vggish/vggish.py

# Copyright 2021 Zilliz. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import logging
import warnings

import os
import sys
import numpy
from pathlib import Path
from typing import List

import torch

from towhee.operator.base import NNOperator
from towhee.models.vggish.torch_vggish import VGG
from towhee import register
from towhee.types.audio_frame import AudioFrame

sys.path.append(str(Path(__file__).parent))
import vggish_input

warnings.filterwarnings('ignore')
log = logging.getLogger()


@register(output_schema=['vec'])
class Vggish(NNOperator):
    """
    """

    def __init__(self, weights_path: str = None, framework: str = 'pytorch') -> None:
        super().__init__(framework=framework)
        self.device = "cuda" if torch.cuda.is_available() else "cpu"
        self.model = VGG()
        if not weights_path:
            path = str(Path(__file__).parent)
            weights_path = os.path.join(path, 'vggish.pth')
        state_dict = torch.load(weights_path, map_location=torch.device('cpu'))
        self.model.load_state_dict(state_dict)
        self.model.eval()
        self.model.to(self.device)

    def __call__(self, data: List[AudioFrame]) -> numpy.ndarray:
        audio_tensors = self.preprocess(data).to(self.device)
        features = self.model(audio_tensors)
        outs = features.to("cpu")
        return outs.detach().numpy()

    def preprocess(self, frames: List[AudioFrame]):
        sr = frames[0].sample_rate
        layout = frames[0].layout
        if layout == 'stereo':
            frames = [frame.reshape(-1, 2) for frame in frames]
            audio = numpy.vstack(frames)
        else:
            audio = numpy.hstack(frames)
            audio = audio.transpose()
        audio = self.int2float(audio)
        try:
            audio_tensors = vggish_input.waveform_to_examples(audio, sr, return_tensor=True)
            return audio_tensors
        except Exception as e:
            log.error("Fail to load audio data.")
            raise e

    def int2float(self, wav: numpy.ndarray, dtype: str = 'float64'):
        """
        Convert audio data from int to float.
        The input dtype must be integers.
        The output dtype is controlled by the parameter `dtype`, defaults to 'float64'.

        The code is inspired by https://github.com/mgeier/python-audio/blob/master/audio-files/utility.py
        """
        dtype = numpy.dtype(dtype)
        assert dtype.kind == 'f'

        if wav.dtype.kind in 'iu':
            ii = numpy.iinfo(wav.dtype)
            abs_max = 2 ** (ii.bits - 1)
            offset = ii.min + abs_max
            return (wav.astype(dtype) - offset) / abs_max
        else:
            log.warning('Converting float dtype from %s to %s.', wav.dtype, dtype)
            return wav.astype(dtype)