Clarification on prediction_mask in MoiraiFinetune

[zqiao11]

Hi @gorold. May I clarify the setup of the prediction mask in validation dataset in MoiraiFinetune?

            + EvalMaskedPrediction(
                mask_length=-prediction_length % patch_size,  # num of patches for prediction?
                target_field="target",
                truncate_fields=("variate_id", "time_id", "observed_mask"),
                optional_truncate_fields=("past_feat_dynamic_real",),
                prediction_mask_field="prediction_mask",
                expected_ndim=3,
            )

Now mask_length is computed as -prediction_length % patch_size. Shouldn't the mask_length be computed as prediction_length // patch_size? As the logic here seems to mask the patches in the prediction range.

    def _generate_prediction_mask(
        self, target: Float[np.ndarray, "var time *feat"]
    ) -> Bool[np.ndarray, "var time"]:
        self.check_ndim("target", target, self.expected_ndim)
        var, time = target.shape[:2]
        prediction_mask = np.zeros((var, time), dtype=bool)
        prediction_mask[:, -self.mask_length :] = True
        return prediction_mask

Please correct me if my understanding is incorrect. Look forward to your reply.

[gorold]

Hey, thanks for catching this. I think you're right, it should be prediction_length // patch_size, will look into this.

[zqiao11]

Hi, it should be math.ceil(prediction_length / patch_size) since the prediction range has been padded.

[gorold]

I think there wouldn't be a difference if the prediction range is already padded?

[zqiao11]

Let's say if prediction length is 96 and patch size is 64, the prediction range will be padded to 128 to make sure there are multiple patches. So, the last 2 patches should be masked as True in prediction_mask.

If we use prediction_length // patch_size, then only the last patch is masked as True.

[gorold]

Right... this makes sense. Thanks a lot!