michal2409 / xview2 Goto Github PK

/data
 ├── train
 │      ├── images
 │      │      └── <image_id>.png
 │      │      └── ...
 │      └── targets
 │             └── <image_id>.png
 │             └── ...
 ├── test
 │      ├── images
 │      │      └── <image_id>.png
 │      │      └── ...
 │      └── targets
 │             └── <image_id>.png
 │             └── ...
 └── holdout
        ├── images
        │      └── <image_id>.png
        │      └── ...
        └── targets
               └── <image_id>.png
               └── ...

python utils/convert2png.py --data DATA_PATH/train

docker run -it --rm --gpus all --shm-size=8g --ulimit memlock=-1 --ulimit stack=67108864 -v RESULTS_PATH:/results -v DATA_PATH:/data xview2 bash

usage: python main.py [--optional arguments] 

optional arguments:
  -h, --help            show this help message and exit
  --exec_mode {train,eval}
                        Execution mode of main script
  --data                Path to the data directory
  --results             Path to the results directory
  --gpus                Number of gpus to use
  --num_workers         Number of subprocesses to use for data loading
  --batch_size          Training batch size
  --val_batch_size      Evaluation batch size
  --precision {16,32}   Numerical precision
  --epochs              Max number of epochs      
  --patience            Early stopping patience
  --ckpt                Path to pretrained checkpoint
  --logname             Name of logging file
  --ckpt_pre            Path to pretrained checkpoint of localization model used to initialize network for damage assesment
  --type {pre,post}     Type of task to run; pre - localization, post - damage assesment
  --seed
  --optimizer {sgd,adam,adamw,radam,adabelief,adabound,adamp,novograd}
  --dmg_model {siamese,siameseEnc,fused,fusedEnc,parallel,parallelEnc,diff,cat}
                        U-Net variant for damage assessment task
  --encoder {resnest50,resnest101,resnest200,resnest269,resnet50,resnet101,resnet152}
                        U-Net encoder
  --loss_str            String used for creation of loss function, e.g focal+dice creates the loss function as sum of focal and dice.
                        Available functions: dice, focal, ce, ohem, mse, coral
  --use_scheduler       Enable Noam learning rate scheduler
  --warmup              Warmup epochs for Noam learning rate scheduler
  --init_lr             Initial learning rate for Noam scheduler
  --final_lr            Final learning rate for Noam scheduler
  --lr                  Learning rate, or a target learning rate for Noam scheduler
  --weight_decay        Weight decay (L2 penalty)
  --momentum            Momentum for SGD optimizer
  --dilation {1,2,4}    Dilation rate for a encoder, e.g dilation=2 uses dilation instead of stride in the last encoder block
  --tta                 Enable test time augmentation
  --ppm                 Use pyramid pooling module
  --aspp                Use atrous spatial pyramid pooling
  --no_skip             Disable skip connections in UNet
  --deep_supervision    Enable deep supervision
  --attention           Enable attention module at the decoder
  --autoaugment         Use imageNet autoaugment pipeline
  --interpolate         Interpolate feature map from encoder without a decoder
  --dec_interp          Use interpolation instead of transposed convolution in a decoder

python main.py --type pre --encoder resnest200 --loss_str ce+dice --deep_supervision --attention --tta --gpus 1 --batch_size 16 --val_batch_size 8 --gpus 1

python main.py --type post --dmg_model siamese --encoder resnest200 --loss_str focal+dice --attention --deep_supervision --tta --gpus 8 --batch_size 16 --val_batch_size 8 

python main.py --exec_mode eval --type {pre,post} --ckpt <path/to/checkpoint> --gpus 1 --val_batch_size 8

python utils/post_process.py

python utils/xview2_metrics.py /results/predictions /results/targets /results/score.json && cat /results/score.json