search

BoyuanJiang / matching-networks-pytorch

Matching Networks for one shot learning
228 stars 67 forks source link

About the testing #2

Open xenuts opened 6 years ago

xenuts commented 6 years ago

I have a little question about the testing.

In the testing part, the query image always comes from the 5 support classes. But, in practice, there is a totally unknown query coming, how we choose support set (because we don't know its class).

Actually I have tested this query with all possible support sets, but I got many high confidences.

This does not make any sense, we can NOT recognise the query by the output confidence.

That makes me confused about how to use matching net in real practice.

Is my understanding all right?

SamujjwalSam commented 5 years ago

Hey @xenuts as per my understanding, you can look into the test set to manually create the support set (S_prime) during testing. Although this will be considered cheating in normal classification scenario, but for One-shot learning this is totally valid.

Hope this answers your question.