Kaggle-FAT

PyTorch implementation for Kaggle Freesound Audio Tagging 2019 Challenge

36th / 880 (top 5%)

Final leaderboard scores (label-weighted label-ranking average precision (lwlrap)):

0.72340 (private) / 0.715 (public)

Brief summary

• Using both curated and noisy data for training from scratch
  • Only creating 5 folds for curated data
  • Using all of noisy data for each fold (but using lower confidence weights)
• Raw audio to PCEN [1] (fixed parameters)
• Augmentations
  • Random crop (2 seconds)
  • mixup [2]
  • SpecAugment (frequency mask and time mask) [3]
• Small models
  • ResNet18 with CBAM [4]
  • MobileNetV3-Large [5]
• Losses
  • Binary cross-entropy loss
  • Lovász hinge loss [6]
• TTA: adaptive and overlapped crops (2 seconds) for each raw audio

Requirements

• pytorch 0.4.1
• torchvision 0.2.0
• librosa
• soundfile
• opencv
• pandas

Usage

Data pre-processing

• Modify the dataset path in config.json
• Run PYTHONPATH=. python loaders/freesound_loader.py to convert to melspectrogram (saved as .npy file)

To train the model

k=5
for((i=0;i<$k;i=i+1))
do 
python train.py --arch resnet18 --dataset freesound --split train \
                --img_rows 128 --img_cols 197 \
                --n_iter 40000 --batch_size 128 --seed 1234 \
                --l_rate 1e-1 --weight_decay 1e-4 --iter_size 1 \
                --num_cycles 0 --print_train_freq 100 --eval_freq 2000 \
                --fold_num $i --num_folds $k --sampling_rate 44100 \
                --dropout_rate 0.5 --gamma_fl 0.0 \
                --use_mix_up --use_spec_aug --use_cbam
done

python train.py -h for more details

To test the model

k=5
for((i=0;i<$k;i=i+1))
do
python test.py --model_path checkpoints/resnet18_freesound_best_128x197_44100_$i-$k_model.pth --dataset freesound \
               --img_rows 128 --img_cols 197 --seed 1234 \
               --batch_size 32 --split test --sampling_rate 44100 \
               --tta 8 --use_cuda --use_cbam
done

python test.py -h for more details

To create final submission

python merge.py --dataset freesound --img_rows 128 --img_cols 197 --seed 1234 --split test

python merge.py -h for more details

References

[1] Per-Channel Energy Normalization: Why and How

[2] mixup: Beyond Empirical Risk Minimization

[3] SpecAugment: A Simple Data Augmentation Method for Automatic Speech Recognition

[4] CBAM: Convolutional Block Attention Module

[5] Searching for MobileNetV3

[6] The Lovász-Softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks