about class weights

[clockwiser]

In the source, class weights are used as below.

def get_model( model_file, model_name, loss_file, loss_name, class_weight, n_encdec, n_classes, in_channel, n_mid, train_depth=None, result_dir=None): model = imp.load_source(model_name, model_file) model = getattr(model, model_name) loss = imp.load_source(loss_name, loss_file) loss = getattr(loss, loss_name)

# Initialize
model = model(n_encdec, n_classes, in_channel, n_mid)
if train_depth:

model = loss(model, class_weight, train_depth)

Strangely, there is no example of setting loss and loss_name. Without proper parameters, class weights should not work.