Triplet Loss Utility for Pytorch Library.

TripletTorch

TripletTorch is a small pytorch utility for triplet loss projects. It provides simple way to create custom triplet datasets and common triplet mining loss techniques.

Install

Install the module using the pip utility ( may require to run as sudo ).

pip3 install triplettorch

Usage

Triplet Dataset

from triplettorch import TripletDataset

# Create a triplet dataset given:
#   * labels  : array of label ( class ) for each sample of the dataset
#   * data_fn : method to access data for a given index in the dataset
#   * size    : number of samples in the dataset
#   * n_sample: number of sample per draw ( to increase probability to
#               contain valid triplets in a batch )
# Do not forget to concatenate batch dimension and sample dimension
# when used with a DataLoader as TripletDataset[ idx ] returns a
# ( batch_size, n_sample, ... ) dimension tensor for labels and data
dataset = TripletDataset( labels, data_fn, size, n_sample )

Triplet Mining

from triplettorch import AllTripletMiner, HardNegativeTripletMiner

# Define the triplet mining loss given:
#   * margin: the margin float value from the triplet loss definition
miner          = AllTripletMiner( .5 ).cuda( )
miner          = HardNegativeTripletMiner( .5 ).cuda( )

# Use the loss in training given:
#   * labels    : array of label ( class ) for each sample of the batch
#   * embeddings: output of the neural network for each sample of the batch
# Returns two values:
#   * loss    : triplet loss value
#   * frac_pos: fraction of positive triplets
#               None ( None HardNegativeTripletMiner )
loss, frac_pos = miner( labels, embeddings )

Example

The repository provides an example application with the MNIST dataset.

References

• FaceNet: A Unified Embedding for Face Recognition and Clustering
• Triplet Loss and Online Triplet Mining in TensorFlow