PyTorch Audio Classification: Urban Sounds

Classification of audio with variable length using a CNN + LSTM architecture on the UrbanSound8K dataset.

Example results:

Contents

• Models
• Inference
• Training
• Evaluation
• To Do

Dependencies

• soundfile: audio loading
• torchparse: .cfg easy model definition
• pytorch/audio: Audio transforms

Features

• Easily define CRNN in .cfg format
• Spectrogram computation on GPU
• Audio data augmentation: Cropping, White Noise, Time Stretching (using phase vocoder on GPU!)

Models

CRNN architecture:

Printing model defined with torchparse:

AudioCRNN(
  (spec): MelspectrogramStretch(num_bands=128, fft_len=2048, norm=spec_whiten, stretch_param=[0.4, 0.4])
  (net): ModuleDict(
    (convs): Sequential(
      (conv2d_0): Conv2d(1, 32, kernel_size=(3, 3), stride=(1, 1), padding=[0, 0])
      (batchnorm2d_0): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (elu_0): ELU(alpha=1.0)
      (maxpool2d_0): MaxPool2d(kernel_size=3, stride=3, padding=0, dilation=1, ceil_mode=False)
      (dropout_0): Dropout(p=0.1)
      (conv2d_1): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=[0, 0])
      (batchnorm2d_1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (elu_1): ELU(alpha=1.0)
      (maxpool2d_1): MaxPool2d(kernel_size=4, stride=4, padding=0, dilation=1, ceil_mode=False)
      (dropout_1): Dropout(p=0.1)
      (conv2d_2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=[0, 0])
      (batchnorm2d_2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (elu_2): ELU(alpha=1.0)
      (maxpool2d_2): MaxPool2d(kernel_size=4, stride=4, padding=0, dilation=1, ceil_mode=False)
      (dropout_2): Dropout(p=0.1)
    )
    (recur): LSTM(128, 64, num_layers=2)
    (dense): Sequential(
      (dropout_3): Dropout(p=0.3)
      (batchnorm1d_0): BatchNorm1d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (linear_0): Linear(in_features=64, out_features=10, bias=True)
    )
  )
)
Trainable parameters: 139786

Usage

Inference

Run inference on an audio file:

./run.py /path/to/audio/file.wav -r path/to/saved/model.pth 

Training

./run.py train -c config.json --cfg arch.cfg

Augmentation

Dataset transforms:

Compose(
    ProcessChannels(mode=avg)
    AdditiveNoise(prob=0.3, sig=0.001, dist_type=normal)
    RandomCropLength(prob=0.4, sig=0.25, dist_type=half)
    ToTensorAudio()
)

As well as time stretching:

TensorboardX

Evaluation

./run.py eval -r /path/to/saved/model.pth

Then obtain defined metrics:

100%|█████████████████████████████████████████████████████████████████████████████████████████████████| 34/34 [00:03<00:00, 12.68it/s]
{'avg_precision': '0.725', 'avg_recall': '0.719', 'accuracy': '0.804'}

10-Fold Cross Validation

Per fold metrics CRNN(Bidirectional, Dropout):

To Do

• commit jupyter notebook dataset exploration
• Switch overt to using pytorch/audio
• use torchaudio-contrib for STFT transforms
• CRNN entirely defined in .cfg
• Some bug in 'infer'
• Run 10-fold Cross Validation
• Switch over to pytorch/audio since the merge
• Comment things