Isolate StandardScalar from load dataset

[akashshah59]

The current implementation of data.py/load_dataset() instantiates a standard scaler by default.

def load_dataset(dataset_dir, batch_size, val_batch_size=None, test_batch_size=None):
    if val_batch_size is None:
        val_batch_size = batch_size

    if test_batch_size is None:
        test_batch_size = batch_size

    data = {}

    for category in ["train", "val", "test"]:
        cat_data = np.load(os.path.join(dataset_dir, category + ".npz"))
        data["x_" + category] = cat_data["x"]
        data["y_" + category] = cat_data["y"]

    scaler = StandardScaler(data["x_train"][..., 0])

    for category in ["train", "val", "test"]:
        data["x_" + category][..., 0] = scaler.transform(data["x_" + category][..., 0])
        data["y_" + category][..., 0] = scaler.transform(data["y_" + category][..., 0])

    data_train = PaddedDataset(batch_size, data["x_train"], data["y_train"])
    data["train_loader"] = DataLoader(data_train, batch_size, shuffle=True)

    data_val = PaddedDataset(val_batch_size, data["x_val"], data["y_val"])
    data["val_loader"] = DataLoader(data_val, val_batch_size, shuffle=False)

    data_test = PaddedDataset(test_batch_size, data["x_test"], data["y_test"])
    data["test_loader"] = DataLoader(data_test, test_batch_size, shuffle=False)

    data["scaler"] = scaler
    return data

The goal is to be able to isolate the scalar from the data loading method, and support more scalars eventually.

[akashshah59]

@klane and @yuqirose Shouldn't scalar be a part of data preprocessing and not part of the model?

Currently, our forward step has a scaler implemented, however, does it make sense to have it defined outside the model step function by the user?

    def _step(self, batch, batch_idx, num_batches):
        x, y = self.prepare_batch(batch)

        if self.training:
            batches_seen = batch_idx + self.current_epoch * num_batches
        else:
            batches_seen = batch_idx

        pred = self(x, y, batches_seen)

        if self.scaler is not None:
            y = self.scaler.inverse_transform(y)
            pred = self.scaler.inverse_transform(pred)