kraina-ai / srai

Spatial Representations for Artificial Intelligence - a Python library toolkit for geospatial machine learning focused on creating embeddings for downstream tasks
https://kraina-ai.github.io/srai/
Apache License 2.0
220 stars 17 forks source link

Question: How to interpret the hex2vec embedder training loss? #449

Open adesso-dominik-chodounsky opened 7 months ago

adesso-dominik-chodounsky commented 7 months ago

I am using the hex2vec embedder and upon looking at the training loss, I see that no matter what hyperparameters I use, the first epoch step has a loss larger than 1.0 (e.g. 1.5) and all the next steps have exactly 1.0.

For f-score, the metric gradually increases, so I guess the model is getting better, but the loss stays exactly the same. Can anyone explain this behaviour please?