• Architecture
• Introduction
• Usage
  • Data preparation
  • Training
  • Testing
• License

Architecture

Introduction

We introduce a Label Noise Robust Learning (LNRL) method for handling label noise in microseismic tasks with small-scale datasets. LNRL aligns feature representation and label representation distribution in multiple feature spaces, learns the correlation between instances and label noise, and mitigates the impact of label noise.

The code of this project is modified based on SeisT.

Usage

Data Preparation

• For training and evaluation

  Create a new file named yourdata.py in the directory dataset/ to read the metadata and seismograms of the dataset. And you need to use @register_dataset decorator to register your dataset.

  (Please refer to the code samples datasets/sos.py)

Training

• Model
  Before starting training, please make sure that your model code is in the directory models/ and register it using the @register_model decorator. You can inspect the models available in the project using the following method:

  >>> from models import get_model_list
  >>> get_model_list()
  ['seist','lnrl']

• Model Configuration
  The configuration of the loss function and model labels is in config.py, and a more detailed explanation is provided in this file.

• Start training
  If you are training with a CPU or a single GPU, please use the following command to start training:

  python main.py \
    --seed 0 \
    --mode "train_test" \
    --model-name "lnrl" \
    --log-base "./logs" \
    --device "cuda:0" \
    --data "/root/data/Datasets/SOS" \
    --dataset-name "sos" \
    --sigma 600 \
    --data-split true \
    --train-size 0.8 \
    --val-size 0.1 \
    --shuffle true \
    --workers 8 \
    --in-samples 6000 \
    --augmentation true \
    --epochs 200 \
    --patience 30 \
    --batch-size 300

  If you are training with multiple GPUs, please use torchrun to start training:

  torchrun \
    --nnodes 1 \
    --nproc_per_node 2 \
    main.py \
      --seed 0 \
      --mode "train_test" \
      --model-name "lnrl" \
      --log-base "./logs" \
      --data "/root/data/Datasets/SOS" \
      --dataset-name "sos" \
      --sigma 600 \
      --data-split true \
      --train-size 0.8 \
      --val-size 0.1 \
      --shuffle true \
      --workers 8 \
      --in-samples 6000 \
      --augmentation true \
      --epochs 200 \
      --patience 30 \
      --batch-size 300

  There are also many other custom arguments, see main.py for more details.

Testing

If you are testing with a CPU or a single GPU, please use the following command to start testing:

python main.py \
  --seed 0 \
  --mode "test" \
  --model-name "lnrl" \
  --log-base "./logs" \
  --device "cuda:0" \
  --data "/root/data/Datasets/SOS" \
  --dataset-name "sos" \
  --data-split true \
  --train-size 0.8 \
  --val-size 0.1 \
  --workers 8 \
  --in-samples 6000 \
  --batch-size 300

If you are testing with multiple GPUs, please use torchrun to start testing:

torchrun \
  --nnodes 1 \
  --nproc_per_node 2 \
  main.py \
    --seed 0 \
    --mode "test" \
    --model-name "lnrl" \
    --log-base "./logs" \
    --data "/root/data/Datasets/SOS" \
    --dataset-name "sos" \
    --data-split true \
    --train-size 0.8 \
    --val-size 0.1 \
    --workers 8 \
    --in-samples 6000 \
    --batch-size 300

It should be noted that the train_size and val_size during testing must be consistent with that during training, and the seed must be consistent. Otherwise, the test results may be distorted.

License

Copyright S.Li et al. 2024. Licensed under an MIT license.