BatchEasyHardMiner Docs

[suchith-sixsense]

As per the docs of the BatchEasyHardMiner,

To implement the loss function described in the paper, use this miner in combination with NTXentLoss(temperature=0.1).

The paper that is in the discussion is this

According to the paper, one of the loss function they are using is this :

which implies we have to use temperature = 1.0 and not temperature = 0.1. So want to know if I am missing something or it is mentioned wrongly in docs.

Thanks.

[KevinMusgrave]

The paper says "The loss function is based on NCA which has a single parameter, temperature, and in all cases we set this parameter to 0.1."

[suchith-sixsense]

Sorry my bad, I didn't go through section 5.1 and was assuming temperature = 1.0 from the expression of loss function. Apologies for the same.