# 持續學習操作手冊

## Environment setup

* **MMDetection official github:** https://github.com/open-mmlab/mmdetection
* **SoftTeacher official github:** https://github.com/microsoft/SoftTeacher

1. [Download mmdetection by official guide](https://github.com/open-mmlab/mmdetection/blob/master/docs/en/get_started.md/#Installation)
2. Download SoftTeacher

    git clone https://github.com/microsoft/SoftTeacher
    make install

3. 將configs_rail放進SofTeacher資料夾底下
## Training
**Fully-supervised**

1. 打開 **faster_rcnn_fully_supervised_rail.py**
2. 將 **ann_file** 路徑改為json檔路徑
3. 將 **img_prefix** 改為圖資資料夾路徑

    data = dict(
        samples_per_gpu=5,
        workers_per_gpu=4,
        train=dict(
            ann_file="/home/wayne/charley/SoftTeacher/dataset/railway/L_100/train/_annotations.coco.json",
            img_prefix="/home/wayne/charley/SoftTeacher/dataset/railway/L_100/train/",
            classes=classes
        ),
        val=dict(
            ann_file='/home/wayne/charley/SoftTeacher/dataset/railway/L_100/val/_annotations.coco.json',
            img_prefix='/home/wayne/charley/SoftTeacher/dataset/railway/L_100/val/',
            classes=classes
        ),
        test=dict(
            ann_file='/home/wayne/charley/SoftTeacher/dataset/railway/L_100/val/_annotations.coco.json',
            img_prefix='/home/wayne/charley/SoftTeacher/dataset/railway/L_100/val/',
            classes=classes
        )
    )

4. 執行 **dist_train.sh**

    bash tools/dist_train.sh configs_rail/soft_teacher/faster_rcnn_fully_supervised_rail.py <NUM_GPUS>

**Semi-supervised**

1. 打開 faster_rcnn_semi_supervised_rail.py
2. 路徑設定和fully-supervised相同
3. sup裡放置labeled的圖資
4. unsup裡放置unlabeled的圖資

    data = dict(
        samples_per_gpu=5,
        workers_per_gpu=4,
        train=dict(
            sup=dict(
                type="CocoDataset",
                ann_file="/home/wayne/charley/SoftTeacher/dataset/railway/L_100/train/_annotations.coco.json",
                img_prefix="/home/wayne/charley/SoftTeacher/dataset/railway/L_100/train/",
                classes=classes
            ),
            unsup=dict(
                type="CocoDataset",
                ann_file="/home/wayne/charley/SoftTeacher/dataset/railway/L_10_UL_90_plus300/train/ann_unlabel.json",
                img_prefix="/home/wayne/charley/SoftTeacher/dataset/railway/L_10_UL_90_plus300/train/",
                classes=classes
            ),
        ),
        val=dict(
            ann_file='/home/wayne/charley/SoftTeacher/dataset/railway/L_100/val/_annotations.coco.json',
            img_prefix='/home/wayne/charley/SoftTeacher/dataset/railway/L_100/val/',
            classes=classes
        ),
        test=dict(
            ann_file='/home/wayne/charley/SoftTeacher/dataset/railway/L_100/val/_annotations.coco.json',
            img_prefix='/home/wayne/charley/SoftTeacher/dataset/railway/L_100/val/',
            classes=classes
        ),
        sampler=dict(
            train=dict(
                sample_ratio=[1, 4],
            )
        ),
    )

5. 將 **dist_train_partially.sh** 改為以下程式碼

    #!/usr/bin/env bash
    set -x

    TYPE=$1
    FOLD=$2
    PERCENT=$3
    GPUS=$4
    PORT=${PORT:-29570}


    PYTHONPATH="$(dirname $0)/..":$PYTHONPATH

    if [[ ${TYPE} == 'baseline' ]]; then
        python -m torch.distributed.launch --nproc_per_node=$GPUS --master_port=$PORT \
            $(dirname "$0")/train.py configs/baseline/faster_rcnn_r50_caffe_fpn_coco_partial_180k.py --launcher pytorch \
            --cfg-options fold=${FOLD} percent=${PERCENT} ${@:5}
    else
        python -m torch.distributed.launch --nproc_per_node=$GPUS --master_port=$PORT \
            $(dirname "$0")/train.py configs_rail/soft_teacher/faster_rcnn_semi_supervised_rail.py --launcher pytorch \
            --cfg-options fold=${FOLD} percent=${PERCENT} ${@:5}
    fi

6. 執行 **dist_train_partially.sh**

    bash tools/dist_train_partially.sh semi 1 10 <NUM_GPUS>

## Inference
1. Image -> 使用demo/image_demo.py
2. Video -> 使用demo/video_demo.py

## 如何產生unlabeled data使用的json檔

1. 使用 **dataset/railway** 資料夾底下之 **img2coco_unlabel.py**
2. 將 **root** 改為放置unlabeled data的資料夾，即可產生訓練需要的json檔

    import os
    import json
    from PIL import Image

    img_list = []
    root = '<unlabeled data路徑>'
    total = os.listdir(root)
    #print(total)
    print('total img:', len(total))
    id=0
    for file in total:
        #print('#####')
        file_path = os.path.join(root,file)
        #print("file: ", file, ' , fold_path:', file_path)
        img = Image.open(file_path)
        w = img.width
        h = img.height
        #print(w,h)
        #id = os.path.splitext(file)[0]
        img_data = {
            'file_name': file,
            'height': h,
            'width': w,
            'id': int(id)
        }
        id+=1
        #print(img_data)
        img_list.append(img_data)

    all_dat = {
        'images': img_list,
        'annotations': [],
        'categories': []
    }
    #print('all_dat:', all_dat)

    with open('ann_unlabel.json', 'w') as f:
        json.dump(all_dat, f)