Auto KGE Feature

[Demirrr]

For a given dataset and a knowledge graph embedding model, make a sequence decisions so that our computation and time budgets are fully utilized, specficaly

1. Find a combination of the batch size and size of embedding vectors that fully uses the available hardware.
2. Through validation set performance, we may study adaptive rates for drop outs

[Demirrr]

3. Detect available GPUs and the size of the KG and move to model paralleisim from data parallelsim

[Demirrr]

Update:

Find a combination of the batch size and size of embedding vectors that fully uses the available hardware.

Possible workflow:

1. Initialize the model and check the CPU usage
2. Perform dummy forward and backward pass to compute memory usage
3. Assume that (2) is a "good" approximation of actual memory usage in CPU/GPU etc.
4. Increment the batch size until out-of-memory issue occurs.

[Demirrr]

Auto batch finder is available here.