MINet

$ tree -L 3
.
├── backbone
│   ├── __init__.py
│   ├── origin
│   │   ├── from_origin.py
│   │   ├── __init__.py
│   │   ├── resnet.py
│   │   └── vgg.py
│   └── wsgn
│       ├── customized_func.py
│       ├── from_wsgn.py
│       ├── __init__.py
│       ├── resnet.py
│       └── resnext.py
├── config.py
├── LICENSE
├── loss
│   ├── CEL.py
│   └── __init__.py
├── main.py
├── module
│   ├── BaseBlocks.py
│   ├── __init__.py
│   ├── MyLightModule.py
│   ├── MyModule.py
│   └── WSGNLightModule.py
├── network
│   ├── __init__.py
│   ├── LightMINet.py
│   ├── MINet.py
│   ├── PureWSGNLightMINet.py
│   └── WSGNLightMINet.py
├── output (These are the files generated when I ran the code.)
│   ├── CPLightMINet_Res50_S352_BS32_E20_WE1_AMPy_LR0.05_LTf3sche_OTf3trick_ALy_BIy_MSy
│   │   ├── cfg_2020-07-23.txt
│   │   ├── pre
│   │   ├── pth
│   │   ├── tb
│   │   ├── te_2020-07-23.txt
│   │   ├── tr_2020-07-23.txt
│   │   └── trainer_2020-07-23.txt
│   └── result.xlsx
├── pyproject.toml
├── readme.md
└── utils
    ├── cal_fps.py
    ├── dataloader.py
    ├── __init__.py
    ├── joint_transforms.py
    ├── metric.py
    ├── misc.py
    ├── pipeline_ops.py
    ├── recorder.py
    ├── solver.py
    └── tensor_ops.py

# CPLightMINet_Res50_S320_BS4_E50_WE1_AMPn_LR0.001_LTpoly_OTf3trick_ALy_BIy_MSn
arg_config = {
    "model": "CPLightMINet_Res50",  # 实际使用的模型，需要在`network/__init__.py`中导入
    "info": "",  # 关于本次实验的额外信息说明，这个会附加到本次试验的exp_name的结尾，如果为空，则不会附加内容。
    "use_amp": False,  # 是否使用amp加速训练
    "resume_mode": "",  # the mode for resume parameters: ['train', 'test', '']
    "use_aux_loss": True,  # 是否使用辅助损失
    "save_pre": False,  # 是否保留最终的预测结果
    "epoch_num": 50,  # 训练周期, 0: directly test model
    "lr": 0.001,  # 微调时缩小100倍
    "xlsx_name": "result.xlsx",  # the name of the record file
    # 数据集设置
    "rgb_data": {
        "tr_data_path": dutstr_path,
        "te_data_list": OrderedDict(
            {
                "pascal-s": pascals_path,
                "ecssd": ecssd_path,
                # "hku-is": hkuis_path,
                # "duts": dutste_path,
                # "dut-omron": dutomron_path,
                # "soc": soc_path,
            },
        ),
    },
    # 训练过程中的监控信息
    "tb_update": 10,  # >0 则使用tensorboard
    "print_freq": 10,  # >0, 保存迭代过程中的信息
    # img_prefix, gt_prefix，用在使用索引文件的时候的对应的扩展名
    "prefix": (".jpg", ".png"),
    # if you dont use the multi-scale training, you can set 'size_list': None
    # "size_list": [224, 256, 288, 320, 352],
    "size_list": None,  # 不使用多尺度训练
    "reduction": "mean",  # 损失处理的方式，可选“mean”和“sum”
    # 优化器与学习率衰减
    "optim": "f3_trick",  # 自定义部分的学习率
    "weight_decay": 5e-4,  # 微调时设置为0.0001
    "momentum": 0.9,
    "nesterov": False,
    "sche_usebatch": False,
    "lr_type": "poly",
    "warmup_epoch": 1,  # depond on the special lr_type, only lr_type has 'warmup', when set it to 1, it means no warmup.
    "lr_decay": 0.9,  # poly
    "use_bigt": True,  # 训练时是否对真值二值化（阈值为0.5）
    "batch_size": 4,  # 要是继续训练, 最好使用相同的batchsize
    "num_workers": 4,  # 不要太大, 不然运行多个程序同时训练的时候, 会造成数据读入速度受影响
    "input_size": 320,
}

F3Net is the most recent work on SOD, and the performance is very good. I think its training strategy is of great reference value. Here, I have changed our training method by learning from its code.

To explore the upper limit of the performance of the model, I tried some ways to improve the performance of the model on a NVIDIA GTX 1080Ti (~11G).

• To achieve a larger batch size:
  • we reduce the number of intermediate channels in AIMs;
  • we apply the checkpoint feature of PyTroch;
    • https://blog.csdn.net/one_six_mix/article/details/93937091
    • https://pytorch.org/docs/stable/checkpoint.html#torch.utils.checkpoint.checkpoint
  • we set the batch size to 32.
• Use more effective training strategies:
  • We apply multi-scale training strategy (borrowed from the code of F3Net).
  • Network parameters are trained in groups with different learning rates (borrowed from the code of F3Net).
  • Multiple learning rate decay strategies with/without the warmup technology.

Results:

• D: DUTS
• DO: DUT-OMRON
• HI: HKU-IS
• E: ECSSD
• PS: PASCAL-S
• MT: Multi-scale Training
• EP: Epoch Number

NOTE: The results here are for reference only. Note that the results here are all tested on the complete test datasets. In fact, some of the results here can be higher if testing in the way of the existing papers. Because the test set in my paper follows the settings of the existing papers, some datasets, such as HKU-IS, are not tested with the complete dataset.

注： 此处结果仅供参考。请注意，这里的结果都是在完整的测试数据集上测试的。事实上，如果按照现有论文的方式进行测试，这里的一些结果可能会更高。由于本文中的测试集遵循现有论文的设置，一些数据集，如HKU-IS，没有使用完整的数据集进行测试。