--Train_Text_Encoder variation #56
-
|
I have noticed that --train_text_encoder has a significant effect on training. However, it is very aggressive, and I have read that on some pages its percentage can be varied. Is that possible? Is there such a command? It would be really important. |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 2 replies
-
|
Looking at TheLastBen colab (https://colab.research.google.com/github/TheLastBen/fast-stable-diffusion/blob/main/fast-DreamBooth.ipynb), the parameter used to control this is '--stop_text_encoder_training=' However the value the parameter takes is not a percentage, so the colab calculates this as follows: TheLastBen/fast-stable-diffusion@3e28d1a Seems like a good option to have, as just as described in the commit, it's easy to overfit when using --train_text_encoder |
Beta Was this translation helpful? Give feedback.
-
|
Thanks!. A doubt. There is a comand there "Train_text_encoder_for=" wich is for %. Can i use that in the parameters or i should use "--stop..." to put the number of steps manually? |
Beta Was this translation helpful? Give feedback.
-
|
No, "Train_text_encoder_for=" is just part of the colab gui, the actual parameter is '--stop_text_encoder_training=', but I don't know if dreambooth-gui supports it. If I'm not mistaken dreambooth-gui is based upon ShivamShrirao's repo, and I don't know if that supports --stop_text_encoder_training I'm currently running a training so I can't try it myself. |
Beta Was this translation helpful? Give feedback.
Looking at TheLastBen colab (https://colab.research.google.com/github/TheLastBen/fast-stable-diffusion/blob/main/fast-DreamBooth.ipynb), the parameter used to control this is '--stop_text_encoder_training='
However the value the parameter takes is not a percentage, so the colab calculates this as follows: TheLastBen/fast-stable-diffusion@3e28d1a
Seems like a good option to have, as just as described in the commit, it's easy to overfit when using --train_text_encoder