Plugin to Predictlive PP Values using a RandomForestRegressor weightned on 112302 scores from Scoresaber.
This mod also requires Counters++ for the menu options.
The RandomForestRegressor runs at about 300 Iterarions (Messured on a Ryzen 5900x) a second while the trained Keras Model runs at 30 Iterations (ong gpu taking up 5 gigs of vram!) a second.
Drag the PPP.dll into your BeatSaber mods folder.
An automatic installation routine will download missing parts.
In the configuration file ((OR COUNTERS++ SETTIGNS)) path/to/config.json, you can choose one of 2 modes. (NAME OF MODE 1) will take your current accuracy current_points / max_points_until_now (max_points_until_now += 115 * max_possible_combo # for every notehit) to predict scores, (NAME OF MODE 2) will use current_points / max_points to get a PP curve like those in Osu! Plugins that builds up over time. that means your PP score will be low for most of the song and climb towards teh end of the song.
The Plugin starts a Flask API written in Python that has been Compiled with Pyinstaller on Port 8080.
The API has one endpoint called /predict which accepts GET requets and URL the parameters stars and acc like /predict?stars=5&acc=0.9000 to predict the PP of a 5 star map with an accuracy of 90%.
The plugin sends this data to the API on each note hit and displays the predicted PP.
On ever note hit event this API is called and the value displayed. Thats it.
Using the data blocks we fitted a RandomForestRegressor and found out that it sucks. Like seriously look at these predicitons.
The lines are Stars (1-14), X = Acc and Y = PP
So we used the dataset to train an AI in Tensorflow that had a WAY better prediciton, but it was slow so we used it to predict a new dataset of 140100 entries. And this one is really accurate.
