app.py

import torch
import torchaudio
import gradio as gr
from demucs import pretrained
from demucs.apply import apply_model
from pyharp import *
from audiotools import AudioSignal

# Available Demucs models
DEMUX_MODELS = ["mdx_extra_q", "mdx_extra", "htdemucs", "mdx_q"]

STEM_CHOICES = {
    "Vocals": 3,   
    "Drums": 0,    
    "Bass": 1,     
    "Other": 2,    
    "Instrumental (No Vocals)": "instrumental"
}


def separate_stem(audio_file_path: str, model_name: str, stem_choice: str):
    """
    Separates an audio file into the chosen stem using a Demucs model.
    Ensures correct stem ordering and supports mono input.
    """
    # Load Demucs model
    model = pretrained.get_model(model_name)
    model.to('cuda' if torch.cuda.is_available() else 'cpu')
    model.eval()

    # Load the audio file
    waveform, sr = torchaudio.load(audio_file_path)

    # Check if input is mono
    is_mono = waveform.shape[0] == 1
    if is_mono:
        waveform = waveform.repeat(2, 1)  # Convert mono to stereo for Demucs

    # Apply Demucs model
    with torch.no_grad():
        stems_batch = apply_model(
            model,
            waveform.unsqueeze(0),
            overlap=0.2,
            shifts=1,
            split=True
        )

    # stems shape: (batch, stems, channels, samples)
    stems = stems_batch[0]  

    print(f"Model '{model_name}' extracted stems shape: {stems.shape}")

    if stem_choice == "Instrumental (No Vocals)":
        stem = stems[0] + stems[1] + stems[2]  # Drums + Bass + Other 
    else:
        stem_index = STEM_CHOICES[stem_choice]
        stem = stems[stem_index]

    # Convert back to mono if the input was originally mono
    if is_mono:
        stem = stem.mean(dim=0, keepdim=True)  # Stereo → Mono

    # Convert to AudioSignal with float32 dtype
    stem_signal = AudioSignal(stem.cpu().numpy().astype('float32'), sample_rate=sr)
    return stem_signal


def process_fn_stem(audio_file_path: str, demucs_model: str, stem_choice: str):
    """
    PyHARP process function:
      - Separates the chosen stem using Demucs.
      - Saves the stem as a .wav file.
    """
    stem_signal = separate_stem(audio_file_path, model_name=demucs_model, stem_choice=stem_choice)
    stem_path = save_audio(stem_signal, f"{stem_choice.lower().replace(' ', '_')}.wav")
    return stem_path, LabelList(labels=[])


# Define the model card
model_card = ModelCard(
    name="Demucs Stem Separator",
    description="Uses Demucs to separate a music track into a selected stem.",
    author="Alexandre Défossez, Nicolas Usunier, Léon Bottou, Francis Bach",
    tags=["demucs", "source-separation", "pyharp", "stems"]
)

# Build Gradio interface with dropdowns for model and stem selection
with gr.Blocks() as demo:
    components = [
        gr.Dropdown(
            label="Select Demucs Model",
            choices=DEMUX_MODELS,
            value="mdx_extra_q"
        ),
        gr.Dropdown(
            label="Select Stem to Separate",
            choices=list(STEM_CHOICES.keys()),
            value="Vocals"

        )
    ]

    app = build_endpoint(
        model_card=model_card,
        components=components,
        process_fn=process_fn_stem
    )

demo.queue()
demo.launch(share=True, show_error=True)