I did a thing! :D
Currently working on a model to label furry artwork, based off the e621 tagging system. It labels 6 different categories: Identity, Species, Body, Clothing, Action, and Rating.
These results are after only one epoch of training on my dataset and was the preliminary test run to see if this would even work. It does 👀👀👀
Here is the architecture of the model. I start with a pre-trained CNN classifier as the base model, and then have multiple prediction heads for each category.
Every parameter is finetuned during training, meaning that the final output layer of the CNN model (the KerasLayer with an output of 1000) serves as a feature vector specifically for furry artwork.
Each prediction head also has a small fully connected Dense layer for category specific labels.
@bersl2 I would not either. *technically* this model is supposed to be able to do safe [0, 0.33), questionable [0.33, 0.67), and explicit [0.67, 1.0], but when I took the screen shot I did only safe [0, 0.5) and explicit [0.5, 1.0].
It probably falls under the questionable category.
@astroorbis soon™️
Once I get a full sized one trained and working, most likely.
zenith
Bersl
Astro