Gerald Woo

Hi @cchallu, this is a follow up to issue #8. I was also wondering if it would be better to use the last value of the lookback window instead of...

calculate_seasonal_error bug for multivariate data

5

## Description ```calculate_seasonal_error``` is buggy when the inputs are multivariate data. ```past_data = extract_past_data(time_series, forecast)``` results in ```(dim, time)``` shaped array rather than a ```(time,)``` shaped array, which is then...

history_size/window_sampling_limit is shorter than input size

1

For certain datasets, e.g. Yearly/Quarterly/Monthly M4 datasets, the quantity `history_size` is set to 1.5, leading `window_sampling_limit` to be 1.5 times of the horizon length. Yet, the input size could be...

fix:Moirai inference in foundation-time-series-arena

4

Thanks for evaluating Moirai. There are some issues with the way inference is being done for Moirai: 1. Using the mean of sample forecasts instead of median. In GluonTS, metrics...

Update README and config files with Moirai-1.1-R and fix Moirai-1.0-R model weights

4

Made PRs on huggingface to fix the regression of Moirai-1.0-R model weights with the recent code changes https://huggingface.co/Salesforce/moirai-1.0-R-small/discussions/7 https://huggingface.co/Salesforce/moirai-1.0-R-base/discussions/6 https://huggingface.co/Salesforce/moirai-1.0-R-large/discussions/6 Please help to test, I only managed to run some...

Issue with BFGS + vmap + diagonal

4

I ran into a bug when using BFGS + vmap + diagonal functions. As can be seen from the minimal reproducible example, using another optimizer like NealderMead works fine. I...

Able to override concrete dunder methods

2

I noticed that I was able to override the `__matmul__` method for lineax's AbstractLinearOperator class, even when strict=True. If I understand correctly, the abstract/final pattern should not allow this? ```python...

The marginal log likelihoods from the recursive filter and parallel filter are different. The below example makes a small modification on the test case from `parallel_inference_test.py` by setting µ0[0] =...

I'm curious why the implementation computes the update step using this approach: https://github.com/probml/dynamax/blob/800fae691edc7a372605a230d91344bd4420fd93/dynamax/linear_gaussian_ssm/inference.py#L303 Is there an advantage in numerical stability or some other advantage in doing so? Would it not...