how to call custom reward_fn

[dvespasiani]

Hi,
How can I parse any other (custom) reward function other than the default auc/aupr?

I am following this example on colab how to use AMBER. Say i want to implement the spearman correlation as the reward function. From the AMBER reward submodule documentation I'd took these few lines of code

from amber.architect.reward import LossAucReard
import scipy.stats as ss
reward_fn = LossAucReward(method=lambda y_true, y_score: ss.spearmanr(y_true, y_score).correlation)

and substituted 'reward_fn': {'method': 'auc'} with 'reward_fn': {'method': reward_fn(model,val_data)} but this doesnt seem to be the right way of doing it. Do you have any suggestions?
Thanks

[zj-zhang]

For using custom reward functions, e.g. spearman correlation, it should be useful to check out how we implemented it here: https://github.com/vli31/CROTON/blob/main/model_creation/model_search/amber_cnn_sumstats.py#L267

Basically, you would want to parse either a Lambda function, or a custom function that takes y_true and y_score as arguments to
'reward_fn': {'method’: lambda y_true, y_score: f(y_true, y_score)}

Will add this to documentation in the future for better reference.

[dvespasiani]

ok yeah it works perfectly indeed, thank you!