Saved files?

[bbecausereasonss]

Awesome work ladies/gents! Where does the gradio interface save generated files?

[john-cornell]

They don't appear to be directly saved, however you can add the following (assuming google.colab here, you may need to find another download method, I'm sure phind.com or GPT can figure it out for you)

I largely cobbled this together from https://huggingface.co/spaces/facebook/MusicGen/blob/main/app.py

from tempfile import NamedTemporaryFile
import gradio as gr
from google.colab import files
#given res is name of output from model.generate, as in the code from 
# https://colab.research.google.com/drive/1fxGqfg96RBUvGxZ1XXN07s3DthrKUl4-?usp=sharing
output = res.detach().cpu().float()[0] 
with NamedTemporaryFile("wb", suffix=".wav", delete=False) as file:
        audio_write(file.name, output, model.sample_rate, strategy="loudness", add_suffix=False)
        waveform_video = gr.make_waveform(file.name)

file.name #If you are at all interested
files.download(file.name)

[BigEnding-8347]

If you run it locally, it is in fact saved directly in temporary folders without changes as of now.

For Windows 11 using miniconda3, check in 'C:\Users\YOURUSERNAMEHERE\AppData\Local\Temp\gradio'. All these folders you will find are the single results.

To find it in a different OS or environment you can use 'inspect element' when you right click the result-file in a browser when using Gradio. Hope it helps. 👍

[SatyaDewangan05]

You can redefine the function display_audio in audiocraft.utils.notebook.

First install some libraries

import locale
locale.getpreferredencoding = lambda: "UTF-8"
!pip install PyDub

Now redefine the function

from pydub import AudioSegment
try:
    import IPython.display as ipd  # type: ignore
except ImportError:
    # Note in a notebook...
    pass

import torch

def display_audio(samples: torch.Tensor, sample_rate: int):
    """Renders an audio player for the given audio samples.

    Args:
        samples (torch.Tensor): a Tensor of decoded audio samples
            with shapes [B, C, T] or [C, T]
        sample_rate (int): sample rate audio should be displayed with.
    """
    assert samples.dim() == 2 or samples.dim() == 3

    samples = samples.detach().cpu()
    if samples.dim() == 2:
        samples = samples[None, ...]
    count = 0
    for audio in samples:
        count += 1
        file_name_mp3 = f'music_00{count}.mp3' # audio file_name

        audio = ipd.Audio(audio, rate=sample_rate) # display audio in HTML
        ipd.display(audio)

        audio = AudioSegment(audio.data, frame_rate=22050, sample_width=2, channels=1)
        audio.export(file_name_mp3, format="mp3", bitrate="64k") # save audio

Now everytime you call the display_audio function, it'll store the audio by the names music_001.mp3

display_audio(res, 32000)

You can modify the file_name by using uuid

[jaminmc]

On my mac, It saved them in the folder: /private/var/folders/sq/05t2tpgd3nvbzklh3n827y2c0000gn/T