AUTOMATIC1111
Adding optional tag shuffling for training
When conducting training, there is a chance that a fixed tag sequence can effect training while working with deepdanbooru, due to the model utilizing attention. As far as i know, in the training case of Waifu Diffusion, a tag has been shuffled to reduce the effect of the tag sequence. Thus I added the following function considering that it would be helpful when training if there is an option that shuffles tag sequence randomly in every text generation. This worked well when I applied this function on my own embedding training
Shouldn't tag shuffle occur for every text token inside a random self.lines, instead of just [filewords] ?
Shouldn't tag shuffle occur for every text token inside a random self.lines, instead of just [filewords] ?
As far as i know, NAI network shown better quality using prompts 'masterpiece, best quality' in front of another tag. I think some tags must be fix in the position as options. if want to shuffle tags instead of just [filewords], then we can modify textual_inversion_templates files.
I'd rather have it as a setting in the settings section for training.
Added the option to apply dropouts to tags to reduce the impact of excessive incorrect tag entry. I think the impact on excessive inaccurate tags can be reduced.