Summary

Data preprocessing

Log-scaled mel-spectrograms is the modern standard way of the data representation in CNN-based audio scene classification. Audio config parameters:

sampling_rate = 32000
hop_length = 345 * 2
fmin = 20
fmax = sampling_rate // 2
n_mels = 128
n_fft = n_mels * 20
min_seconds = 10  #Seconds

Augmentations

Several augmentations were applied on spectrograms during the training stage using library audiomentations. lists augmentation techniques:

        train_aug = aaug.Compose([
            aaug.AddGaussianNoise(p=0.5),
            aaug.AddGaussianSNR(p=0.5),
            aaug.Gain(p=0.5),
            aaug.AddBackgroundNoise(sounds_path = noise_dir, p=0.5)
            #aaug.Normalize(p=0.3)
            #aaug.Shift(p=0.2)
            #aaug.PitchShift(min_semitones=-4, max_semitones=4, p=0.3),
            #aaug.TimeStretch(p = 0.3)
       ])

Training

• 5 stratifiedkfold
• Loss: BCE with auxiliary
• Optimizer: AdamW with initial LR 0.0001 and weight decay Wd 1e-5
• CosineAnnealingLR scheduler: factor 10
• Use ImbalanceData Sampler for sampling curated and noisy data
• Training with BCE on noisy samples with a high lwlrap score by previous models
• Mixed precision training

References

[1] PANNs: Large-Scale Pretrained Audio Neural Networks for Audio Pattern Recognition

[2] Introduction to Sound Event Detection

[3] Audiomentations Python library for audio augmentation

[4] PyTorch Image Models

[5] 6th place solution to the Cornell Birdcall Identification Challenge