SiameseUtils
This repository provides a sampler to construct uniform batches from a dataset consisting of multiple labels with
multiple samples for each label. In addition, it provides a function select_triplets which - given a batch of
embeddings - selects triplets according to the batch hard strategy (https://omoindrot.github.io/triplet-loss#batch-hard-strategy).
Installation
Make sure you have numpy and torch installed, then install via:
pip install git+https://github.com/alebeck/SiameseUtilsExample usage
from torch.utils.data import DataLoader
from siamese_utils import SiameseSampler, select_triplets
train_sampler = SiameseSampler(dataset_labels, num_labels, num_samples, validation_size=val_size)
train_loader = DataLoader(dataset, batch_sampler=train_sampler)
for x, labels in train_loader:
embeddings = model(x)
loss = loss_fn(*select_triplets(embeddings, labels))
...Dataset requirements
For the triplet selection strategy to work properly, dataset[i] should return a 2-tuple consisting of the i-th
element as well as the label of the i-th element.