Cornell Birdcall 36th place solution

https://www.kaggle.com/c/birdsong-recognition/overview

Transcript

1. 1D Features
   Resources:
   TITAN RTX, 1080Ti x 2, 2080Ti x 2 *1
   Cornell Birdcall Identification
   Kaggle
   XenoCant
   Extended
   XenoCant*2
   Copyright 2020 @ Maxwell_110
   *1 My room became tropical.
   *2 Credit to Vopani
   *3 Kerneler-kun had become
   a notebook expert! Check
   his profile.
   Feature Extraction
   Training / Prediction
   BirdVox
   ff1010
   fs2019
   ESC50
   264 birds
   nocall
   Trim silent parts
   (librosa.effects.trim)
   Load Audio
   (librosa.load)
   Resample with
   22.05 kHz
   remove only
   silent start/end parts
   Log Mel Spectrogram (2D)
   ( librosa.feature.melspectrogram
   librosa.core.power_to_db )
   Audio Data (1D)
    5 - 10 (s) variable audio length
    64 nmel, 10ms hop, 80ms sfft
    Event aware extraction
    5 (s) constant audio length
    Event aware extraction
   64
   500 - 1000
   5 x 22050
   1
   Augmentation
    p: 0.5
    width / height shift:
   0.2 / 0.1
    Scale: -0.05 / +0.05
   2
   Random Eraser
    p: 0.5
    erase num: 1
    width: [0, 0.1]
    height: [0.1, 0.3]
    fill with -1
   Standardize
   [- 1, + 1]
   To 2D Models
   To 1D Models
    p: 0.5
    width shift: 0.2
    NoiseInjection
   Augmentation Standardize
   [- 1, + 1]
   cut out
   at random
   2D
   Features
   ResNet 18
   nocall
   Binary Model
   (call / nocall )
   3
   4 Multi-Label Model (264 types)
   Multi-Task-Learning (MTL) for primary and noisy background labels
   call
   2D Models
   2048
   nodes
   +
   BN
   +
   ReLu
   512
   nodes
   +
   BN
   +
   ReLu
   2D: GAP
   /
   1D: MAP*4
   2048
   nodes
   +
   BN
   +
   ReLu
   512
   nodes
   +
   BN
   +
   ReLu
   2D: GAP
   /
   1D: MAP*4
   264
   nodes
   primary
   labels
   264
   nodes
   background
   labels
   ResNet 18
   2D Models
   MTL
   Loss
   3 stage scratch learning
   1. primary only
   -
   ,
   = [, ]
   - 2D: 200 epochs + Early Stopping (ES)
   Adam, CyclicLR 1e-4 ~ e-3
   - 1D: 100 epochs + ES
   SGD, CosineAnnealing 1e-1 ~ e-6
   => Adam, ReduceLROnP 1e-4
   2. + background
   -
   ,
   = [, ]
   - Adam (5e-5)
   - ReduceLROnP (x 0.25)
   3. + Psuedo Labeling
   -
   ,
   = [, ]
   - Adam (5e-5)
   - ReduceLROnP (x 0.25)
   - Predictions of backgrounds more
   than 0.15 are added to primary labels
   as soft labels in primary branch. All
   values are clipped between 0 and 1.
   primary branch
   background branch
   PANNs 1D
   1D Model
   Blending
   Public: 36 th (0.623)
   Private: 39 th (0.580)
   *3
   • Class-Wise Blending
   - For each bird class
   - Optimize blend coefficients with BCE loss
   • Class-Wise threshold optimization
   - For each bird class
   - Maximize macro-F1 (not sample-wise)
   • 3 Epoch ensemble for PANNs 1D
   • 5 Fold ensemble for ResNet18 , PANNs 1D
   Model Architecture
   *4 mean on time axis,
   max and mean on freq axis
   

   View Slide

2. 1 Event Aware Extraction
   Example: XC341516.mp3 (ebird_code: brespa, 60 sec)
   brespa
   bug
   Naïve Random Extraction (5 sec)
   waveform
   logmel
   logmel
   freq-wise
   max
   nocall, no event
   => mislabeling
   logmel freq-wise max
   aware extraction
   noisy event / call
   => sometimes works, other times fails.
   logmel freq-wise max
   aware extraction
   with removing lower frequencies
   bug
   brespa
    Using logmel spectrum, extract audio chunks which contain
   logmel frequency-wise max intensity over the threshold.
    Because almost of all Bird call frequencies range from 1 ~ 8 kHz,
   get the better chunks for training via removing lower frequencies (~ 300Hz).
    Variable extraction length, 5 - 10 sec, also helps to get a good signal,
   meanwhile 2D models are well generalized to variable length.
   threshold
   

   View Slide