Skip to content

Latest commit

 

History

History
152 lines (112 loc) · 4.21 KB

GET_STARTED.md

File metadata and controls

152 lines (112 loc) · 4.21 KB
Raw

MPViT on ImageNet classification

Main results on ImageNet-1K

🚀 These all models are trained for 300 epochs on ImageNet-1K with the same training recipe as DeiT and CoaT.

model resolution acc@1 #params FLOPs weight
MPViT-T 224x224 78.2 5.8M 1.6G weight
MPViT-XS 224x224 80.9 10.5M 2.9G weight
MPViT-S 224x224 83.0 22.8M 4.7G weight
MPViT-B 224x224 84.3 74.8M 16.4G weight

Preparation

Data

Please refer to DeiT for data preparation.

We recommend to symlink the path(data/ImageNet) to the ImageNet dataset path to data/ as follows

# symlink the ImageNet dataset
cd MPViT
mkdir data
ln -s /path_to_imagenet_dataset data/ImageNet

Check the data/ImageNet path points to /path_to_imagenet correctly.

# ImageNet -> /path_to_imagenet_dataset
ll data

Required packages installation

We use Pytorch==1.7.0, torchvision==0.8.1, cuda==10.1 einops and pytorch-image-models (timm).

Please refer to the official pytorch installation guide.

# Install Pytorch and torchvision
conda install pytorch==1.7.0 torchvision==0.8.1 -c pytorch

# Install timm
pip install timm

# Install einops
pip install einops

Evaluation

To evaluate a pre-trained MPViT on ImageNet val with a single GPU, run:

MPViT-base:

python main.py --eval --resume https://dl.dropbox.com/s/la8w31m0apj2830/mpvit_base.pth --model mpvit_base

This should give

* Acc@1 84.292 Acc@5 96.820 loss 0.916
  Accuracy of the network on the 50000 test images: 84.3%

Here you'll find the command-lines to reproduce the inference results for MPViT models.

MPViT-Tiny 
python main.py --eval --resume https://dl.dropbox.com/s/1cmquqyjmaeeg1n/mpvit_tiny.pth --model mpvit_tiny

giving

* Acc@1 78.210 Acc@5 94.270 loss 1.112
  Accuracy of the network on the 50000 test images: 78.2%
MPViT-Xsmall 
python main.py --eval --resume https://dl.dropbox.com/s/vvpq2m474g8tvyq/mpvit_xsmall.pth --model mpvit_xsmall

giving

* Acc@1 80.896 Acc@5 95.352 loss 1.039
Accuracy of the network on the 50000 test images: 80.9%
MPViT-Small 
python main.py --eval --resume https://dl.dropbox.com/s/y3dnmmy8h4npz7a/mpvit_small.pth --model mpvit_small

giving

* Acc@1 82.960 Acc@5 96.154 loss 0.893
Accuracy of the network on the 50000 test images: 83.0%

Training from scratch

To train MPViT models on ImageNet on a single node with 8 gpus for 300 epochs, run:

python -m torch.distributed.launch --nproc_per_node <num-of-gpus-to-use>  main.py \ 
--model <mpvit_model> --data-path <imagenet-path> --batch-size <batch-size-per-gpu> --output <output-directory>

MPViT-Tiny:

python -m torch.distributed.launch --nproc_per_node=8 --use_env main.py \
--model mpvit_tiny --batch-size 128 --data-path <imagenet-path> --output_dir <output-directory>

MPViT-Xsmall:

python -m torch.distributed.launch --nproc_per_node=8 --use_env main.py \
--model mpvit_xsmall --batch-size 128 --data-path <imagenet-path> --output_dir <output-directory>

MPViT-Small:

python -m torch.distributed.launch --nproc_per_node=8 --use_env main.py \
--model mpvit_small --batch-size 128 --drop-path 0.05 --data-path <imagenet-path> --output_dir <output-directory>

MPViT-Base:

python -m torch.distributed.launch --nproc_per_node=8 --use_env main.py \
--model mpvit_base --batch-size 128 --drop-path 0.3 --data-path <imagenet-path> --output_dir <output-directory>