How mangio621/mangio-rvc-fork works
Main Components
• UI Layer
– Web UI: Gradio app in infer-web.py & app.py
– CLI Interface: stateful menu in infer-web.py with --is_cli, batch scripts (go-web.bat, run.sh)
• Core Engine
– Inference Pipeline
• F0 extraction modules (lib/infer_pack/modules/F0Predictor/*)
• Feature packing (lib/infer_pack/transforms.py, infer-pack/commons.py)
• Retrieval index lookup (Faiss .index files under logs/…)
• Conversion model execution (lib/infer_pack/models*.py, lib/uvr5_pack/lib_v5/*)
• Post-processing (resample, volume mix)
– Training Pipeline
• Data preprocessing (train/data_utils.py, mel_processing.py)
• Feature extraction (train/losses.py, train/utils.py)
• Model training (VITS + cycle discriminators)
• Logging & checkpointing (logs/…, TensorBoard via Makefile)
• Configuration & Localization
– config.py, JSON files under configs/
– i18n loader (i18n.py, i18n/*.json, i18n/locale_diff.py)
• Data & Storage
– Audio assets (audios/, audio-outputs/)
– Model artifacts (pretrained/, pretrained_v2/, weights/, uvr5_weights/)
– Retrieval indexes & logs (logs/…/*.index, logs/…/mute/)
• External Dependencies
– FFmpeg/ffprobe binaries
– Python libraries (PyTorch, Faiss, NumPy, Gradio, etc.)
– Docker container (Dockerfile) & CI/CD pipelines (.github/workflows/*.yml)
Relationships & Data Flows
• User → UI: Browser (Gradio) or Terminal (CLI)
• UI → Core Engine: passes user parameters (model path, audio path, f0 method, etc.)
• Inference Engine:
– calls F0 extract → packs features → queries Faiss → runs conversion model → writes output WAV
– returns result to UI/CLI
• Training Engine:
– ingests raw audio → preprocesses → extracts features → trains model → writes checkpoints and logs
– TensorBoard watches logs folder
• Config & i18n feed parameters into both pipelines
• External tools (FFmpeg) handle audio I/O
Architectural Patterns
• Modular Separation between UI, Inference, Training, Config, Utils
• Plugin-style F0Predictors (Harvest, Dio, Crepe, hybrid) under lib/infer_pack/modules
• Retrieval-based conversion using a Faiss index for top-1 matching
• Stateful CLI Navigator implementing a finite-state menu
Technologies
• Python 3.8+
• Gradio for Web UI
• PyTorch & ONNX (optional)
• Faiss for indexing
• FFmpeg for audio processing
• Docker & GitHub Actions for CI/CD
Putting It All Together
• Top Row: User Clients (Browser / CLI) → UI Layer
• Middle Row:
– Config & i18n loader → splits into Inference vs Training pipelines
– Inference: F0 Extract → Retrieval → Model → Post-proc → Output
– Training: Preproc → Feature Extract → Model Train → Checkpoint & Logs
• Bottom Row: Data Stores & External Tools
– Model repositories, indexes, audio folders, TensorBoard, FFmpeg