Skip to content
/ opal Public

Rhythm Game Score Estimation through Neural Collaborative Filtering

Notifications You must be signed in to change notification settings

Eve-ning/opal

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

modelsize version pyversions https://img.shields.io/pypi/dm/opal-net

Test Model Pipeline Inference

☄️ opal-net

opal is an accuracy-prediction model.

It uses a Matrix Factorization branch & Multi-layered Perceptron branch to learn associations between user and maps, then use those associations to predict new scores never before seen.

⏳ Project Status

Currently, it's in its early access, that means, it'll have many problems! However, we're working on it to minimize these issues o wo)b

⏬ Dataset Used

I used the top 10K mania users data from https://data.ppy.sh. After preprocessing, we use

  • ~10m scores for training
  • ~1m scores for validation and testing each

After preprocessing, we found ~30K valid users, ~10K valid maps This models can thus help predict ~300m unplayed scores!

Users

We deem a player on separate years as a different user. This is to reflect the improvement of the player after time.

🔆 Usage

To use this, install opal-net

pip install opal-net

Then in a python script

Tip: GPU doesn't speed this up significantly, you can use a CPU.

from opal import OpalNet

# Load in the model
# You can explicitly specify map_location='cpu' or 'cuda' in map_location=...
opal = OpalNet.load()

# You can predict a single instance.
#
# The 1st arg: "<USER_ID>/<YEAR>",
# The 2nd arg: "<MAP_ID>/<SPEED>" 
#   <YEAR> is the year of the user to test.
#   <SPEED> can be {-1, 0, or 1} for {HT, NT, DT}
#
# For example: 
# Predict Evening on Year 2020, on the map Triumph & Regret [Regret] at Double Time
pred = opal.predict("2193881/2020", "767046/1")

# You can predict multiple entries at the same time. This is much faster that looping the above.
# Note that both lists must be of the same length!
# Note: If you're predicting millions, partition the predictions to reduce GPU memory usage!
preds = opal.predict(["2193881/2020", "2193881/2017"], ["767046/0", "767046/1"])

# Note that if the prediction doesn't exist, then it'll raise a ValueError
try:
    opal.predict("2193881/2018", "767046/0")
except ValueError:
    print("Prediction Failed!")

🧠 AlphaOsu!

Currently, opal doesn't provide recommendations, however, you can try out AlphaOsu!.

Annex

Why not Score Metric?

Score is not straightforward to calculate, and may be difficult to debug. Furthermore, score isn't of interest when calculating performance points anymore.

osu!mania ScoreV1 Reference