TinyDAS - Tinygrad meets the PubDAS dataset

How to get data

See this pdf

Or sign up to GLOBUS and go here

Modules

The following is an explanation of the project structure

Dataset

Loads the HDF5 data in from the data folder and exports it to a pytorch esc dataset

Dataloader

Uses parallel workers to load single datafiles in parallel

Models

See examples in the tinydas/models folder

All autoencoders are based on the BaseAE class

Finding anomalies

Will upload jupyter notebooks soon

Hyperparameters

They are stored in yaml files under the configs directory. Name of the config is the name of the model in lowercase

How to run

python main.py -t train -m ae

or alternatively

python main.py -t detect -m ae

NOTES:

• Utils for loss scaling and clipping exist in this repo, but is kinda wonky for training certain models. However, F16 inference is easy:

1. Select model
2. Load model
3. model.half()