app.py

import json
import shlex
import subprocess

import gradio as gr
import numpy as np
import requests
import timm
import torch
import torch.nn.functional as F
from torchaudio.compliance import kaldi

TAG = "gaunernst/vit_base_patch16_1024_128.audiomae_as2m_ft_as20k"
MODEL = timm.create_model(f"hf_hub:{TAG}", pretrained=True).eval()

LABEL_URL = "https://huggingface.co/datasets/huggingface/label-files/raw/main/audioset-id2label.json"
AUDIOSET_LABELS = list(json.loads(requests.get(LABEL_URL).content).values())

SAMPLING_RATE = 16_000
MEAN = -4.2677393
STD = 4.5689974


def resample(x: np.ndarray, sr: int):
    cmd = f"ffmpeg -ar {sr} -f s16le -i - -ar {SAMPLING_RATE} -f f32le -"
    proc = subprocess.run(shlex.split(cmd), capture_output=True, input=x.tobytes())
    return np.frombuffer(proc.stdout, dtype=np.float32)


def preprocess(x: torch.Tensor):
    x = x - x.mean()
    melspec = kaldi.fbank(x.unsqueeze(0), htk_compat=True, window_type="hanning", num_mel_bins=128)
    if melspec.shape[0] < 1024:
        melspec = F.pad(melspec, (0, 0, 0, 1024 - melspec.shape[0]))
    else:
        melspec = melspec[:1024]
    melspec = (melspec - MEAN) / (STD * 2)
    return melspec.view(1, 1, 1024, 128)


def predict(audio, start):
    sr, x = audio
    if x.shape[0] < start * sr:
        raise gr.Error(f"`start` ({start}) must be smaller than audio duration ({x.shape[0] / sr:.0f}s)")

    x = resample(x[int(start * sr) :], sr)
    x = torch.from_numpy(x)

    with torch.inference_mode():
        logits = MODEL(preprocess(x)).squeeze(0)

    topk_probs, topk_classes = logits.sigmoid().topk(10)
    return [[AUDIOSET_LABELS[cls], prob.item() * 100] for cls, prob in zip(topk_classes, topk_probs)]


gr.Interface(
    fn=predict,
    inputs=["audio", "number"],
    outputs="dataframe",
    examples=[["LS_female_1462-170138-0008.flac", 0], ["LS_male_3170-137482-0005.flac", 0]],
).launch()