Skip to content

Latest commit

 

History

History
47 lines (43 loc) · 3.18 KB

README.md

File metadata and controls

47 lines (43 loc) · 3.18 KB

Main Results on ImageNet with Pretrained Models

We train Semantic FPN models on the ADE20K dataset for 40,000 iterations using standard training settings from the MMSegmentation toolkit. The input images are cropped to a resolution of 512×512 during training.

Model Params(M) Latency(ms) mIoU Ckpt. Log
iFormer-M 8.9 4.00 42.4 40000 iters 40000 iters
iFormer-L 14.7 6.60 44.5 40000 iters 40000 iters
iFormer-L2 24.5 9.06 46.2 40000 iters 40000 iters
  • As detection, we have lost some original checkpoints corresponding to the performance reported in the paper due to certain errors. As a result, the performance may not be consistent with those presented in the paper. In some cases, our reproduced results surpass those reported in the paper.

Getting Started

Requirements

Please see the official MMSegmentation for installation.

Here we provide ours installation process:

Download mmcv

pip install mmcv-2.1.0-cp38-cp38-manylinux1_x86_64.whl

Download the MMSegmentation source code and compile from the source.

git clone https://github.com/open-mmlab/mmsegmentation.git
cd mmsegmentation
git checkout b040e147adfa027bbc071b624bedf0ae84dfc922
pip install -v -e .

Our specific configuration details:

mmcv                      2.1.0
mmdet                     3.3.0
mmengine                  0.10.5
mmsegmentation            1.2.2

Data Preparation

Prepare the challenging ADE20K dataset according to the instructions in MMSeg.

Training

cd iFormer/segmentation
sh tools/dist_train.sh configs/sem_fpn/fpn_iformer_m_ade20k_40k.py 8 --work-dir=./output/seg_m_0

Evaluation

cd iFormer/detection
checkpoint_path=your checkpoint path
sh tools/dist_test.sh configs/sem_fpn/fpn_iformer_m_ade20k_40k.py $check_path 1 --work-dir=./output/seg_m_0