app.py

from audiocraft.models import MusicGen
from audiocraft.data.audio import audio_write
import gradio as gr
import numpy as np
import warnings
# from google.cloud import storage


class MusicGenHandler():
    def __init__(self, init_model_path='createsafe/grimes-stem-model', generation_duration=30.0):
        self.model_path = init_model_path
        self.generation_duration = generation_duration
        self._setup_model()

    def _setup_model(self):
        self.model = MusicGen.get_pretrained(self.model_path)
        self.model.set_generation_params(duration=self.generation_duration) 

    def inference(self, prompts):
        """turns prompt or list of prompts into audio"""
        if not isinstance(prompts, list):
            prompts = list(prompts)
        return self.model.generate(prompts).numpy()
    
    def update_model(self, new_model_path):
        if not new_model_path == self.model_path:
            try:
                self.model_path = new_model_path
                self._setup_model()
            except:
                warnings.warn(f"could not setup model located at {new_model_path}")

model = MusicGenHandler()

def slider_val_to_text(val):
    if val == 0:
        return "none"
    elif val == 0.1:
        return "minimal"
    elif val == 0.2:
        return "little"
    elif val == 0.3:
        return "not much"
    elif val == 0.4:
        return "just below mean"
    elif val == 0.5:
        return "mean"
    elif val == 0.6:
        return "just above mean"
    elif val == 0.7:
        return "sufficient"
    elif val == 0.8:
        return "ample"
    elif val == 0.9:
        return "great"
    elif val == 1:
        return "maximal"


def text_to_music(text, instrument, brightness, percusiveness, business, variance, temperature, bass, mids, highs, tempo, noisiness):
    dsp_feature_string = ""
    if text:
        dsp_feature_string += text + ". "
    if instrument:
        dsp_feature_string += instrument + ". "
    if brightness:
        dsp_feature_string += 'brightness ' + slider_val_to_text(brightness) + ', '
    if percusiveness:
        dsp_feature_string += 'percusiveness ' + slider_val_to_text(percusiveness) + ', '
    if business:
        dsp_feature_string += 'business ' + slider_val_to_text(business) + ', '
    if variance:
        dsp_feature_string += 'variance ' + slider_val_to_text(variance) + ', '
    if temperature:
        dsp_feature_string += 'temperature ' + slider_val_to_text(temperature) + ', '
    if bass:
        dsp_feature_string += 'bass ' + slider_val_to_text(bass) + ', '
    if mids:
        dsp_feature_string += 'mids ' + slider_val_to_text(mids) + ', '
    if highs:
        dsp_feature_string += 'highs ' + slider_val_to_text(highs) + ', '
    if tempo:
        dsp_feature_string += 'tempo ' + slider_val_to_text(tempo) + ', '
    if noisiness:
        dsp_feature_string += 'noisiness ' + slider_val_to_text(noisiness)

    if instrument == "all-stems":
        model.update_model(new_model_path='./cs-pretrained/stem_model')
    elif instrument == "drums":
        model.update_model(new_model_path='./cs-pretrained/drums_model')
    
    audio = model.inference(prompts=[dsp_feature_string])

    # convert to 16 bit PCM
    if np.max(np.abs(audio)) > 0.0:
        audio /= np.max(np.abs(audio))
    audio *= 32767
    audio = audio.astype(int)

    return (32000, audio)


def run():
    iface = gr.Interface(fn=text_to_music, inputs=[
        gr.Textbox(
            label="Text prompt"
        ),
        gr.Dropdown(
            ["all-stems", "drums", "keys", "bass"], label="Instrument"
        ),
        gr.Slider(
            0, 1, step=0.1, label="Brightness"
        ),
        gr.Slider(
            0, 1, step=0.1, label="Percussiveness"
        ),
        gr.Slider(
            0, 1, step=0.1, label="Business"
        ),
        gr.Slider(
            0, 1, step=0.1, label="Variance"
        ),
        gr.Slider(
            0, 1, step=0.1, label="Temperature"
        ),
        gr.Slider(
            0, 1, step=0.1, label="Bass"
        ),
        gr.Slider(
            0, 1, step=0.1, label="Mids"
        ),
        gr.Slider(
            0, 1, step=0.1, label="Highs"
        ),
        gr.Slider(
            0, 1, step=0.1, label="Tempo"
        ),
        gr.Slider(
            0, 1, step=0.1, label="Noisiness"
        ),
    ], outputs="audio")

    iface.launch()

# def __init__(self, init_model_path='./cs-pretrained/stem_model', generation_duration=30.0):

if __name__ == "__main__":
    run()