Tabular Data Weighted Custom Loss Function Idea

[darek.kleczek]

This worked for me:

How to use class weights in loss function for imbalanced dataset

from torch import nn weights = [0.4, 1] class_weights=torch.FloatTensor(weights).cuda() learn.crit = nn.CrossEntropyLoss(weight=class_weights)