A loss function based on the distances between anchor, positive and negative embeddings used in "FaceNet: A Unified Embedding for Face Recognition and Clustering" http://arxiv.org/abs/1503.03832 . The module finds positive and negative embeddings within a current mini-batch on-the-fly, so it does not require additional space to save embeddings. This is basically a simpler version of https://github.com/Atcold/torch-TripletEmbedding .
Install this module via luarocks
luarocks install https://raw.githubusercontent.com/jhjin/triplet-criterion/master/rocks/triplet-scm-1.rockspec
The loss function can be used in the same way as other criterions except few parameters as follows.
require('triplet')
local loss = nn.TripletCriterion(samples, blocks, norm, margin) samples: the number of faces sampled from each identity in a batchblocks: the number of identities in a batch (samplesxblocks<batchSize)norm: Lp-norm for distances between embeddings (default 2)margin: a hypersphere margin between anchor-positive and anchor-negative pairs (default 0.2)
In a mini-batch, samples from the same identity should be prepared in a consecutive ordering by thier batch index.
In the case of 2 samples and 3 blocks with a batchSize of 8, for example, the batch should be prepared in
| Batch index | Identity |
|---|---|
| 1 | Person A |
| 2 | Person A |
| 3 | Person B |
| 4 | Person B |
| 5 | Person C |
| 6 | Person C |
| 7 | Person randomly sampled |
| 8 | Person randomly sampled |
From the example, anchor and positive embeddings are selected from
the first samples x blocks region (batch index 1~6)
while negative embeddings are selected from
the rest (batchSize - samples x blocks) of the region (batch index 7,8).
A large size of batch is preferred in order to let the training converge to a higher score/accuracy.