Ba/r_nse_loss #179
Ba/r_nse_loss #179
Conversation
|
I was thinking that maybe we should just called them |
| l_init_train, | ||
| l_init_val, | ||
| training_loss, | ||
| loss_types[1], |
There was a problem hiding this comment.
I don't know how I feel about this. I thinks this is already breaking :/
There was a problem hiding this comment.
I think it is not breaking - it is just for the trainboard. The default behavior stays the same (training_loss = :mse, loss_types = [:mse, :r2])
| # one minus nse | ||
| function loss_fn(ŷ, y, y_nan, ::Val{:nse}) | ||
| return sum((ŷ[y_nan] .- y[y_nan]).^2) / sum((y[y_nan] .- mean(y[y_nan])).^2) | ||
| return one(eltype(ŷ)) - (sum((ŷ[y_nan] .- y[y_nan]).^2) / sum((y[y_nan] .- mean(y[y_nan])).^2)) |
There was a problem hiding this comment.
oh yes, it should be like this for a proper loss.
| return one(eltype(ŷ)) .- (cor(ŷ[y_nan], y[y_nan])) | ||
| end | ||
|
|
||
| function loss_fn(ŷ, y, y_nan, ::Val{:nseLoss}) |
There was a problem hiding this comment.
maybe this is only ::Val{:nse} and the next one is the actual :Val{:nseLoss}.
Yes, performance_metrics would be good. This would be indeed breaking - maybe my proposed changes are also breaking. We could have performance_metrics where the highest value gives the best performance (NSE) and what we have so far returns the model with lowest metric. So at the moment the train function would only work with losses (decreasing) |
make losses losses