Skip to content

Latest commit

 

History

History
120 lines (94 loc) · 4.12 KB

README.md

File metadata and controls

120 lines (94 loc) · 4.12 KB
Raw

ActiveMLP for Semantic Segmentation

This is the official impelmentation of ActiveMLP for semantic segmentation.

Results and models

ADE20K

UperNet

Backbone Method #iters mIoU mIoU (ms) #params. FLOPs (512x2048) model
Active-xT UperNet 160K 44.27 45.44 45.3M 889.5G
Active-T UperNet 160K 46.46 47.55 57.1M 926.9G
Active-S UperNet 160K 48.42 49.49 68.6M 988.3G
Active-B UperNet 160K 48.63 49.86 82.1M 1055.9G
Active-L UperNet 160K 50.05 51.10 107.9M 1105.6G model/log

Semantic FPN

Backbone Method #iters mIoU #params. FLOPs (512x512) model
Active-xT Semantic FPN 40K 43.03 19.1M 33.1G
Active-T Semantic FPN 40K 45.84 30.9M 42.4G
Active-S Semantic FPN 40K 47.30 42.4M 57.8G
Active-B Semantic FPN 40K 47.73 55.9M 74.7G
Active-L Semantic FPN 40K 48.05 79.8M 86.6G

Usage

Install and data preparation

  • Install required packages following ActiveMLP.
  • Install mmcv and mmsegmentation with:
pip install mmcv-full==1.3.0 mmsegmentation==0.11.0
  • 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" ./
  • Prepare ADE20K following the guidelines in mmsegmentation.

Evaluation

  • To evaluate a pre-trained segmentation model, run with:
# single-scale mIoU
python -m torch.distributed.launch --nproc_per_node <num-gpus> \
    tools/test.py <config-file> <checkpoint.pth> \
    --launcher pytorch --eval mIoU

# multi-scale mIoU
python -m torch.distributed.launch --nproc_per_node <num-gpus> \
    tools/test.py <config-file> <checkpoint.pth> \
    --launcher pytorch --eval mIoU --aug-test

For example, to evaluate the multi-scale mIoU of ActiveMLP-Large+UperNet with two GPUs distributedly:

python -m torch.distributed.launch --nproc_per_node 2 \
    tools/test.py configs/activemlp/active_large_ade20k_upernet_160k.py \
    active_large_upernet_160k.pth --launcher pytorch --eval mIoU --aug-test

This should give:

+--------+------+-------+-------+
| Scope  | mIoU | mAcc  | aAcc  |
+--------+------+-------+-------+
| global | 51.1 | 61.95 | 83.86 |
+--------+------+-------+-------+
  • To get the FLOPs and parameters of ActiveMLP-Large+UperNet, run with:
# FLOPs with input 512x2048
python tools/get_flops.py 512 2048 --config configs/activemlp/active_large_ade20k_upernet_160k.py

# output
>>> model: ActiveLarge from with: configs/activemlp/active_large_ade20k_upernet_160k.py | shape=[512, 2048]
    FLOPs: 1105.632G
    #param: 107.892M

Training

To train with pre-trained models, run with:

python -m torch.distributed.launch --nproc_per_node <num-gpus> \
    tools/train.py <config-file> \
    --launcher pytorch \
    --options model.pretrained=<checkpoint.pth> [other options]

For example, to train Active-Large+UperNet on 8 gpus, run with:

python -m torch.distributed.launch --nproc_per_node 8 \
    tools/train.py configs/activemlp/active_large_ade20k_upernet_160k.py \
    --launcher pytorch \
    --options model.pretrained=activemlp_large.pth

Citing

@article{wei2022activemlp,
    title={ActiveMLP: An MLP-like Architecture with Active Token Mixer},
    author={Wei, Guoqiang and Zhang, Zhizheng and Lan, Cuiling and Lu, Yan and Chen, Zhibo},
    journal={arXiv preprint arXiv:2203.06108},
    year={2022}
}