catchMinor

• all about imbalanced-learning, anomaly-detection (tabular, time series, graph)
• focus on developing models based on torch, lightning

Example Code

import numpy as np
import torch
from pytorch_lightning import Trainer
from pytorch_lightning.loggers import TensorBoardLogger
from torch.utils.data import DataLoader

from catchMinor.data_load.dataset import tabularDataset
from catchMinor.tabular_model.AutoEncoder.ae_config import (
    AutoEncoder_config,
    AutoEncoder_loss_func_config,
    AutoEncoder_optimizer_config,
)
from catchMinor.tabular_model.AutoEncoder.lit_ae import LitBaseAutoEncoder
from catchMinor.utils.data import normal_only_train_split_tabular

# set data
data = np.load(data_path)
X, y = data["X"], data["y"]

(
    normal_X_train,
    mix_X_test,
    normal_y_train,
    mix_y_test,
) = normal_only_train_split_tabular(X, y, 0.8)

train_dataset = tabularDataset(normal_X_train, deepcopy(normal_X_train))
valid_dataset = tabularDataset(mix_X_test, deepcopy(mix_X_test))

train_loader = DataLoader(train_dataset, batch_size=512)
valid_loader = DataLoader(valid_dataset, batch_size=512)

# config
model_config = AutoEncoder_config(features_dim_list=[9, 4])
optim_config = AutoEncoder_optimizer_config()
loss_func_config = AutoEncoder_loss_func_config(loss_fn="MSELoss")

# Lit model
model = LitBaseAutoEncoder(model_config, optim_config, loss_func_config)

# trainer
TensorBoard_logger = TensorBoardLogger(
    save_dir="./log", name="AutoEncoder", version="0.1"
)

early_stopping_callback = EarlyStopping(monitor="valid_loss", mode="min", patience=2)

trainer = Trainer(
    log_every_n_steps=1,
    accelerator=config.cuda,
    logger=TensorBoard_logger,
    max_epochs=config.epochs,
    deterministic=True,
    callbacks=[early_stopping_callback],
    check_val_every_n_epoch=1,
)

# fit the model
trainer.fit(model, train_loader, valid_loader)

Installation (not yet)

• recommend to use pyenv virtualenv

python -m pip install catchMinor

Benchmark

• tabluar data: ADBench
  • (paper) ADBench: Anomaly Detection Benchmark
• time series data: UCR Time Series Anomaly Archive
  • (paper) Current Time Series Anomaly Detection Benchmarks are Flawed and are Creating the Illusion of Progress

Implemented Algorithms

Anomaly Detection

model	data	desc
AutoEncoder	tabular, time series	linear, reconstruction-based
VAE	tabular, time series	linear, reconstruction-based
GAN	tabular, time series	linear, reconstruction-based
AnomalyTransformer	time series	transformer
DLienar	time series	linear, pred-based
NLienar	time series	linear, pred-based

Imbalanced Learning

Contribute

• follow gitflow & forkflow
• branch
  • master: production branch
  • develop: develop branch
  • feature/{work}: feature branch
  • release/{version}: temporary branch before release
  • hotfix: bugfix branch based on master branch
• merge (main contributor)
  • merge from feature branch to develop branch: merge squash
  • else: merge --no-ff
• example

# 1. fork repository
# 2. clone your origin and add remote upstream
git clone {your_origin_repo}
git add remote upstream {catchMinor_origin_repo}
# 3. make your origin develop branch as default branch
# 4. make feature branch
git checkout -b feature/{your_work} develop
# 5. do something to contribute
# 6. add, commit (use cz c)
git add {your_work}
git commit -m "{your_work_message}"
# 7. push to your origin (develop branch)
git push origin feature/{your_work}
# 8. PR in github