How charantejagoud2006/audio_inpainting works
DPAI is a research audio-inpainting application: users provide audio and a gap mask (or request synthetic gaps), the system reconstructs missing audio using a deep-prior-trained MultiResUNet, then saves artifacts and metrics. The graph also preserves the benchmark/summarization and interactive web-demo capabilities documented in the README. Relationships reflect the excerpts and SOURCE INDEX; unsampled benchmark internals and some web UI wiring are not expanded.
Main entry points: inpainter.py Inpainting pipeline · app.py CLI commands · server.py Web API and jobs · total_loss.py Combined loss
Components 21
Application entry
- CLI commands
app.py - Benchmark suite
src/audio/benchmark.py
Audio preparation
- Audio loading
src/audio/loader.py - Gap masks
src/audio/corruption.py - STFT and masks
src/audio/stft.py
Reconstruction
- Inpainting pipeline
src/pipeline/inpainter.py - Deep-prior fitting
src/training/deep_prior.py - Input noise
src/training/noise.py - MultiResUNet
src/models/multires_unet.py - Harmonic convolution
src/models/harmonic_conv.py - Combined loss
src/losses/total_loss.py - Masked MSE
src/losses/masked_mse.py - Multiscale loss
src/losses/multiscale_loss.py
Evaluation and artifacts
- Run metrics
src/pipeline/metrics.py - NMSE and PESQ
src/evaluation/pesq.py - Run artifacts
src/pipeline/io.py - Plots and previews
src/utils/visualization.py
Web demo
- Web API and jobs
src/web/server.py - Interactive interface
src/web/static/app.js
Other components
- CLI user
- Browser user
Connections 27
- CLI user → CLI commands: runs commands
README.md - CLI commands → Audio loading: loads audio
app.py - CLI commands → Gap masks: creates gaps
app.py - CLI commands → Inpainting pipeline: starts inpainting
app.py - CLI commands → Benchmark suite: runs benchmark
app.py - CLI commands → Run artifacts: saves run
app.py - Inpainting pipeline → STFT and masks: transforms audio
src/pipeline/inpainter.py - Inpainting pipeline → MultiResUNet: builds model
src/pipeline/inpainter.py - Inpainting pipeline → Deep-prior fitting: optimizes model
src/pipeline/inpainter.py - Inpainting pipeline → Input noise: creates input
src/pipeline/inpainter.py - Inpainting pipeline → Combined loss: configures loss
src/pipeline/inpainter.py - Deep-prior fitting → Input noise: perturbs input
src/training/deep_prior.py - Deep-prior fitting → Masked MSE: uses default loss
src/training/deep_prior.py - MultiResUNet → Harmonic convolution: uses layers
src/models/multires_unet.py - Combined loss → Masked MSE: combines term
src/losses/total_loss.py - Combined loss → Multiscale loss: combines term
src/losses/total_loss.py - Combined loss → STFT and masks: converts spectrogram
src/losses/total_loss.py - Multiscale loss → STFT and masks: computes spectrograms
src/losses/multiscale_loss.py - Run artifacts → Plots and previews: writes plots
src/pipeline/io.py - Run artifacts → Run metrics: records metrics
src/pipeline/io.py - Run metrics → NMSE and PESQ: computes scores
src/pipeline/metrics.py - Browser user → Web API and jobs: uses demo
README.md - Web API and jobs → Audio loading: loads uploads
src/web/server.py - Web API and jobs → Gap masks: creates gaps
src/web/server.py - Web API and jobs → Inpainting pipeline: queues inpainting
src/web/server.py - Web API and jobs → Run artifacts: saves runs
src/web/server.py - Interactive interface → Web API and jobs: calls API