Inconsistency in default parameters of Triplet Loss and Triplet Loss miner

[Sorrow321]

Hi,

In TripletMarginLoss you have default margin set to 0.05:

class TripletMarginLoss(BaseMetricLossFunction):
    """
    Args:
        margin: The desired difference between the anchor-positive distance and the
                anchor-negative distance.
        swap: Use the positive-negative distance instead of anchor-negative distance,
              if it violates the margin more.
        smooth_loss: Use the log-exp version of the triplet loss
    """

    def __init__(
        self,
        margin=0.05,
        swap=False,
        smooth_loss=False,
        triplets_per_anchor="all",
        **kwargs
    ):

While in TripletMarginMiner it is 0.2

class TripletMarginMiner(BaseTupleMiner):
    """
    Returns triplets that violate the margin
    Args:
        margin
        type_of_triplets: options are "all", "hard", or "semihard".
                "all" means all triplets that violate the margin
                "hard" is a subset of "all", but the negative is closer to the anchor than the positive
                "semihard" is a subset of "all", but the negative is further from the anchor than the positive
            "easy" is all triplets that are not in "all"
    """

    def __init__(self, margin=0.2, type_of_triplets="all", **kwargs):
        super().__init__(**kwargs)
        self.margin = margin
        self.type_of_triplets = type_of_triplets
        self.add_to_recordable_attributes(list_of_names=["margin"], is_stat=False)
        self.add_to_recordable_attributes(
            list_of_names=["avg_triplet_margin", "pos_pair_dist", "neg_pair_dist"],
            is_stat=True,
        )

I think it's better to set them to some common value, because I guess that most of people want to use them out of box

[Sorrow321]

Or maybe I'm mistaken and it is intentional to mine "hardest" samples?

[KevinMusgrave]

I think the miner margin was set based on some paper. I don't know about the loss margin. Setting them to the same value could be reasonable.