3rd place solution for 国立国会図書館の画像データレイアウト認識

My machine

• Intel(R) Core(TM) i7-6850K CPU @ 3.60GHz
• CUDA 10.1, python3.6.9, pytorch 1.3.1
• 2 1080ti

Step 0: Set up environment

1. install anaconda
2. conda create -n fuxian python=3.6.9
3. conda activate fuxian; cd code
4. pip install -r requirements.txt
5. python setup.py develop
6. put all data under the directory data/, and unzip them

Step 1: Prepare dataset

 python tools/library/prepare_testdataset_split.py

Step 2: Inference

1. bash predict.sh
2. python tools/library/merge_bbox_gudian.py
3. python tools/library/merge_bbox_jindai.py 
4. python tools/library/prepare_submit_split_sort.py

• you can find the final json in data/final_submit.json

Step 3: Retrain

bash data/download.sh
python tools/library/prepare_dataset_split.py
bash train.sh

• because I use coco-pretrained weights, so we have to donwload them
• the training configs use 2 gpu, if you have only 1, you have to change the learning rate to half manually. when training is finished, you can find weight in data/retrained/models/