Question on inference library

[haoyanc]

Hi, i have recently trained a model using the triplet margin loss notebook as baseline reference with my custom dataset. How does the result from the is_match function different from the accuracy calculated from the triplet loss notebook. The is_match function seems to be performing worse when trying to match data of the same labels.

[KevinMusgrave]

The accuracy calculation uses k-nearest-neighbors, whereas is_match uses thresholding with distance or similarity.

For example:

• k-nearest-neighbors with k = 1: for each image, find the 1st closest neighbor in all of the training data, and if it's the same class, then give this image a score of 1.
• is_match: given two images, if the distance between them is less than the threshold, then it's a match.