The ~INT_BAYESIAN~ flag does not seem to work

[cerebrai]

Hi,

I have a custom loss function that has a hyper-parameter that only takes integer values. I used the following command

--loss_funcs~OVERRIDE~ {metric_loss: {MyCustomLoss: {t_k~INT_BAYESIAN~: [0, 10]}}},

however, the first bayesian iteration sets a float value of around 8.3 for t_k. What am I missing here?

Note: As a workaround, I have type-casted t_k to be int in my code but I am afraid it might affect the optimization process.

[KevinMusgrave]

I'm not sure why that is happening. However, I think your workaround might be equivalent, according to this issue: facebook/Ax#133