Tensorflow2 implementation of Data-driven Harmonic Filters for Audio Representation Learning

Data-driven Harmonic Filters for Audio Representation Learning

Minz Won, Sanghyuk Chun, Oriol Nieto, and Xavier Serra

ICASSP, 2020

Reference

• pdf
• pytorch

Prerequisited

• Tensorflow (>=2.2)
• Kapre : pip install kapre , for more information

Usage

• Requirements

$ conda env create -n {ENV_NAME} --file environment.yaml
$ conda activate {ENV_NAME}

• Preprocessing

$ python -u preprocess.py run ../dataset
$ python -u split.py run ../dataset

• Training

$ python train.py

• Options

'--conv_channels', type=int, default=128
'--sample_rate', type=int, default=16000
'--n_fft', type=int, default=512
'--n_harmonic', type=int, default=6
'--semitone_scale', type=int, default=2
'--learn_bw', type=str, default='only_Q', choices=['only_Q', 'fix']
'--input_legnth', type=int, default=80000
'--batch_size', type=int, default=16
'--log_step', type=int, default=19
'--model_save_path', type=str, default='./../saved_models'
'--gpu', type=str, default='0'
'--data_path', type=str, default='./../../tf-harmonic-cnn/dataset'

To be changed

tf.keras.optimizers.adam -> tfa.optimizers.AdamW
tf.keras.optimizers.sgd -> tfa.optimizers.SGDW

Author

• Jaehwan Lee @jaehwlee
• jaehwlee@gmail.com