osu!mania score prediction through AI collaborative filtering
Project description
:arrow_forward: Try Out Opal on Streamlit
:comet: opal-net
opal is an accuracy-prediction model.
It uses Neural Collaborative Filtering to learn associations between user and maps, then using those associations to predict new scores never before seen.
Performance Error Graph
:hourglass_flowing_sand: 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
:arrow_double_down: 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.
:high_brightness: Usage
To use this, install opal-net
pip install opal-net
Then in a python script
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!")
:brain: 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.
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Hashes for opal_net-2023.8.4a0-py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | ca1d94b504a5793d0eac4b573b32d04cd7217f5dfd0e4df3c06424d8a7a3adf0 |
|
MD5 | 90e705a4bdacc1609f5388d7fd539a0f |
|
BLAKE2b-256 | 5a29e6cb725fe6080a28e0354a86929b07b7401cec3dea288f576491634d7194 |