Jafagervik / TinyDAS

Tinygrad and PubDAS

TinyDAS - Tinygrad meets PubDAS

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

Generates an iterator for the dataset of a certain batchsize

Models

class MyAwesomeModel(BaseAE):
    def __init__(self, kwargs**):
        self.M = kwargs["data"]["M"]
        self.N = kwargs["data"]["N"]

        self.net = [Linear(M*N, 1), Tensor.tanh]

    def __call__(self, x: Tensor) -> Tuple[Tensor, ...]:
        """In certain models, a forward pass returns multiple values"""
        y = x.sequential(self.net)
        return y,

    @property
    def convolutional(self) -> bool:
        """If set to true, data needs to be reshaped in a different manner"""
        return False

    @abstractmethod
    def criterion(self, x: Tensor) -> Dict[str, Tensor]:
        """Stored as a dict in case you want to track several losses"""
        (y, ) = self(x)
        loss = mse(y, x)
        return {"loss": loss}

Most models are based on this abstractclass

Utils

Finding anomalies

Hyperparameters

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

epochs: 10
lr: 0.001
batch_size: 16

How to run

python main.py -t train -m ae

or alternatively

python main.py -t detect -m ae

NOTES: