MINet

Folders & Files

• backbone: Store some code for backbone networks.
• loss: The code of the loss function.
• module: The code of important modules.
• network: The code of the network.
• output: It saves all results.
• utils: Some instrumental code.
  • dataloader.py: About creating the dataloader.
  • …
• main.py: I think you can understand.
• config.py: Configuration file for model training and testing.

My Environment

Latest Env Info

The yaml file exported by my latest work environment is code/ minet.yaml. You can refer to the version information of each package in it.

For Apex:

```shell script $ # For performance and full functionality, we recommend installing Apex with CUDA and C++ extensions via $ git clone https://github.com/NVIDIA/apex $ cd apex $ pip install -v –no-cache-dir –global-option=”–cpp_ext” –global-option=”–cuda_ext” ./

## Note

<details>
<summary>Configraturn Demo</summary>

```python
# 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,
}

</details>

• (2020-07-23 update) There are some important configuration items in the configuration system:
  • model in config.py: It is the name of the model that will be used. (Note: you must import the model into the ./network/__init__.py in advance.)
    • For example, if we can set model to MINet_VGG16, we will use the model MINet_VGG16 from the folder network.
  • info in config.py: An additional string used to illustrate this experiment.
  • exp_name in main.py: The complete string used to describe this experiment, which is obtained by the function function construct_exp_name in utils/misc.py by parsing config.py. For more information, please see utils/ misc.py.
    • Based on the demo configuration, the exp_name generated by our code is CPLightMINet_Res50_S320_BS4_E50_WE1_AMPn_LR0.001_LTpoly_OTf3trick_ALy_BIy_MSn.
    • The corresponding folder in output, where all files generated during the running of the model will be saved, will have the name CPLightMINet_Res50_S320_BS4_E50_WE1_AMPn_LR0.001_LTpoly_OTf3trick_ALy_BIy_MSn.
  • resume_mode: It indicates whether you want to use the existing parameter files (‘XXX.pth’) stored in the pth folder in the model folder corresponding to exp_name.
    • If you set it to test, and your exp_name is CPLightMINet_Res50_S320_BS4_E50_WE1_AMPn_LR0.001_LTpoly_OTf3trick_ALy_BIy_MSn, the .pth file should be placed in ./output/CPLightMINet_Res50_S320_BS4_E50_WE1_AMPn_LR0.001_LTpoly_OTf3trick_ALy_BIy_MSn (Please be careful to change the name of the .pth file to state_final.pth if you use our pretrained parameter file.).
    • If your training process is interrupted, you can set it to train, and it will automatically load the existing checkpoint_final.pth.tar file from the pth folder to resume the training process.
    • In the case of training the model from scratch (that is, start_epoch=0), you just need to set it to "".

Train

1. You can customize the value of the arg_config dictionary in the configuration file.
   • The first time you use it, you need to adjust the path of every dataset.
   • Set model to the model name that has been imported in network/__init__.py.
   • Modify info according to your needs, which will be appended to the final exp_name. (The function construct_exp_name in utils/misc.py will generate the final exp_name.)
   • Set the item resume_mode to "".
   • And other setting in config.py, like epoch_num, lr and so on…
2. In the folder code, run the command python main.py.
3. Everything is OK. Just wait for the results.
4. The test will be performed automatically when the training is completed.
5. All results will be saved into the folder output, including predictions in folder pre (if you set save_pre to True), .pth files in folder pth and other log files.

If you want to test the trained model again…

Our pre-training parameters can also be used in this way.

1. In the output folder, please ensure that there is a folder corresponding to the model (See Note), which contains the pth folder, and the .pth file of the model is located here and its name is state_final.pth.
2. Set the value of model of arg_config to the model you want to test.
3. Set the value of te_data_list to your dataset path.
4. Set the value of resume_mode to test.
5. In the folder code, run python main.py.
6. You can find predictions from the model in the folder pre of the output.

If you want to inference the trained model on your own dataset…

• Configure the arguments in the script tools/inference.py
• Run it.

Evaluation

• We evaluate results of all models by lartpang/sal_eval_toolbox.
  • lartpang/sal_eval_toolbox is based on ArcherFMY/sal_eval_toolbox. But, we add the code about E-measure and weighted F-measure and update the related code in our forked repository lartpang/sal_eval_toolbox. Welcome to use it in your code :star:!
• 2021-06-06: A better evaluation toolbox for RGB/RGB-D/Co-RGB SOD https://github.com/lartpang/PySODEvalToolkit

More

If there are other issues, you can create a new issue.