search

Zyin055 / Inspect-Embedding-Training

Python script to analyze textual inversion embedding files used with AI image generators
MIT License
102 stars 11 forks source link

Learn rate on graph is offset #19

Closed Xyem closed 1 year ago

Xyem commented 1 year ago

The vertical lines and labels for learning rates are offset (very obvious with cyclical learn profiles). Perhaps just a fence post error?

In the attachment, the high learn rate (0.05) is shown as being a small block at the start of a low learn (0.0005) section immediately after the "busy" section, instead of surrounding the busy section like it should be.

vd-596-vector

Zyin055 commented 1 year ago

This visual issue happens because your embedding files and loss rate aren't being saved to disk at the same rate. Likely you have your embeddings saved every X steps, but loss rate saved every Y steps (the learn rate is extracted from the loss rate .csv file).

Xyem commented 1 year ago

Ah, I see. Indeed, setting the CSV to be saved every step makes the lines render in the correct place. Thanks for explaining!