1.Baseline

Library

• Segmentation Models

2. Things worth trying

2-1. Cautions

1. Debugging Mode

   • In order to check if our environment is set correctly
     • Sample data set
     • Train for 1~2 Epochs
     • Check if the loss

2. Seed fix

   • Torch, Numpy, OS Seed Fix

   # seed 고정
   random_seed = 21
   torch.manual_seed(random_seed)
   torch.cuda.manual_seed(random_seed)
   torch.cuda.manual_seed_all(random_seed) # if use multi-GPU
   torch.backends.cudnn.deterministic = True
   torch.backends.cudnn.benchmark = False
   np.random.seed(random_seed)
   random.seed(random_seed)
   os.environ["PYTHONHASHSEED"] = str(random_seed)
   

   • Validation set seed fix

3. Recording results

4. Experimenting one things at a time

   • In order to figure out what led to what

5. Role distribution

2-2. Validation

1. Evaluate model without submitting
2. So that the model does not over-fit the public train set

• Hold Out

• K-fold

• Stratified K-fold