CrowdAI mapping challenge 2018 solution
So far just the command for training the best model First
cd src
Then
# 3 GPU training, ~10 images per GPU
python3 train_weighted.py \
--arch linknet152 --img_size 288 --lr 1e-3 \
--batch-size 30 --workers 8 \
--epochs 100 --start-epoch 0 --epoch_fraction 0.1 \
--m0 5 --m1 5 --m2 20 \
--ths 0.5 --tensorboard True --tensorboard_images True \
--do_augs True --do_more_augs False --aug_prob 0.1 \
--do_energy_levels False --do_boundaries False \
--optimizer adam \
--do_running_mean False --bce_weight 0.5 --dice_weight 0.5 \
--print-freq 20 \
--lognumber ln152_wboth_small_augs_boost_dice \
--do_remove_small_on_borders True \
--do_produce_sizes_mask True --do_produce_distances_mask True \
--w0 5.0 --sigma 10.0 \
Also make sure that your data directory looks like this
data/
├── test_images
├── train
│ └── images
└── val
└── images
```