Triplet loss on Acoustic Scene Classification(ASC)-PyTorch

PyTorch implementation of triplet networks for learning embeddings.

Triplet networks are useful to learn mapping from input to a compact
Euclidean space where distances correspond to a measure of similarity.

Installation

Requires pytorch 0.4 with torchvision 0.2.1

Code structure

• data_manager folder

  • Please see the data_manager/README.md for details.

• networks.py

  • some network classes. e.g. vggish_bn. it’s VGG-like network architecture.

• losses.py

  • OnlineTripletLoss class -triplet loss for triplets of embeddings.

• metrics.py

  • Sample metrics that can be used with fit function from trainer.py

• trainer.py

  • fit - unified function for training a network with different number of
    inputs and different types of loss functions.

• utils.py

  • FunctionNegativeTripletSelector class -generating triplets based
    on embeddings and ground truth class labels.
  • plot_embeddings, extract_embeddings are function of learned embeddings visualization.

• experiment folder

  • classification_baseline.py - A baseline of classification code.
  • hard_triplet_baseline.py - A random hard selection of triplets baseline code.

• jupyter_script

  • some experiment results.

TODO

1. change network architecture.
2. novel tripets selection strategy
3. pickup classifier.
4. verification on embedding metric.