data/
    ├── DIV2K_train_HR/                 # 训练集高分辨率数据
    └── DIV2K_valid_HR/                 # 测试集高分辨率数据

# 单卡训练
# 用法：bash run_standalone_train_gpu.sh [SCALE] [DATASET_GT_PATH]
bash run_standalone_train_gpu.sh 4 /home/nonroot/IRN/data/DIV2K_train_HR

# 分布式训练
# 用法：bash run_distribute_train_gpu.sh [DEVICE_NUM] [SCALE] [DATASET_PATH]
# 样例：DEVICE_NUM等于2、4、8，分别对应2、4、8卡分布式
bash run_distribute_train_gpu.sh 8 4 /home/nonroot/IRN/data/DIV2K_train_HR

# 单卡评估
# 用法：bash run_eval_gpu.sh [SCALE] [DATASET_PATH] [CHECKPOINT_PATH]
bash run_eval_gpu.sh 4 /home/nonroot/IRN/data/DIV2K_valid_HR /home/nonroot/IRN/ckpt/latest.ckpt

# 单卡训练
# 用法：bash run_standalone_train_ascend.sh [SCALE] [DATASET_GT_PATH]
bash run_standalone_train_ascend.sh 4 /home/nonroot/IRN/data/DIV2K_train_HR

# 8卡分布式训练
# 用法：bash run_distribute_train_ascend.sh [RANK_TABLE_FILE] [SCALE] [DATASET_PATH]
bash run_distribute_train_ascend.sh rank_table_file.json 4 /home/nonroot/IRN/data/DIV2K_train_HR

# 单卡评估
# 用法：bash run_eval_ascend.sh [SCALE] [DATASET_PATH] [CHECKPOINT_PATH]
bash run_eval_ascend.sh 4 /home/nonroot/IRN/data/DIV2K_valid_HR /home/nonroot/IRN/ckpt/latest.ckpt

.
├── README.md                               # 说明文档
├── scripts
│   ├── run_distribute_train_ascend.sh      # Ascend处理器环境多卡训练脚本
│   ├── run_distribute_train_gpu.sh         # GPU处理器环境多卡训练脚本
│   ├── run_eval_ascend.sh                  # Ascend处理器环境评估脚本
│   ├── run_eval_gpu.sh                     # GPU处理器环境评估脚本
│   ├── run_standalone_train_ascend.sh      # Ascend处理器环境单卡训练脚本
│   └── run_standalone_train_gpu.sh         # GPU处理器环境单卡训练脚本
├── src
│   ├── data
│   │   ├── dataset.py                      # 数据集处理
│   │   └── util.py                         # 数据集读取图片缩放等
│   ├── network
│   │   ├── Invnet.py                       # IRN网络定义
│   │   ├── net_with_loss.py                # 自定义loss
│   │   └── util.py                         # 网络初始化等
│   ├── optim
│   │   ├── adam_clip.py                    # 梯度裁剪
│   │   ├── warmup_cosine_annealing_lr.py   # 余弦退火学习率算法
│   │   └── warmup_multisteplr.py           # 多步学习率算法
│   ├── options
│   │   ├── options.py                      # 配置文件读取
│   │   ├── test
│   │   │   ├── test_IRN_x2.yml             # 2倍缩放测试配置文件
│   │   │   └── test_IRN_x4.yml             # 4倍缩放测试配置文件
│   │   └── train
│   │       ├── train_IRN_x2.yml            # 2倍缩放训练配置文件
│   │       └── train_IRN_x4.yml            # 4倍缩放训练配置文件
│   └── utils
│       └── util.py                         # 评价指标计算
├── train.py                                # 训练网络
├── export.py                               # 导出网络
├── requirements.txt                        # 环境需求文件
└── val.py                                  # 测试网络

datasets:
  train:
    name: DIV2K
    mode: LQGT
    batch_size: 8

#### network structures
network_G:
  which_model_G:
      subnet_type: DBNet
  in_nc: 3
  out_nc: 3
  block_num: [8, 8]
  scale: 4
  init: xavier

#### training settings: learning rate scheme, loss
train:
  lr_G: !!float 2e-4
  beta1: 0.9
  beta2: 0.999
  epochs: 5001
  warmup_iter: -1  # no warm up

  lr_scheme: MultiStepLR
  lr_steps: [100000, 200000, 300000, 400000]
  lr_gamma: 0.5


  pixel_criterion_forw: l2
  pixel_criterion_back: l1

  manual_seed: 10

  val_freq: !!float 5e3

  lambda_fit_forw: 16.
  lambda_rec_back: 1
  lambda_ce_forw: 1
  weight_decay_G: !!float 1e-5
  gradient_clipping: 10

# python train.py -h
usage: train.py [--scale {2,4}] [--dataset_GT_path {path of intended GT dataset}] [--dataset_LQ_path {path of intended LQ dataset}] [--resume_state {path of the checkpoint}] [--device_target {Ascend,GPU,CPU}] [--device_num DEVICE_NUM] [--run_distribute RUN_DISTRIBUTE]


IRN for image rescaling.

optional arguments:
  -h, --help            show this help message and exit
  --scale {2,4}
                        Rescaling parameter
  --dataset_GT_path
                        Path of intended GT dataset
  --dataset_LQ_path None
                        Path of intended LQ dataset
  --resume_state None
                        Path of the checkpoint
  --device_target {Ascend,GPU,CPU}
                        Type of device(s) where the model would be deployed
                        to.
  --device_num DEVICE_NUM
                        The number of device(s) to be used for training.
  --run_distribute RUN_DISTRIBUTE
                        Whether to train the model in distributed mode or not.

# python test.py -h
usage: eval.py  [--scale {2,4}] [--dataset_GT_path {path of intended GT dataset}] [--dataset_LQ_path {path of intended LQ dataset}] [--resume_state {path of the checkpoint}] [--device_target {Ascend,GPU,CPU}]

IRN for image rescaling.

optional arguments:
  -h, --help            Show this help message and exit
  --scale {2,4}
                        Rescaling parameter
  --dataset_GT_path
                        Path of intended GT dataset
  --dataset_LQ_path None
                        Path of intended LQ dataset
  --resume_state
                        Path of the checkpoint
  --device_target {Ascend,GPU,CPU}
                        Type of device(s) where the model would be deployed
                        to.

# python export.py -h
usage: export.py [-h] [--scale {2,4}] [--device_id DEVICE_ID] --checkpoint_path
                 CHECKPOINT_PATH [--file_name FILE_NAME]
                 [--file_format {AIR,ONNX,MINDIR}]
                 [--device_target {Ascend,GPU,CPU}]

IRN with AutoAugment export.

optional arguments:
  -h, --help            Show this help message and exit
  --scale {2,4}
                        ResIcaling parameter
  --device_id DEVICE_ID
                        Device id.
  --checkpoint_path CHECKPOINT_PATH
                        Checkpoint file path.
  --file_name FILE_NAME
                        Output file name.
  --file_format {AIR,ONNX,MINDIR}
                        Export format.
  --device_target {Ascend,GPU,CPU}
                        Device target.