Page 1 of 1
Possibility of Turning Off "Learn Mask" on an Active Training Model
Posted: Sun Dec 19, 2021 12:37 pm
by ianstephens
We have been playing with different masks and settings recently.
We have a model that's at 1M iterations with "learn mask" enabled.
We'd like to switch off the learned mask and allow the trainer to work with the original mask in the PNG files.
Is this possible with an active training session?
We simply tried to set
in the existing JSON file but it hit us back with an error when we tried to continue training on the model.
Is this something that can be done?
Thanks for any help, suggestions, and tips in advance.
Re: Possibility of Turning Off "Learn Mask" on an Active Training Model
Posted: Sun Dec 19, 2021 7:20 pm
by bryanlyon
Learn Mask changes the model itself in order to let it learn the mask. Unfortunately once a model is started, it can't be changed. You're going to be stuck with the mask being learned, that said, once convert comes up, you can just not use the predicted mask.
Re: Possibility of Turning Off "Learn Mask" on an Active Training Model
Posted: Sun Dec 19, 2021 7:35 pm
by ianstephens
Thanks for that [mention]bryanlyon[/mention] - I did figure that was the case as it was throwing up errors about the model itself (can't remember the exact message) when we tried to manually edit the JSON config.
The issue we came across is that when using the original mask in the preview tool (after training with predicted), the model had only covered details learned what was inside the predicted mask.
And when using the predicted, there was a noticeable lack of coverage of the face area (the mask had been made smaller by the prediction algo).
Either way, it looks bad but the original mask is better (although has shading/black areas at the edges where the predicted learned model cuts off).
If you have any tips/suggestions that would be great - we're 1.2M iter. in and about to switch on nowarp.
Re: Possibility of Turning Off "Learn Mask" on an Active Training Model
Posted: Wed Jan 19, 2022 3:23 am
by torzdf
There is a fix for this, but I can only tell you top level, I can't take you through this step-by-step as it's involved.
It is possible to create a new model which does not have "learn mask" enabled, and then port your required weights from your existing model into this new model.
If you can find someone with intermediate to advanced experience of tf/keras, they should be able to do this for you.
Alternatively, I'm currently taking contracts 