you can download it from http://image-net.org/. We provide the following two ways to load data:

  • For standard folder dataset, move validation images to labeled sub-folders. The file structure should look like:

    $ tree data
imagenet
├── train
│   ├── class1
│   │   ├── img1.jpeg
│   │   ├── img2.jpeg
│   │   └── ...
│   ├── class2
│   │   ├── img3.jpeg
│   │   └── ...
│   └── ...
└── val
    ├── class1
    │   ├── img4.jpeg
    │   ├── img5.jpeg
    │   └── ...
    ├── class2
    │   ├── img6.jpeg
    │   └── ...
    └── ...

  • To boost the slow speed when reading images from massive small files, we also support zipped ImageNet, which includes four files:

    • train.zip, val.zip: which store the zipped folder for train and validate splits.
    • train_map.txt, val_map.txt: which store the relative path in the corresponding zip file and ground truth label. Make sure the data folder looks like this:
    $ tree data
data
└── ImageNet-Zip
    ├── train_map.txt
    ├── train.zip
    ├── val_map.txt
    └── val.zip

$ head -n 5 data/ImageNet-Zip/val_map.txt
ILSVRC2012_val_00000001.JPEG  65
ILSVRC2012_val_00000002.JPEG  970
ILSVRC2012_val_00000003.JPEG  230
ILSVRC2012_val_00000004.JPEG  809
ILSVRC2012_val_00000005.JPEG  516

$ head -n 5 data/ImageNet-Zip/train_map.txt
n01440764/n01440764_10026.JPEG  0
n01440764/n01440764_10027.JPEG  0
n01440764/n01440764_10029.JPEG  0
n01440764/n01440764_10040.JPEG  0
n01440764/n01440764_10042.JPEG  0

Evaluation

To evaluate a pre-trained MS-MLP on ImageNet val, run:

python -m torch.distributed.launch --nproc_per_node <num-of-gpus-to-use> --master_port 12345 main.py --eval \
--cfg <config-file> --resume <checkpoint> --data-path <imagenet-path>

For example, to evaluate the MS-MLP-Tiny with a single GPU:

python -m torch.distributed.launch --nproc_per_node 1 --nnodes=1 --master_port 12345 main.py --eval \
--cfg configs/msmlp_tiny_patch4_shift5_224.yaml --resume <msmlp-tiny.pth> --data-path <imagenet-path>

Training from scratch

To train a MS-MLP on ImageNet from scratch, run:

python -m torch.distributed.launch --nproc_per_node <num-of-gpus-to-use> --master_port 12345  main.py \ 
--cfg <config-file> --data-path <imagenet-path> [--batch-size <batch-size-per-gpu> --output <output-directory> --tag <job-tag>]

Notes:

  • To use zipped ImageNet instead of folder dataset, add --zip to the parameters.
    • To cache the dataset in the memory instead of reading from files every time, add --cache-mode part, which will shard the dataset into non-overlapping pieces for different GPUs and only load the corresponding one for each GPU.
  • When GPU memory is not enough, you can try the following suggestions:
    • Use gradient accumulation by adding --accumulation-steps <steps>, set appropriate <steps> according to your need.
    • Use gradient checkpointing by adding --use-checkpoint, which can save a lot of GPU memory. Please refer to this page for more details.
    • We recommend using multi-node with more GPUs for training very large models, a tutorial can be found in this page.
  • To change config options in general, you can use --opts KEY1 VALUE1 KEY2 VALUE2, e.g., --opts TRAIN.EPOCHS 100 TRAIN.WARMUP_EPOCHS 5 will change total epochs to 100 and warm-up epochs to 5.
  • For additional options, see config and run python main.py --help to get detailed message.

For example, to train MS-MLP with 8 GPU on a single node for 300 epochs, run:

MS-MLP-Tiny:

python -m torch.distributed.launch --nproc_per_node=8 --nnodes=1 main.py \
--cfg configs/msmlp_tiny_patch4_shift5_224.yaml --data-path <imagenet-path> --batch-size 128 --cache-mode no \
--accumulation-steps 0 --output <output-path>

MS-MLP-Small:

python -m torch.distributed.launch --nproc_per_node=8 --nnodes=1 main.py \
--cfg configs/msmlp_small_patch4_shift5_224.yaml --data-path <imagenet-path> --batch-size 128 --cache-mode no \
--accumulation-steps 0 --output <output-path>

MS-MLP-Base:

python -m torch.distributed.launch --nproc_per_node=8 --nnodes=1 main.py \
--cfg configs/msmlp_base_patch4_shift5_224.yaml --data-path <imagenet-path> --batch-size 64 --cache-mode no \
--accumulation-steps 2 --output <output-path>

For multi-node training, please add --node_rank, --master_addr, --master_port options. For example:

python -m torch.distributed.launch --nproc_per_node=8 --nnodes=2 --node_rank=$RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT  main.py \
--cfg configs/msmlp_base_patch4_shift5_224.yaml --data-path <imagenet-path> --batch-size 64 --cache-mode no \
--accumulation-steps 0 --output <output-path>

Throughput

To measure the throughput, run:

python -m torch.distributed.launch --nproc_per_node 1 --master_port 12345  main.py \
--cfg <config-file> --data-path <imagenet-path> --batch-size 64 --throughput --amp-opt-level O0

Citation

If you find this repo useful to your project, please consider to cite it with following bib:

@misc{zheng2022mixing,
  title={Mixing and Shifting: Exploiting Global and Local Dependencies in Vision MLPs}, 
  author={Huangjie Zheng and Pengcheng He and Weizhu Chen and Mingyuan Zhou},
  year={2022},
  eprint={2202.06510},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}

Acknowledgement

Our codebase is built based on Swin-Transformer, AS-MLP and Focal-Transformer. We thank the authors for the nicely organized code!