This is an official GitHub repository for the paper [Lee D, Aune E., Langet N. and Eidsvik J., 2022, "Ensemble and self-supervised learning for improved classification of seismic signals from the ̊Aknes rockslope"].
The paper is implemented in PyTorch in this repository.

Quick Start

🚀 Run the self-supervised learning (SSL) of VNIbCReg:

• python sslearn.py --config_ssl configs/ssl.yaml

🚀 Run i) training from scratch, or ii) fine-tuning, or iii) linear evaluation:

• python finetune.py --config_ft configs/finetune.yaml

🚀 Run the implementation of the relevant previous study [1] as described in our paper [2].

• python original_implementation/train.py --configs original_implementation/config.yaml
  

Note:

• configs/ssl.yaml is set to run VNIbCReg by default.
• configs/finetune.yaml is set to run training a model (ResNet34) from scratch on 80% of the dataset.
• original_implementation/config.yaml is set to run training a model (ResNet34) from scratch on 80% of the dataset following the implementation of the relevant previous study [1].
• All configuration files can be edited to suit your experimental purpose.
• The dataset is automatically downloaded once you run any of the above codes. The downloading code is defined in utils/dataset.py. Or, it can be found on https://doi.org/10.6084/m9.figshare.21340101.v1.

Configuration Files

Basic annotations are already available in the configuration files. Here, details of some parameters that might not be clear are explained.

🔧 ssl.yaml

• model_params

  • in_channels: input channel size
  • out_size_enc: output channel size from an encoder
  • out_size_enc: output channel size from an encoder
  • proj_hid: hidden size of the projector in VIbCReg
  • proj_out: output size of the projector in VIbCReg
  • backbone_type: a type of backbone. Available backbone types are ResNet18Encoder, ResNet34Encoder, ResNet50Encoder, and ResNet152Encoder. (for ResNet50Encoder, and ResNet152Encoder, out_size_enc needs to 2048.)

• exp_params

  • LR: learning rate
  • model_save_ep_period: a period for saving a model (epoch)

• trainer_params

  • gpus: indices for gpus to be used.

• dataset

  • num_workers: num_workers in torch.utils.data.DataLoader.
  • return_single_spectrogram_train: use of the ensemble prediction during training
  • return_single_spectrogram_train: use of the ensemble prediction during testing

🔧 finetune.yaml

Parameters that are described above are not specified.

• load_encoder

  • ckpt_fname: none for training from scratch. checkpoints/some_saved_model.ckpt for loading a pretrained encoder.

• exp_params

  • freeze_encoders: freeze the encoder if True to conduct linear evaluation
  • freeze_bn_stat_train: if True, encoder.eval() is set during training to use the averaged statistics of BatchNorm. This is useful when finetuning with a very-small dataset.

• dataset

  • train_data_ratio: can be adjusted to 0.05 or 0.1 for the fine-tuning evaluation in a small-dataset regime.

🔧 original_implementation/config.yaml

• backbone_with_clf_type: available backbone types are: AlexNet, ResNet18, and ResNet34

References

[1] Langet et al., 2022, "Automated classification of seismic signals recorded on the Åknes rockslope, Western Norway, using a Convolutional Neural Network"
[2] Lee et al., 2022, "Ensemble and self-supervised learning for improved classification of seismic signals from the ̊Aknes rockslope"