GitHubEdit on GitHub

Visualization Tools

Overview

Amphion provides visualization tools to help understand and analyze audio generation models. These tools offer interactive visualizations of model architectures, training processes, and generated outputs.

SingVisio

SingVisio is a visualization tool for understanding diffusion models in singing voice conversion:

Features

  • Interactive visualization of the diffusion process
  • Real-time audio playback
  • Step-by-step denoising visualization
  • Model architecture visualization

Usage

from amphion.visualization import SingVisio

# Initialize visualizer
visualizer = SingVisio(
    model_path="path/to/model",
    device="cuda"
)

# Visualize conversion process
visualizer.visualize(
    source="path/to/source.wav",
    target="path/to/target.wav"
)

Model Architecture Visualization

Visualize model architectures with detailed layer information:

from amphion.visualization import visualize_model

# Visualize model architecture
visualize_model(
    model_type="fastspeech2",
    save_path="architecture.png"
)

Training Visualization

Monitor training progress with real-time visualizations:

from amphion.visualization import TrainingVisualizer

visualizer = TrainingVisualizer(
    log_dir="path/to/logs",
    update_interval=100
)

# Add to training loop
visualizer.update(
    loss=current_loss,
    step=current_step
)

Audio Visualization

Visualize audio features and transformations:

from amphion.visualization import plot_audio

# Plot waveform and spectrogram
plot_audio(
    audio_path="path/to/audio.wav",
    features=["waveform", "mel", "f0"],
    save_path="visualization.png"
)

Interactive Dashboard

Launch an interactive dashboard for model analysis:

from amphion.visualization import Dashboard

dashboard = Dashboard(
    model="path/to/model",
    port=8080
)

dashboard.serve()
Datasets