Severstal

Scripts for Severstal Kaggle Competition

Results.

1. The script scores 47th on the public leading board
2. The script scores 326th on the private leading board

We suspect the shape or size the defects in steel has changed, but since the labels for the test images are not provided, it is not easy to verify this.

Training phase

1. Train a model in a script under FixCheck branch

   cd Severstal/src

   python main.py --model 'se_resnet50' --sch 2 --loss 2 --output 2 --augment 2 -e 40 --wlovasz 0.2

You will expect the first few epochs result similiar to

Epoch 0 :Train_loss:1.373 Train_dice:0.465 Train_other:0.849 Valid_loss:0.855 Valid_dice:0.782 Valid_other:0.854
Improving val_dice from 0.782 to 0.867, saving the model

Epoch 1 :Train_loss:0.661 Train_dice:0.852 Train_other:0.879 Valid_loss:0.536 Valid_dice:0.867 Valid_other:0.913
Improving val_dice from 0.867 to 0.892, saving the model

Epoch 2 :Train_loss:0.474 Train_dice:0.874 Train_other:0.901 Valid_loss:0.435 Valid_dice:0.892 Valid_other:0.913
Improving val_dice from 0.892 to 0.906, saving the model

Inference Phase

1. Execute the code cd Severstal/src

   python main.py --load_mod --model 'se_resnet50' --sch 2 --loss 2 --output 2 --augment 2 -e 40 --wlovasz 0.2

2. The second command will evaluate the model with postprocessing methods

    Category 1: Mean 0.9749, True Area[Neg,0.0000; Pos,0.0127], Pred Dice[Neg,0.9979; Pos,0.6183], Dice Diff[Neg,4.000; Pos,46.562]
    
    Category 2: Mean 0.9796, True Area[Neg,0.0000; Pos,0.0093], Pred Dice[Neg,1.0000; Pos,0.0000], Dice Diff[Neg,0.000; Pos,41.000]
    
    Category 3: Mean 0.8690, True Area[Neg,0.0000; Pos,0.0634], Pred Dice[Neg,0.9750; Pos,0.7240], Dice Diff[Neg,29.000; Pos,234.348]
    
    Category 4: Mean 0.9817, True Area[Neg,0.0000; Pos,0.0870], Pred Dice[Neg,0.9989; Pos,0.7346], Dice Diff[Neg,2.000; Pos,34.773]
   
   
    Final SWA Dice 0.951
    
    ==============SWA Predict===============
    
    Four labels ratio (1~4): 0.057,0.000,0.416,0.063
    Label Num (Zero labels~four Labels): 999,946,66,0,0
    Label ratio (Zero labels~four Labels): 0.4968,0.4704,0.0328,0.0000,0.0000
    
    ==============True===============
    
    Four labels ratio (1~4): 0.061,0.020,0.422,0.065
    Label Num (Zero labels~four Labels): 941,997,73,0,0