How sweetcocoa/pop2piano works
Pop2Piano generates piano-cover MIDI from pop audio, with inference as the user-to-result path: a user runs test-set inference, which loads each sample, invokes the generation wrapper, and saves composer-specific MIDI. Its dataset workflow downloads media and preprocesses beat timing, quantized MIDI, and aligned audio; evaluation measures generated melody accuracy. The graph includes documented but unsampled preprocessing branches by directory path; relationships for those branches are omitted where source evidence does not establish calls.
Main entry points: transformer_wrapper.py Transformer Generation · bpm_quantize.py Tempo Quantization · inference_testset.py Test-set Inference · midiaudiopair.py Sample Metadata
Components 18
Cover Generation
- Test-set Inference
evaluate/inference_testset.py - Transformer Generation
transformer_wrapper.py - MIDI Tokenizer
midi_tokenizer.py - Audio Features
layer/input.py - Generated MIDI
Dataset Preparation
- Media Downloader
download/download.py - Original Media
- Beat Quantization
preprocess/beat_quantizer.py - Tempo Quantization
preprocess/bpm_quantize.py - Audio-MIDI Alignment
preprocess/pop_align.py - Stem Separation
preprocess/split_spleeter.py - Melody Accuracy Prep
preprocess/melody_accuracy.py
Evaluation
- Test-set Evaluation
evaluate/evaluate_testset.py - Melody Accuracy
evaluate/midi_melody_accuracy.py
Shared Data and DSP
- Sample Metadata
midiaudiopair.py - Audio DSP
utils/dsp.py - Beat Features
Other components
- User
Connections 19
- User → Test-set Inference: runs
evaluate/inference_testset.py - Test-set Inference → Sample Metadata: loads sample
evaluate/inference_testset.py - Test-set Inference → Transformer Generation: invokes generation
evaluate/inference_testset.py - Transformer Generation → Generated MIDI: writes MIDI
- Transformer Generation → MIDI Tokenizer: uses tokens
transformer_wrapper.py - Transformer Generation → Audio Features: uses features
transformer_wrapper.py - Transformer Generation → Beat Quantization: uses beat quantization
transformer_wrapper.py - Transformer Generation → Audio DSP: uses DSP
transformer_wrapper.py - User → Media Downloader: runs downloader
README.md - Media Downloader → Original Media: downloads media
- Tempo Quantization → Sample Metadata: loads sample
preprocess/bpm_quantize.py - Tempo Quantization → Beat Quantization: extracts and quantizes
preprocess/bpm_quantize.py - Tempo Quantization → Audio DSP: mixes audio
preprocess/bpm_quantize.py - Tempo Quantization → Beat Features: saves beat data
preprocess/bpm_quantize.py - Audio-MIDI Alignment → Sample Metadata: loads sample
preprocess/pop_align.py - Audio-MIDI Alignment → Audio DSP: uses audio helpers
preprocess/pop_align.py - Test-set Evaluation → Sample Metadata: loads sample
evaluate/evaluate_testset.py - Test-set Evaluation → Melody Accuracy: scores melody
evaluate/evaluate_testset.py - Test-set Inference → Melody Accuracy: evaluates melody
evaluate/inference_testset.py