How to configure a list of complex objects?

[lukasbm]

Hey everyone,
thanks for the great library, i've been using it for the past few weeks.
It has quite a steep learning curve and theres things i don't quite understand yet.
For example i have a store which configures my loss configs:

loss_store = store(group="criterion")
loss_store(InversionLoss, name="inversion")
loss_store(TripletLoss, name="triplet")
loss_store(CosineSimilarityLoss, name="cosine")

which works great so far.
However I have a more complex loss like this, which is composed of other losses:

class CombinedLoss(Loss):
    def __init__(self, losses: list[Loss], weights: list[float]):
        super().__init__()
        self.losses = losses
        self.weights = weights

Now I am running into troubles, as I don't know how to configure the list of losses.
It doesn't need to be fully configurable, I just want to store some presets.
I tried setting them up with "just", but this does not work as CombinedLoss is not a primitive object.
My current solution involves setting up a new function that created the loss:

def build_loss_temp() -> Loss:
    return CombinedLoss(
        losses=[InversionLoss(), CosineSimilarityLoss()],
        weights=[0.7, 0.3],
    )
# and then saving it
loss_store(build_loss_temp, name="combined")

Though this is tideous and would make it harder to programatically create configs.

Ideally I'm aiming for something like this:

loss_store(builds(CombinedLoss, weights=[0.7, 0.3],  hydra_defaults=["_self_", {"losses": ["inversion", "cosine"]}]), name="combined")

Is this possible?

Thanks in advance :)

[rsokl]

Hydra does not treat lists very nicely -- it is hard to overwrite specific elements in lists. I think your build_loss_temp is the most straightforward approach for now. The alternative, more general approach, will require a lot more work on your end and a lot of explanation on my end. And I need to see if there is possibly a way for me to simplify things for you before you go down that path.