This repository has been archived by the owner on Nov 18, 2023. It is now read-only.

Continuous attribute embedding#102

Merged

jmsfltchr merged 25 commits intotypedb:masterfrom jmsfltchr:continuous-attribute-embedding

Oct 7, 2019

+253-78

Commits on Sep 19, 2019

Introduce a base class for all Attribute embedding models
James Fletcher
committed

Commits on Sep 20, 2019

A basic implementation of embedding for continuous attributes
James Fletcher
committed

Commits on Sep 25, 2019

Update pipeline
James Fletcher
committed

Commits on Sep 26, 2019

Merge branch 'master' into continuous-attribute-embedding
James Fletcher
committed
Change the control flow of graph transformation in the pipeline and build continuous attribute embedders
James Fletcher
committed

Commits on Sep 27, 2019

Add synthetic generation of continuous double attributes
James Fletcher
committed
Alter flow of embedder construction to accommodate continuous attribute embedders amd to be more robust
James Fletcher
committed
Add some basic histograms for the embedding of attributes
James Fletcher
committed
Results improve when dropout is introduced for continuous attribtues, preventing overfitting.
James Fletcher
committed

Commits on Sep 30, 2019

Use a loss function that ignores preexisting graph elements, and adds corresponding reporting metrics.
James Fletcher
committed
Alter loss function to only penalise for incorrect nodes, not edges
James Fletcher
committed
Merge branch 'learning-stability' into continuous-attribute-embedding
James Fletcher
committed

Commits on Oct 3, 2019

Commits on Oct 4, 2019

Create histograms of the gradients of the variables during training
James Fletcher
committed
Normalise type embeddings in order to increase their magnitudes in line with the attribute embeddings
James Fletcher
committed
Implement gradient clipping, which seems to greatly improve the stability and speed of convergence of the model
James Fletcher
committed
Change the implementation of dropout for continuous attributes
James Fletcher
committed
Reduce number of training iterations
James Fletcher
committed
Change control flow of learning for the case where a log directory is not specified and therefore summaries should not be written
James Fletcher
committed

Commits on Oct 7, 2019

Fix embedding tests
James Fletcher
committed
Reduce to 1000 training iterations default
James Fletcher
committed
Correct test name
James Fletcher
committed