dice-embeddings 1.0.4

Features

1. KvsSample technique implemented. KvsSample is KvsAll with selected tail entities. This technique reduces the memory usage during training as we can select the number of tail entities.
2. Sharded Training tested

Maintenance

1. Use Python 3.9
2. More tests are added
3. ReadMe is structured

Todos for the next release

1. Explicit Class Kronecker Decomposition at retriving embeddings

dice-embeddings 1.0.3

Features

Self-supervised Learning module: Pseudo-Labelling and Conformal Credal Self-Supervised Learning implemented.

Maintenance

1. Documentation & Instrations are improved.
2. Use Python 3.10 due to PEP 635

Todos for the next release

2. Consider using Weights & Biases
3. Study [Raymond Hettinger](https://twitter.com/raymondh/status/1533369936739016705) 's talk about Structural Pattern Matching in the Real World: New tooling, real code, problems solved.
4. Explicit Class Kronecker Decomposition at retriving embeddings

Dice-Embeddings

Features

Batch Relaxation training strategy started.
Seed selection for the computation is available.
Input KG size reduction: Entities that do not occur X times can be removed.
Lower memory usage through selecting most efficient index type.
swifter is included to do dataframe().apply() via using all CPUs
QMult with 11.4 B on DBpedia is succesfuly trained and deployed.

Maintenance

The title of the repo. has been changed.
Repo name has been changed.
Testing with three pytest setting is documented
Regression tests are extended.
More functions and classes are documented.

Todos for the next release

1. Use Python 3.10 to use PEP 635
2. Use Python 3.10 to benefit from 10% performance increase and https://bugs.python.org/issue42093 <3
3. Consider using Weights & Biases

KGE at scale

Features

1. 1vsAll, KvsAll, Negative Sampling strategy available.
2. Continues training (training pre-trained model) is implemented
3. Deployment and opensourcing a pre-trained KG is implemented.
4. Using multi-CPUs during pre-processing is available. This is important to scale on large KGs
5. Artificial noise can be added into input KG. The rate of added noisy triples can be used to reguirlizer the selected KG, as well as, compare link prediction performances of KGE models under noisy data.

Maintenance

1. The title of the repo. has been changed
2. Manuel Instalation added.
3. Links for pre-trained models stored on Hobbit Data are provided.
4. How to cite section is added.
5. Regression tests are added.