Beatmap Generator for osu!

Reference Repo: https://github.com/kotritrona/osumapper

Data: https://pan.baidu.com/s/1xqTLtfbMd7DMc1sWZxGCBw

.osu Format: https://osu.ppy.sh/help/wiki/osu!_File_Formats/Osu_(file_format)

Video for Generated Beatmap: https://drive.google.com/file/d/1N71juERfY-CZtFXWFYUWlTLsuG3EGOxQ/view?usp=sharing

Currently only using: Standard -> Beatmap Pack -> 800-899

How to use:

To read osu file, evaluate its rhythm, train a GAN model and get generated output, simply run the scipt.sh(Linux or MacOS) or script.bat(Windows). Remember to modify the path to the corresponding file in the script before running.

To train the GAN model, simply run the follow scripts:

$ python GAN1.py rhythm_data.npz flow_dataset.npz

or

$ python GAN.py rhythm_data.npz flow_dataset.npz

where GAN1.py contains the modified GAN architecture and training setup and GAN.py contains the original setup from reference repo.

Data structure

• "lst" (transformed data): table of [TICK, TIME, NOTE, IS_CIRCLE, IS_SLIDER, IS_SPINNER, IS_SLIDER_END, IS_SPINNER_END, SLIDING, SPINNING, MOMENTUM, ANGULAR_MOMENTUM, EX1, EX2, EX3], length MAPTICKS
• "wav" (wav data): np.array, shape of [len(snapsize), MAPTICKS, 2, fft_size//4]
• "flow" (flow data): table of [TICK, TIME, TYPE, X, Y, IN_DX, IN_DY, OUT_DX, OUT_DY] notes only

About MOMENTUM & ANGULAR_MOMENTUM

Two new variables defined by that author to capture the information about the moving speed of the mouse. Probably they are not necessary.

DONE:

1. Beatmap resources
2. Create maplist: maplist_Normal.txt, maplist_Hard.txt, maplist_Easy.txt
3. use osureader.py to parse the beatmap
4. save data to .npz files (one for each): transformed_data, wav_data, flow_data
5. CRNN, (model1 and model2)
6. GAN

TODO:

Sorry for bringing up such a big project...Let's see where we can get.

Data Preparation

1. think more about the input and output

Network Architecture (generate the 'transformed' data)

1. model1: almost no circles; deal with unbalanced training data
2. replace div data with something built with madmom downbeat tracking...

GAN (generate the 'flow' data)

1. write into python script
2. replace with my model (See predict_CRNN.py and notebook 6)

Evaluation Metrics

1. F-score (That's why we need to deal with the unbalanced issue)

Deliverables

• Report
• Demo Video: Generate with one click -> Play:)