Training only B side

If training is failing to start, and you are not receiving an error message telling you what to do, tell us about it here

Forum rules

Read the FAQs and search the forum before posting a new topic.

This forum is for reporting errors with the Training process. If you want to get tips, or better understand the Training process, then you should look in the Training Discussion forum.

Please mark any answers that fixed your problems so others can find the solutions.

Post Reply
User avatar
Posts: 7
Joined: Thu Apr 20, 2023 1:16 pm

Training only B side

Post by jode »

Maybe stupid question but is it possible to train only B side? I often use same B side to different A sides. Hoping I could make B side "ready" for next project and then just load weights to make training faster. Is it possible with Phaze-A trainer?

User avatar
Posts: 182
Joined: Thu May 26, 2022 6:02 am
Has thanked: 159 times
Been thanked: 13 times

Re: Training only B side

Post by MaxHunter »

The short answer is, no.

What you can do, if you use Phaze A, is freeze the B side, and reuse the weights. As you continually re-use the weights in theory it should take less time to train.

The model needs both sides to build properly.

Last edited by MaxHunter on Tue Sep 19, 2023 2:08 am, edited 1 time in total.
User avatar
Posts: 2520
Joined: Fri Jul 12, 2019 12:53 am
Answers: 149
Has thanked: 125 times
Been thanked: 576 times

Re: Training only B side

Post by torzdf »

Absolutely you can do this, to an extent. It will save time for the next training, but not completely eliminate.

You would

  • create a new model in Phaze-A, same structure as the initial model.
  • In the "load weights" section, load in weights for every part of the model that are "shared", "both" or "B" only
  • Freeze the same weights as you selected for "load weights"
  • In the main training window, "Load Weights" from the model that has the B data you want to keep
  • Start training

The A side should start to catch up pretty quickly with the B-Side (you can train higher batch-size too, as less of the model is training).
Once you are happy that the A has got as far as it is going to get, unfreeze the whole model to let the rest of it adapt to the new data.

My word is final

Post Reply