Several comments in my previous post on Gaussian process modelling said it was nice but too slow because the initial step requires a matrix inversion that takes O(n³) steps if there are n points in the light curve.
\n\nWell, this is not true if you use the celerite package. As noted by organizers in the kernels they shared with us at the start of this competition, celerite is way faster than other. Indeed, celerite inverts the matrix in O(n) steps!
\n\nAfter some tuning I can now compute max log likelihood Gaussian process for all train data in about 30 minutes with a single thread. This makes computation on test data affordable. Next is to fidn how to make a useful use of it ...
\n\nTo be clear, I am not saying GP is useful, I am just saying that computation speed is not really an issue. I wanted to make that point to give justice to celerite authors who did a remarkable job.
\n\ncelerite documentation: https://celerite.readthedocs.io/en/stable/
", "messages": [ { "id": "435696", "postDate": "12/08/2018 15:24:06", "content": "Several comments in my previous post on Gaussian process modelling said it was nice but too slow because the initial step requires a matrix inversion that takes O(n³) steps if there are n points in the light curve.
\n\nWell, this is not true if you use the celerite package. As noted by organizers in the kernels they shared with us at the start of this competition, celerite is way faster than other. Indeed, celerite inverts the matrix in O(n) steps!
\n\nAfter some tuning I can now compute max log likelihood Gaussian process for all train data in about 30 minutes with a single thread. This makes computation on test data affordable. Next is to fidn how to make a useful use of it ...
\n\nTo be clear, I am not saying GP is useful, I am just saying that computation speed is not really an issue. I wanted to make that point to give justice to celerite authors who did a remarkable job.
\n\ncelerite documentation: https://celerite.readthedocs.io/en/stable/
", "rawMarkdown": "Several comments in my previous post on Gaussian process modelling said it was nice but too slow because the initial step requires a matrix inversion that takes O(n³) steps if there are n points in the light curve.\n\nWell, this is not true if you use the celerite package. As noted by organizers in the kernels they shared with us at the start of this competition, celerite is way faster than other. Indeed, celerite inverts the matrix in O(n) steps!\n\nAfter some tuning I can now compute max log likelihood Gaussian process for all train data in about 30 minutes with a single thread. This makes computation on test data affordable. Next is to fidn how to make a useful use of it ...\n\nTo be clear, I am not saying GP is useful, I am just saying that computation speed is not really an issue. I wanted to make that point to give justice to celerite authors who did a remarkable job.\n\ncelerite documentation: https://celerite.readthedocs.io/en/stable/", "votes": null }, { "id": "435781", "postDate": "12/08/2018 18:54:50", "content": "Thank you for this comments
", "rawMarkdown": "Thank you for this comments", "votes": null } ], "comments": [ { "id": 435781, "author_name": "blondinka", "author_url": "", "post_date": "12/08/2018 18:54:50", "content": "Thank you for this comments
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "435696": "Several comments in my previous post on Gaussian process modelling said it was nice but too slow because the initial step requires a matrix inversion that takes O(n³) steps if there are n points in the light curve.\n\nWell, this is not true if you use the celerite package. As noted by organizers in the kernels they shared with us at the start of this competition, celerite is way faster than other. Indeed, celerite inverts the matrix in O(n) steps!\n\nAfter some tuning I can now compute max log likelihood Gaussian process for all train data in about 30 minutes with a single thread. This makes computation on test data affordable. Next is to fidn how to make a useful use of it ...\n\nTo be clear, I am not saying GP is useful, I am just saying that computation speed is not really an issue. I wanted to make that point to give justice to celerite authors who did a remarkable job.\n\ncelerite documentation: https://celerite.readthedocs.io/en/stable/", "435781": "Thank you for this comments" }, "source": "meta" }