nnfp/nn_fingerprint.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
from pathlib import Path
from typing import List, Union

import torch
import torchaudio
import numpy

from towhee.operator.base import NNOperator
from towhee import register
from towhee.types.audio_frame import AudioFrame
from towhee.models.nnfp import NNFp
from towhee.models.utils.audio_preprocess import preprocess_wav, MelSpec
# from towhee.dc2 import accelerate

from .configs import default_params, hop25_params, distill_params

warnings.filterwarnings('ignore')
log = logging.getLogger('nnfp_op')
log.setLevel(logging.ERROR)


# @accelerate
class Model:
    def __init__(self, params, device='cpu', model_path=None):
        self.device = device
        log.info('Loading model...')
        try:
            state_dict = torch.jit.load(model_path, map_location=self.device)
        except Exception:
            state_dict = torch.load(model_path, map_location=self.device)
        if isinstance(state_dict, torch.nn.Module):
            self.model = state_dict
        else:
            dim = params['dim']
            h = params['h']
            u = params['u']
            f_bin = params['n_mels']
            n_seg = int(params['segment_size'] * params['sample_rate'])
            t = (n_seg + params['hop_length'] - 1) // params['hop_length']
            log.info('Creating model with parameters...')
            self.model = NNFp(
                dim=dim, h=h, u=u,
                in_f=f_bin, in_t=t,
                fuller=params['fuller'],
                activation=params['activation']
            ).to(self.device)
            self.model.load_state_dict(state_dict)
        self.model.eval()
        log.info('Model is loaded.')

    def __call__(self, *args, **kwargs):
        new_args = []
        new_kwargs = {}
        for x in args:
            new_args.append(x.to(self.device))
        for k, v in kwargs.items():
            new_kwargs[k] = v.to(self.device)
        return self.model(*new_args, **new_kwargs)


@register(output_schema=['vecs'])
class NNFingerprint(NNOperator):
    """
    Audio embedding operator using Neural Network Fingerprint
    """
    def __init__(self,
                 model_name: str = 'nnfp_default',
                 model_path: str = None,
                 framework: str = 'pytorch',
                 device: str = None
                 ):
        super().__init__(framework=framework)
        if device is None:
            device = 'cuda' if torch.cuda.is_available() else 'cpu'
        self.device = device
        self.model_name = model_name

        if model_name == 'nnfp_default':
            self.params = default_params
        elif model_name == 'nnfp_hop25':
            self.params = hop25_params
        elif model_name == 'nnfp_distill':
            self.params == distill_params
        else:
            raise ValueError('Invalid model name. Accept value from ["nnfp_default", "nnfp_hop25", "nnfp_distill"]')

        if model_path is None:
            path = str(Path(__file__).parent)
            model_path = os.path.join(path, 'saved_model', 'nnfp_fma.pt')
        self.model = Model(params=self.params, device=self.device, model_path=model_path)

    def __call__(self, data: Union[str, List[AudioFrame]]) -> numpy.ndarray:
        audio_tensors = self.preprocess(data)
        # print(audio_tensors.shape)
        features = self.model(audio_tensors)
        outs = features.detach().cpu().numpy()
        return outs

    @property
    def _model(self):
        return self.model.model

    def preprocess(self, frames: Union[str, List[AudioFrame]]):
        if isinstance(frames, str):
            audio, sr = torchaudio.load(frames)
        else:
            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).transpose()
            else:
                audio = numpy.hstack(frames)
                if len(audio.shape) == 1:
                    audio = audio[None, :]
            audio = self.int2float(audio)
            audio = torch.from_numpy(audio)
        assert len(audio.shape) == 2

        if sr != self.params['sample_rate']:
            resampler = torchaudio.transforms.Resample(sr, self.params['sample_rate'], dtype=audio.dtype)
            audio = resampler(audio)
            # import resampy
            # audio = audio.detach().cpu().numpy()
            # audio = resampy.resample(audio, sr, self.params['sample_rate'])

        wav = preprocess_wav(audio,
                             segment_size=int(self.params['sample_rate'] * self.params['segment_size']),
                             hop_size=int(self.params['sample_rate'] * self.params['hop_size']),
                             frame_shift_mul=self.params['frame_shift_mul'])
        wav = wav.to(torch.float32)
        mel = MelSpec(sample_rate=self.params['sample_rate'],
                      window_length=self.params['window_length'],
                      hop_length=self.params['hop_length'],
                      f_min=self.params['f_min'],
                      f_max=self.params['f_max'],
                      n_mels=self.params['n_mels'],
                      naf_mode=self.params['naf_mode'],
                      mel_log=self.params['mel_log'],
                      spec_norm=self.params['spec_norm'])
        wav = mel(wav)
        return wav

    @staticmethod
    def int2float(wav: numpy.ndarray, dtype: str = 'float64'):
        """
        Convert audio imgs 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
            if wav.dtype != 'int16':
                wav = (wav >> 16).astype(numpy.int16)
            assert wav.dtype == 'int16'
            wav = (wav / 32768.0).astype(dtype)
            return wav
        else:
            log.warning('Converting float dtype from %s to %s.', wav.dtype, dtype)
            return wav.astype(dtype)

    @property
    def supported_formats(self):
        return ['onnx']

    def save_model(self, format: str = 'pytorch', path: str = 'default'):
        if path == 'default':
            path = str(Path(__file__).parent)
            path = os.path.join(path, 'saved', format)
            os.makedirs(path, exist_ok=True)
            name = self.model_name.replace('/', '-')
            path = os.path.join(path, name)
            if format in ['torchscript', 'pytorch']:
                path = path + '.pt'
            elif format == 'onnx':
                path = path + '.onnx'
            else:
                raise ValueError(f'Invalid format {format}.')
        dummy_input = torch.rand(
            (1,) + (self.params['n_mels'], self.params['u'])
        )
        if format == 'pytorch':
            torch.save(self._model, path)
        elif format == 'torchscript':
            try:
                try:
                    jit_model = torch.jit.script(self._model)
                except Exception:
                    log.warning(
                        'Failed to directly export as torchscript.'
                        'Using dummy input in shape of %s now.', dummy_input.shape)
                    jit_model = torch.jit.trace(self._model.to('cpu'), dummy_input, strict=False)
                torch.jit.save(jit_model, path)
            except Exception as e:
                log.error('Fail to save as torchscript: %s.', e)
                raise RuntimeError(f'Fail to save as torchscript: {e}.')
        elif format == 'onnx':
            try:
                torch.onnx.export(self._model.to('cpu'),
                                  dummy_input,
                                  path,
                                  input_names=['input'],
                                  output_names=['output'],
                                  opset_version=12,
                                  do_constant_folding=True,
                                  dynamic_axes={'input': {0: 'batch_size'},
                                                'output': {0: 'batch_size'}
                                                }
                                  )
            except Exception as e:
                log.error(f'Fail to save as onnx: {e}.')
                raise RuntimeError(f'Fail to save as onnx: {e}.')
            # todo: elif format == 'tensorrt':
        else:
            log.error(f'Unsupported format "{format}".')
        return Path(path).resolve()

    def train(self, **kwargs):
        from .train_nnfp import train_nnfp
        config_json_path = kwargs['config_json_path']
        train_nnfp(
            self._model,
            config_json_path=config_json_path
        )