loss convergence

[legendhua]

Hi, I arranged my equipment environment follow your code. But the train loss is unchanged, then I add val loss and find the val loss is also keeping unchanged. I changed the learning rate bigger, but still not effect. Have you encountered this problem?

[assassint2017]

In my experiments, at the beginning of the training process, the loss value dropped very quickly and quickly reached about 0.1.

[assassint2017]

Which loss function did you use? Focal loss or Dice loss, i use Dice loss in my experiments

[legendhua]

Focal loss is my choose, I will try the dice loss, and did you apply the train process with focal loss. Thanks for you advice.

[assassint2017]

I haven't had time to try focal loss. The specific implementation of focal loss is quite variable. Hyperparameters also affect performance, and the coefficients that normalize the loss function are also worthy of attention. In my code, I use the same strategy as the original paper. But for different problems, in order to get good performance you may need to try other different strategies. In short, using focal loss and achieving good results is not an easy task, and that is why I prefer Dice loss.

[legendhua]

Thanks for you detailed reply, 非常感谢！！

[assassint2017]

~=o(^▽^)o~♪

[John1231983]

@assassint2017 : What is your rank in leaderboard? Thanks