Studying Backward Compatible Embeddings

A repository created for studying different aspects of Backward Compatible Embeddings DL architecture for machine learning tasks.

Problem formulation

Problem formulation can be found in the original paper repository bc-emb.

Modules

torch_bce contains several main modules:

1. containers
   1. ListModelContainer - container for used models
   2. WeightsContainer - module for Backward Transformation storage, @deprecated(0.1.0)
2. interfaces - main interfaces to inherit from
   1. ModelInterface - interface for PyTorch backward compatible models to subclass
   2. Saveable - save/load interface
   3. TrainerInterface - interface for trainer
3. losses - different useful losses.
   1. GraphSAGELoss - loss for unsupervised embeddings learning on graphs as described in paper.
   2. AlignmentLoss - loss for alignment of different embeddings that has its own state and should be optimized.
   3. BackwardTransformation - backward transformation for embeddings.
4. models
   1. intended
      1. GraphSAGE
   2. unintended
      1. MLP
5. trainers
   1. GSAlignmentTrainer - for training of intended models in unified manner.
6. tests - tests with help of unittest
7. utils - some useful utilities, such as:
   1. datasets
   2. metrics
   3. samplers

Classes Diagram

Reproducibility

Docker

For now there is Dockerfile that enables to start all tests. Just execute ./build_test.sh from inside the directory.

Setup

To set up the framework, execute pip install -e . from inside the directory. There is also requirements file, however some dependencies (e.g., PyTorch Geometric and PyTorch Geometric Temporal must be installed manually, as it is done in Dockerfile.

Example