doubaodexiaomuwu / xview2unet Goto Github PK

Untar train.tar and tier3.tar and test.tar from https://xview2.org/dataset

Arrange data as follows:

data
├── test
│   └── images
└── train
    ├── images1024
    ├── labels1024
    └── targets1024

For fp16 training using Volta or Turing GPU install Nvidia apex python only version from https://github.com/NVIDIA/apex . (Note the automatically applied dynamic loss scaling feature may help with stability of training)

You may need to modify batch sizes in trainlocunet.py and traindamgeunet.py

pip install -r requirements.txt 
python preprocess.py
python trainlocunet.py
python traindamageunet.py
tree workspace # to see your checkpoints and tensorboard logs 
python testdamage.py

Results will be in

results
├── predictions          # pixel values in range (0-1) or (0-4) valid for submission (zip this folder for submission) 
└── vizpredictions       # pixel values in range (0-255) for easy viewing

Just missed out on top 50 leaderboard despite joining the competition very late and entering submissions on last day only

(weighted overall, loc, dmg) .68 / .78 / .63

Feel free to experiment with the code and post issues.