clmr/clmr_magnatagatune.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 os
import sys
import logging
from pathlib import Path
from typing import Union

import torchaudio
import torch
import numpy

from towhee.operator import NNOperator
from towhee import register

sys.path.append(str(Path(__file__).parent))
from clmr_checkpoint import load_encoder_checkpoint
from sample_cnn import SampleCNN

log = logging.getLogger()


@register(output_schema=['vecs'])
class ClmrMagnatagatune(NNOperator):
    """
    Pretrained clmr
    """

    def __init__(self, framework="pytorch") -> None:
        super().__init__(framework=framework)
        self.device = "cuda" if torch.cuda.is_available() else "cpu"

        weight_path = os.path.join(str(Path(__file__).parent),
                                   'clmr_checkpoint_10000.pt')
        state_dict = load_encoder_checkpoint(weight_path, 1)
        encoder = SampleCNN(strides=[3, 3, 3, 3, 3, 3, 3, 3, 3], supervised=False, out_dim=1)
        encoder.load_state_dict(state_dict)

        new_encoder = torch.nn.Sequential(*(list(encoder.children())[:-1]))
        x = list(new_encoder[0][:10].children())
        y = torch.nn.Sequential(*list(new_encoder[0][10].children())[:-1])
        x.append(y)
        self.model = torch.nn.Sequential(*x)
        self.model.eval()
        self.model.to(self.device)

    def __call__(self, audio: Union[str, numpy.ndarray], sample_rate: int = None) -> numpy.ndarray:
        _sr = 22050
        audio_length = 59049

        if isinstance(audio, str):
            source = os.path.abspath(audio)
            audio, sr = torchaudio.load(source)
        elif isinstance(audio, numpy.ndarray):
            sr = sample_rate
            audio = torch.tensor(audio).to(torch.float32)

        if sr != _sr:
            transform = torchaudio.transforms.Resample(orig_freq=sr, new_freq=_sr)
            audio = transform(audio)

        with torch.no_grad():
            batch = torch.split(audio, audio_length, dim=1)
            batch = torch.cat(batch[:-1])
            batch = batch.unsqueeze(dim=1)
            batch = batch.to(self.device)
            features = numpy.squeeze(self.model(batch))

        embeddings = features.to("cpu")
        return embeddings.detach().numpy()


# if __name__ == "__main__":
#     encoder = ClmrMagnatagatune()
#
#     # audio_path = "/audio/path/or/link"
#     # vec = encoder(audio_path)
#
#     audio_data = numpy.zeros((2, 441344))
#     sample_rate = 44100
#     vec = encoder(audio_data, sample_rate)
#
#     print(vec.shape)