ADE20k Semantic Segmentation with BEiT

Getting Started

Install the mmsegmentation library and some required packages.

bash

pip install mmcv-full==1.3.0 mmsegmentation==0.11.0
pip install scipy timm==0.3.2

Install apex for mixed-precision training

bash

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" ./

Follow the guide in mmseg to prepare the ADE20k dataset.

Fine-tuning

Command format:

tools/dist_train.sh <CONFIG_PATH> <NUM_GPUS>  --work-dir <SAVE_PATH> --seed 0  --deterministic --options model.pretrained=<IMAGENET_CHECKPOINT_PATH/URL>

Using a BEiT-base backbone with UperNet:

bash

bash tools/dist_train.sh \
    configs/beit/upernet/upernet_beit_base_12_512_slide_160k_21ktoade20k.py 8 \
    --work-dir /path/to/save --seed 0  --deterministic \
    --options model.pretrained=https://github.com/addf400/files/releases/download/BEiT-v2/beitv2_base_patch16_224_pt1k_ft21k.pth

Using a BEiT-large backbone with UperNet:

bash

bash tools/dist_train.sh \
    configs/beit/upernet/upernet_beit_large_24_512_slide_160k_21ktoade20k.py 8 \
    --work-dir /path/to/save --seed 0  --deterministic \
    --options model.pretrained=https://github.com/addf400/files/releases/download/BEiT-v2/beitv2_large_patch16_224_pt1k_ft21k.pth

Evaluation

Command format:

tools/dist_test.sh  <CONFIG_PATH> <CHECKPOINT_PATH> <NUM_GPUS> --eval mIoU

For example, evaluate a BEiT-large backbone with UperNet:

bash

bash tools/dist_test.sh configs/beit/upernet/upernet_beit_large_24_512_slide_160k_21ktoade20k.py \
    https://github.com/addf400/files/releases/download/BEiT-v2/beitv2_large_patch16_224_pt1k_ft21ktoade20k.pth  4 --eval mIoU

Expected results:

+--------+-------+-------+-------+
| Scope  | mIoU  | mAcc  | aAcc  |
+--------+-------+-------+-------+
| global | 57.54 | 68.78 | 86.22 |
+--------+-------+-------+-------+

Acknowledgment

This code is built using the mmsegmentation library, Timm library, the Swin repository, XCiT and the SETR repository.