README.md

Pre-train

• Environment Setup

  • The required packages and the corresponding versions please refer to environment.yml.

  • Integrate Adan optimizer into timm:

    1. Please move the adan.py under the python environment package timm/optim/ directory.

    2. Add the following lines to the timm/optim/optim_factory.py file:

       # Import the Adan optimizer
       from .adan import Adan
       ### ...
       
       
       
       # change the create_optimizer_v2 function
       # ....
       elif opt_lower == 'fusednovograd':  
           opt_args.setdefault('betas', (0.95, 0.98))
           optimizer = FusedNovoGrad(parameters, **opt_args)
       elif opt_lower == "adan":                      # New
           optimizer = Adan(parameters, **opt_args)   # New

       ( The details of the Adan optimizer please refer to the official repo. )

• Dataset

  • MillionAID dataset: Please learn the details and download the dataset from MillionAID
  • For pre-training, we utilize the test-set of MillionAID dataset and load the data by PyTorch ImageFolder . So, please make sure all the images are under folder in the input data path (e.g, all the images are under MillionAidDataPath\test\ . The data path is assigned MillionAidDataPath ).

• Distributed training

  • Pre-train ResNet50 on MillionAID dataset with 4 GPU:

    python -m torch.distributed.launch --nproc_per_node=4 
    					main_pretrain.py 
      				-c ./pt_config.yaml \
    					--batch-size 64 \
    					--data-path ${MillionAIDDataPath}  \
    					--eval-data-path None \
    					--workers 8 \
    					--accelerator gpu \
    					--output ${OutPutPath} \
    					--enable_amp True \

  • Pre-train Swin-B on MillionAID dataset with 4 GPU:

    python -m torch.distributed.launch --nproc_per_node=4 
    					main_pretrain.py 
      				-c ./pt_swin_config.yaml \
    					--batch-size 64 \
    					--data-path ${MillionAIDDataPath}  \
    					--eval-data-path None \
    					--workers 8 \
    					--accelerator gpu \
    					--output ${OutputPath} \
    					--enable_amp True \

• Fine-tuning

  • We use the UCM dataset for fine-tuning on classification task. We firstly merge all images together, and then split them to training and test sets, where their information are separately recoded in train.txt and test.txt.

    The form in train.txt is exemplified as

    sparseresidential55.tif 18
    mediumresidential78.tif 12
    ....

    Here, 18 is the category for corresponded image.

    Finally, the construction of the data folder is:

    img
     |-img1.tif
     |-img2.tif
     ...
    train.txt
    test.txt

  • Fine-tune the pre-trained Swin-B on UCM dataset with 1 GPU:

    python main_finetune.py -c F:\\pumpkinCode\\Code\\IndexNetV2\\ft_config.yaml \
     					--batch-size 16 \
     					--data-path ${DataPath}  \
     					--workers 8 \
     					--accelerator gpu \
     					--output ${OutputPath} \
     					--enable_amp True \
     					--infloader True \
     					--checkpoint_path ${CheckpointPath}