🧩 BEiT Semantic Segmentation

Semantic segmantation with CUBOX dataset

Original Work https://github.com/microsoft/unilm/tree/master/beit/semantic_segmentation

Related Paper https://arxiv.org/abs/2106.08254

🐋 Docker Image

yoorachoi/beit:6.0-dist

🙋‍♀️ Maintainer**

Yura Choi

📌 Table Of Contents

Experiment Environment
- Hardware
- Setup
- Structure
Scripts
- Train Script
- Test Script
Acknowledgement

💻 Experiment Environment

⚙️ Hardware

Both - nvidia-docker pre-installed

Trained on: - GPU - 4 x GeForce RTX 2080 Ti (11GB) - RAM - 256GB - vision lab server 112

Tested on: - GPU - 2 x GeForce RTX 2080 Ti (11GB) - SSD - Samsung MZ7LH3T8 (3.5 TB)

⛳ Setup

Download Docker Image
```
docker pull yoorachoi/beit:dist
```
Download Dataset to [PATH_TO_DATASET]

Execute Docker Container

docker run -it --gpus all --ipc host \
--mount type="bind",source=[PATH_TO_DATASET],target="/dataset" \
yoorachoi/beit:dist /bin/bash

📁 Structure

/unilm
├── backbone
|   └── beit.py
├── configs
|   ├── _base_
|   |    ├── datasets
|   |    |    ├── cubox.py
|   |    |    ├── cubox_test.py
|   |    |    ├── cubox_test_none.py
|   |    |    └── ...
|   |    ├── models
|   |    |    └── upernet_beit_cubox.py
|   |    └── schedules
|   |    |    ├── schedule_160k.py
|   |    |    └── schedule_320k.py
|   ├── beit/upernet
|   |    ├── upernet_beit_base_12_256_slide_160k_ade20k_pt2ft.py
|   |    └── upernet_beit_base_12_256_slide_160k_ade20k_pt2ft_test.py
|   └── test_configs
|        ├── test_all.py
|        ├── test_none.py
|        └── ...
├── mmcv_custom 
└── tools



/dataset
├── images
|   ├── train
|   |   ├── none
|   |   |   ├── class1
|   |   |   |   ├── image1
|   |   |   |   └── image2
|   |   |   |   └── ...
|   |   |   └── class2
|   |   |   └── class3
|   |   |   └── ...
|   |   └── semitransparent
|   |   |   ├── class1
|   |   |   └── class2
|   |   |   └── class3
|   |   |   └── ...
|   |   └── wiredense
|   |   |   ├── class1
|   |   |   └── class2
|   |   |   └── class3
|   |   |   └── ...
|   |   └── wireloose
|   |   |   ├── class1
|   |   |   └── class2
|   |   |   └── class3
|   |   |   └── ...
|   |   └── wiremedium
|   |   |   ├── class1
|   |   |   └── class2
|   |   |   └── class3
|   |   |   └── ...
|   └── validation # same structure as train
|   └── test        # same structure as train
|
└──  seg_map          # same structure as images

📜 Scripts (inside docker container)

1️⃣ Train Script

Available Config files

configs/beit/upernet/upernet_beit_base_12_256_slide_160k_ade20k_pt2ft.py

Command format:

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

Train with total train set & validate with total test set:

bash tools/dist_train.sh \
    configs/beit/upernet/upernet_beit_base_12_256_slide_160k_ade20k_pt2ft.py 4 \
    --work-dir /path/to/save --seed 0  --deterministic \
    --options model.pretrained=https://unilm.blob.core.windows.net/beit/beit_base_patch16_224_pt22k_ft22k.pth

More config files can be found at configs/beit/upernet in original repository.

2️⃣ Evaluation Script