.
└──data
    ├── FaceImageCroppedWithAlignment               # 训练数据集 MS-Celeb-1M
    │   ├── m.0_0zl
    │   ├── m.0_0zy
    │   ├── m.01_06j
    │   ├── m.0107_f
    │   ...
    │
    ├── lfw                                         # 测试数据集 LFW
    │   ├── image
    │   │   ├── Aaron_Eckhart
    │   │   ├── Aaron_Guiel
    │   │   ├── Aaron_Patterson
    │   │   ├── Aaron_Peirsol
    │   │   ├── Aaron_Pena
    │   │   ...
    │   │
    │   ├── image_list_for_blufr.txt                # lfw BLUFR protocols 测试集列表，需用户生成，方法见上文
    │   └── image_list_for_lfw.txt                  # lfw 6,000 pairs 测试集列表，需用户生成，方法见上文
    │
    └── MS-Celeb-1M_clean_list.txt                  # 清洗后的训练集列表

from easydict import EasyDict as edict
lightcnn_cfg = edict({
    # training setting
    'network_type': 'LightCNN_9Layers',
    'epochs': 80,
    'lr': 0.01,
    'num_classes': 79077,
    'momentum': 0.9,
    'weight_decay': 1e-4,
    'batch_size': 128,
    'image_size': 128,
    'save_checkpoint_steps': 60000,
    'keep_checkpoint_max': 40,
    # train data location
    'data_path': '/data/MS-Celeb-1M/FaceImageCroppedWithAlignment', # 绝对路径（需要修改）
    'train_list': '/data/MS-Celeb-1M_clean_list.txt',               # 绝对路径（需要修改）
    # test data location
    'root_path': '/data/lfw/image',                                 # 绝对路径（需要修改）
    'lfw_img_list': 'image_list_for_lfw.txt',                       # 文件名
    'lfw_pairs_mat_path': 'mat_files/lfw_pairs.mat',                # 运行测试脚本位置的相对路径
    'blufr_img_list': 'image_list_for_blufr.txt',                   # 文件名
    'blufr_config_mat_path': 'mat_files/blufr_lfw_config.mat'       # 运行测试脚本位置的相对路径
})

# 进入根目录
cd LightCNN/

# 运行单卡训练
# DEVICE_ID: Ascend处理器的id，需用户指定
bash scripts/train_standalone.sh DEVICE_ID

cd LightCNN/scripts

# 运行2卡或4卡训练
# hccl.json: Ascend配置信息，需用户自行配置，与八卡不同，详见官网教程
# DEVICE_NUM应与train_distribute.sh中修改device_ids的长度相同
# 需进入train_distribute.sh 修改device_ids=(id1 id2) 或 device_ids=(id1 id2 id3 id4)
bash train_distribute.sh hccl.json DEVICE_NUM

# 运行8卡训练
# hccl.json: Ascend配置信息，需用户自行配置
bash train_distribute_8p.sh hccl.json

# 进入根目录
cd LightCNN/

# 评估LightCNN在lfw 6,000 pairs上的表现
# DEVICE_ID: Ascend处理器id
# CKPT_FILE: checkpoint权重文件
bash scripts/eval_lfw.sh DEVICE_ID CKPT_FILE

# 评估LightCNN在lfw BLUFR protocols上的表现
# DEVICE_ID: Ascend处理器id
# CKPT_FILE: checkpoint权重文件
bash scripts/eval_blufr.sh DEVICE_ID CKPT_FILE

.
├── README_CN.md
├── ascend310_infer
├── mat_files
│   ├── blufr_lfw_config.mat                        # lfw BLUFR protocols测试配置文件
│   └── lfw_pairs.mat                               # lfw 6,000 pairs测试配置文件
├── scripts
│   ├── eval_blufr.sh                               # lfw BLUFR protocols测试脚本
│   ├── eval_lfw.sh                                 # lfw 6,000 pairs测试脚本
│   ├── convert.sh                                  # 训练数据集格式转换脚本
│   ├── run_infer_310.sh                            # 启动Ascend 310 推理shell脚本
│   ├── train_distribute_8p.sh                      # 8卡并行训练脚本
│   ├── train_distribute.sh                         # 多卡（2卡/4卡）并行训练脚本
│   └── train_standalone.sh                         # 单卡训练脚本
├── src
│   ├── config.py                                   # 训练参数配置文件
│   ├── convert.py                                  # 训练数据集转换脚本
│   ├── dataset.py                                  # 加载训练数据集
│   ├── get_list.py                                 # 获取测试集列表
│   ├── lightcnn.py                                 # LightCNN模型文件
│   └── lr_generator.py                             # 动态学习率生成脚本
│
├── eval_blufr.py                                   # lfw BLUFR protocols测试脚本
├── eval_lfw.py                                     # lfw 6,000 pairs测试脚本
├── export.py                                       # 模型转换脚本
├── postprocess.py                                  # 310推理后处理及测试脚本
├── preprocess.py                                   # 310推理数据预处理脚本
└── train.py                                        # 训练脚本

'network_type': 'LightCNN_9Layers',                 # 模型名称
'epochs': 80,                                       # 总训练epoch数
'lr': 0.01,                                         # 训练学习率
'num_classes': 79077,                               # 分类总类别数量
'momentum': 0.9,                                    # 动量
'weight_decay': 1e-4,                               # 权重衰减
'batch_size': 128,                                  # batch size
'image_size': 128,                                  # 输入模型的图像尺寸
'save_checkpoint_steps': 60000,                     # 保存checkpoint的间隔step数
'keep_checkpoint_max': 40,                          # 只保存最后一个keep_checkpoint_max检查点

# trian_standalone.sh
python3 train.py \
          --device_target Ascend \
          --device_id "$DEVICE_ID" \
          --ckpt_path ./ckpt_files > train_standalone_log.log 2>&1 &

# train_distribute_8p.sh
for ((i = 0; i < ${DEVICE_NUM}; i++)); do
    export DEVICE_ID=$i
    export RANK_ID=$i
    rm -rf ./train_parallel$i
    mkdir ./train_parallel$i
    cp ../*.py ./train_parallel$i
    cp *.sh ./train_parallel$i
    cp -r ../src ./train_parallel$i
    cd ./train_parallel$i || exit
    echo "start training for rank $RANK_ID, device $DEVICE_ID"
    env >env.log
    python3 train.py \
        --device_target Ascend \
        --device_id "$DEVICE_ID" \
        --run_distribute 1 \
        --ckpt_path ./ckpt_files > train_distribute_8p.log 2>&1 &
    cd ..
done

# train_distribute.sh

# distributed devices id
device_ids=(0 1 2 3)

for ((i = 0; i < ${DEVICE_NUM}; i++)); do
    export DEVICE_ID=${device_ids[i]}
    export RANK_ID=$i
    rm -rf ./train_parallel$i
    mkdir ./train_parallel$i
    cp ../*.py ./train_parallel$i
    cp *.sh ./train_parallel$i
    cp -r ../src ./train_parallel$i
    cd ./train_parallel$i || exit
    echo "start training for rank $RANK_ID, device $DEVICE_ID"
    env >env.log
    python3 train.py \
        --device_target Ascend \
        --device_id $DEVICE_ID \
        --run_distribute 1 \
        --ckpt_path ./ckpt_files > train_distribute.log 2>&1 &
  cd ..
done

# 单卡训练结果
epoch: 1 step: 39451, loss is 4.6629214
epoch time: 4850141.061 ms, per step time: 122.941 ms
epoch: 2 step: 39451, loss is 3.6382508
epoch time: 4148247.801 ms, per step time: 105.149 ms
epoch: 3 step: 39451, loss is 2.9592063
epoch time: 4146129.041 ms, per step time: 105.096 ms
epoch: 4 step: 39451, loss is 3.6300964
epoch time: 4128986.449 ms, per step time: 104.661 ms
epoch: 5 step: 39451, loss is 2.9682
epoch time: 4117678.376 ms, per step time: 104.374 ms
epoch: 6 step: 39451, loss is 3.2115498
epoch time: 4139044.713 ms, per step time: 104.916 ms
...

# 分布式训练结果（8P）
epoch: 1 step: 4931, loss is 8.716646
epoch time: 1215603.837 ms, per step time: 246.523 ms
epoch: 2 step: 4931, loss is 3.6822505
epoch time: 1038280.276 ms, per step time: 210.562 ms
epoch: 3 step: 4931, loss is 1.8040423
epoch time: 1033455.542 ms, per step time: 209.583 ms
epoch: 4 step: 4931, loss is 1.6634097
epoch time: 1047134.763 ms, per step time: 212.357 ms
epoch: 5 step: 4931, loss is 1.369437
epoch time: 1053151.674 ms, per step time: 213.578 ms
epoch: 6 step: 4931, loss is 1.3599608
epoch time: 1064338.712 ms, per step time: 215.846 ms
...

# 进入根目录
cd LightCNN/

# 评估LightCNN在lfw 6,000 pairs上的表现
# DEVICE_ID: Ascend处理器id
# CKPT_FILE: checkpoint权重文件
bash scripts/eval_lfw.sh DEVICE_ID CKPT_FILE

# 评估LightCNN在lfw BLUFR protocols上的表现
# DEVICE_ID: Ascend处理器id
# CKPT_FILE: checkpoint权重文件
bash scripts/eval_blufr.sh DEVICE_ID CKPT_FILE

# eval_lfw.sh
# ${DEVICE_ID}: Ascend处理器id
# ${ckpt_file}: checkpoint权重文件，由用户输入
# eval_lfw.log：保存的测试结果
python3 eval_lfw.py \
            --device_target Ascend \
            --device_id "${DEVICE_ID}" \
            --resume "${ckpt_file}" > eval_lfw.log 2>&1 &

# eval_blufr.sh
# ${DEVICE_ID}: Ascend处理器id
# ${ckpt_file}: checkpoint权重文件，由用户输入
# eval_blufr.log：保存的测试结果
# Tips：在eval_blufr.py中，可以使用numba库加速计算。如果引入了numba库，可以用'@jit'语法糖进行加速，去掉注释即可
python3 eval_blfur.py \
          --device_target Ascend \
          --device_id "${DEVICE_ID}" \
          --resume "${ckpt_file}" > eval_blufr.log 2>&1 &

python export.py --ckpt_file [CKPT_PATH] --file_format MINDIR

# Ascend310 inference
bash run_infer_310.sh [MINDIR_PATH] [DATASET_PATH] [DEVICE_ID]