/buzz-models

Repository for bee sound models training

Primary LanguagePython

smartula-analysis

Repository for seraching optimal bee-sound represenation with use of Autoencoders:

  • Vanilla Autoencoder (ae)
  • Convolutional Autoencoder (conv_ae)
  • Variational Autoencoder (vae)
  • Convolutional Autoencoder (conv_vae)
  • Contrastive Autoencoder (cvae)
  • Contrastive Convolutional Autoencoder (conv_cvae)

and sound representation features:

  • Periodogram (periodogram)
  • spectrogram (spectrogram)
  • MelSpectrogram (melspectrogram)
  • MFCC (mfcc)
  • Bioacustic indicies (indicies)
    • Acustic Complexity Index (aci - configured by config.json)
    • Acoustic Diversity Index (adi - configured by config.json)
    • Acoustic Evenness Index (aei - configured by config.json)
    • Bioacustic Index (bi - configured by config.json)

Training defined model

non contrastive

train.py scripts trains choosen architecture with use config from json file. E.g. command will invoke training Variational Autoencoder on periodogram data.

python train.py vae periodogram ..\\measurements\\smartulav2 --config_file=example_config.json

constrastive

user should specify target and background data through options

python train.py vae periodogram ..\\measurements\\smartulav2 --config_file=example_config.json --target smrpiclient6 --background smrpiclient3 smrpiclient7

also there is option to use discriminator in contrastive training by using --discriminator option