Re-train BlogCatalog with embedding dimension 2

[wehlutyk]

So that we have the training history, not only the checkpoint.

[wehlutyk]

Training done, must create a results notebook.

[wehlutyk]

Results are in 509b78a, see the projects/scale/blogcatalog-dim_ξ=2-results.ipynb.

Highlights (see again figures below):

• nodes (or at least the average values for the gaussian distributions) are aligned on a line in the embedding, just like some of the projections for Train BlogCatalog with higher embedding dimension #31. But in this case it's the only representation that is produced, and it's rather poor.
• adjacency reconstruction is bad
• feature loss seems to go back up a little (after epoch 2000), while adjacency loss still goes down, so maybe the decoder is having a hard time accommodating the two constraints.

[wehlutyk]

Additional things to explore:

• train with no feature reconstruction, and id features on input → Train BlogCatalog with no feature reconstruction (and Id input features) #40
• train with a much higher (e.g. 10000) weight on adjacency reconstruction → Train BlogCatalog with a much higher weight (e.g. 1000) on adjacency reconstruction #41