{ "id": 56172, "title": "computation power", "url": "/competitions/trackml-particle-identification/discussion/56172", "author_name": "Jacek Poplawski", "post_date": "2018-05-07T09:36:23.701000", "votes": 1, "comment_count": 6, "views": 0, "content": "

This competition looks very cool, but I wonder - can we even start without big cluster of computers? It looks very computation intensive task. What are your thoughts?

", "messages": [ { "id": 324415, "postDate": "2018-05-07T17:20:25.900Z", "content": "

You can definitely start with a standard computer, for instance with the DB scan kernel. For going further, it all depends on the method you choose. Two important ideas:

\n\n", "rawMarkdown": "You can definitely start with a standard computer, for instance with the DB scan kernel. For going further, it all depends on the method you choose. Two important ideas: \n \n\n - The dataset is very large. It does not mean that you have to use all\n of it! For instance, if you plan to learn a model, you might be able\n to learn it only from a fraction of the data. It's up to you. \n - The ultimate goal is to have fast inference (applying the model to the\n test data) on standard processors, although in the accuracy phase,\n there is no consideration of speed.\n", "votes": 3, "replies": [ { "id": 324420, "postDate": "2018-05-07T17:24:12.527Z", "content": "

so there is still a hope :) thank you

", "rawMarkdown": "so there is still a hope :) thank you" } ] }, { "id": 325363, "postDate": "2018-05-08T10:52:19.637Z", "content": "

@Jacek, I think, 10 events for training is enough to reach the score ~0.6, so we can start with one computer and a few good ideas.

", "rawMarkdown": "@Jacek, I think, 10 events for training is enough to reach the score ~0.6, so we can start with one computer and a few good ideas. ", "votes": 4, "replies": [ { "id": 325366, "postDate": "2018-05-08T10:53:54.040Z", "content": "

Thanks Grzegorz - and congratulations on your LB score :)

", "rawMarkdown": "Thanks Grzegorz - and congratulations on your LB score :)" }, { "id": 330743, "postDate": "2018-05-19T15:52:48.283Z", "rawMarkdown": "", "isDeleted": true }, { "id": 330841, "postDate": "2018-05-19T20:32:48.217Z", "content": "

@Ram, My score 0.42 (~0.2 above the benchmark) was obtained using one personal computer, one simple idea, one event and run time below one hour, so I think the score 0.6 was rather a safe approximation for one computer, a few good ideas and ten events. The emphasis is still on the number and the quality of ideas (feature engineering) and not on the number of events.

", "rawMarkdown": "@Ram, My score 0.42 (~0.2 above the benchmark) was obtained using one personal computer, one simple idea, one event and run time below one hour, so I think the score 0.6 was rather a safe approximation for one computer, a few good ideas and ten events. The emphasis is still on the number and the quality of ideas (feature engineering) and not on the number of events.", "votes": 7 } ] }, { "id": 324157, "postDate": "2018-05-07T09:36:23.700Z", "content": "

This competition looks very cool, but I wonder - can we even start without big cluster of computers? It looks very computation intensive task. What are your thoughts?

", "rawMarkdown": "This competition looks very cool, but I wonder - can we even start without big cluster of computers? It looks very computation intensive task. What are your thoughts?", "votes": 1 } ], "comments": [ { "id": 324415, "author_name": "CecileGermain", "author_url": "", "post_date": "2018-05-07T17:20:25.900000", "content": "

You can definitely start with a standard computer, for instance with the DB scan kernel. For going further, it all depends on the method you choose. Two important ideas:

\n\n", "votes": 3, "replies": [ { "id": 324420, "author_name": "Jacek Poplawski", "author_url": "", "post_date": "2018-05-07T17:24:12.527000", "content": "

so there is still a hope :) thank you

", "votes": 0, "replies": [] } ] }, { "id": 325363, "author_name": "Grzegorz Sionkowski", "author_url": "", "post_date": "2018-05-08T10:52:19.637000", "content": "

@Jacek, I think, 10 events for training is enough to reach the score ~0.6, so we can start with one computer and a few good ideas.

", "votes": 4, "replies": [ { "id": 325366, "author_name": "Jacek Poplawski", "author_url": "", "post_date": "2018-05-08T10:53:54.040000", "content": "

Thanks Grzegorz - and congratulations on your LB score :)

", "votes": 0, "replies": [] }, { "id": 330743, "author_name": "", "author_url": "", "post_date": "2018-05-19T15:52:48.283000", "content": "", "votes": 0, "replies": [] }, { "id": 330841, "author_name": "Grzegorz Sionkowski", "author_url": "", "post_date": "2018-05-19T20:32:48.217000", "content": "

@Ram, My score 0.42 (~0.2 above the benchmark) was obtained using one personal computer, one simple idea, one event and run time below one hour, so I think the score 0.6 was rather a safe approximation for one computer, a few good ideas and ten events. The emphasis is still on the number and the quality of ideas (feature engineering) and not on the number of events.

", "votes": 7, "replies": [] } ] } ], "raw_markdown_by_id": { "324415": "You can definitely start with a standard computer, for instance with the DB scan kernel. For going further, it all depends on the method you choose. Two important ideas: \n \n\n - The dataset is very large. It does not mean that you have to use all\n of it! For instance, if you plan to learn a model, you might be able\n to learn it only from a fraction of the data. It's up to you. \n - The ultimate goal is to have fast inference (applying the model to the\n test data) on standard processors, although in the accuracy phase,\n there is no consideration of speed.\n", "325363": "@Jacek, I think, 10 events for training is enough to reach the score ~0.6, so we can start with one computer and a few good ideas. ", "324157": "This competition looks very cool, but I wonder - can we even start without big cluster of computers? It looks very computation intensive task. What are your thoughts?" } }