Base info

1、比赛链接

Steel Defect Detection https://www.kaggle.com/c/severstal-steel-defect-detection/overview

2、训练baseline

https://www.kaggle.com/rishabhiitbhu/unet-starter-kernel-pytorch-lb-0-88

3、前向baseline

https://www.kaggle.com/rishabhiitbhu/unet-pytorch-inference-kernel

4、可能会用到的trick

MixMatch https://www.kaggle.com/c/freesound-audio-tagging-2019/discussion/91827

Duplicates & Pattern https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/107053#latest-618971

model ensemble

test time augmentation（horizontal flip）

5、可供参考的discussion

Single Model Segmentation Score https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/109174#latest-628801

train instruction template

deeplabv3

'''bash nohup python -u train.py
--gpu 2 --epoch 21 --arch deeplabv3_resnet50
--downsample 1 --val_batch 32 --train_batch 8
--work_dir /DATA5_DB8/data/yanjia/results/deeplab_20-40
--resume_from /DATA5_DB8/data/yanjia/results/deeplab/model_20.pth > workdir/20190926_1938.log 2>&1 & '''

Unet

nohup python -u train.py
--gpu 1 --epoch 21 --val_batch 32
--work_dir workdir/Unet_resnet18_Adam_baseline_40-60epoch
--resume_from workdir/Unet_resnet18_Adam_baseline_20-40epoch_v1/model_20.pth > workdir/20190926_0952.log 2>&1 &

test intruction template

deeplabv3

python valid.py
--gpu 3 --val_batch 32 --downsample 1
--arch deeplabv3_resnet50 --ckpt_path /DATA5_DB8/data/yanjia/results/deeplabv3_dilation2/model_best.pth

Unet

python valid.py
--val_batch 32 --gpu 0
--ckpt_path workdir/Unet_resnet18_Adam_baseline_20-40epoch_v1/model_best.pth