Algorithm A1. The pseudocodes of model training.

Input: load CNN model name MODEL, mini-batch size N, pre-trained config CONFIG, optimizer function OPTIM, loss function LOSS, data path PATH, training iteration EPOCH

1  model = LoadModel(MODEL)

2  if CONFIG has pre-trained parameters then

3    Load pre-trained parameters

4  if CONFIG need freeze some layers then

5    Set requires_grad = False

6  model.to(‘cuda:0’)

7  if CONFIG has optimizer parameters then

8    Set parameters of optimizer (include Adam or SGD)

9    optimizer = OPTIM(learning_rate=0.1/0.01)

10  if CONFIG need adjust learning rate then

11    Set lr_scheduler to adjust learning rate

12  if CONFIG has more fine-tuning setting then

13    Add other fine-tuning setting (e.g. Batch Normalization)

14  criterion = LOSS()

15  if PATH is valid then

16    Prepare train data loader

17    train_loader = DataLoader(batch_size=N)

18    Prepare validation data loader

19    valid_loader = DataLoader(batch_size=N)

20  for an epoch in EPOCH do

21    for traing data in train_loader do

22     train batch-size training data

23     zero gradients buffers

24     calculate training loss

25     backpropagate the error

26     update weight

27     if log training history then

28      Log accuracy and loss of each epoch in history

29     Test the trained model in validation data set

30     model.eval()

31     for validation data in valid_loader do

32      calculate the best accuracy

33  Save the trained model

34  Save procedure history