I really like this competition and have been in it since the beginning. The dataset is collected from Hangzhou, the city where I spent my best years and met my wife. I saw the malls I've been to and my favorite restaurants when processing the data. It's a wonderful feeling. Hope the pandemic could end soon and people around the world could travel freely. Hangzhou is amazing. I recommend it to everyone.
\nI would like to give some tips if you are new to kaggle. Hope you find them useful.
\nDo not tune the models based on the leaderboard score. The public leaderboard has 15% test data and that's less than 1% of the training data. The cross-validation is much more representative of the model's performance.
Use the leaderboard as a way of debugging. If the CV score improves and the LB score worsens, do not jump to the conclusion \"just trust the CV\". There are so many steps in the post-processing pipeline, maybe something goes wrong with test data.
Study the public kernels but don't use the results as submissions. The kernels are wonderful and have huge potentials. But unfortunately, the endless dependencies are out of control. It would be much safer to run the code line by line and watch how it affects the CV score.
Improve the efficiency, I spent 20 days refactoring the code just to speed up the run time end-to-end without any submissions. I recommend the FeatureStore which is very fast to reload the part needed instead of reading the whole thing.
Training base models are painfully slow. Start early and don't give up on it. You can always study post-processing on the side. It is more likely that some good post-processing kernels (rather than good base models) are shared in the final weeks. If you have good base models, you will benefit from such sharing instead of getting hurt.
There are always randomness and luck. Be prepared that the results could be disappointing despite every effort we put in.
Last but not least, I would like to thank the top teams for sending me invitations. Very flattered. 😊
", "messages": [ { "id": 1277803, "postDate": "2021-04-19T08:32:40.900Z", "content": "I really like this competition and have been in it since the beginning. The dataset is collected from Hangzhou, the city where I spent my best years and met my wife. I saw the malls I've been to and my favorite restaurants when processing the data. It's a wonderful feeling. Hope the pandemic could end soon and people around the world could travel freely. Hangzhou is amazing. I recommend it to everyone.
\nI would like to give some tips if you are new to kaggle. Hope you find them useful.
\nDo not tune the models based on the leaderboard score. The public leaderboard has 15% test data and that's less than 1% of the training data. The cross-validation is much more representative of the model's performance.
Use the leaderboard as a way of debugging. If the CV score improves and the LB score worsens, do not jump to the conclusion \"just trust the CV\". There are so many steps in the post-processing pipeline, maybe something goes wrong with test data.
Study the public kernels but don't use the results as submissions. The kernels are wonderful and have huge potentials. But unfortunately, the endless dependencies are out of control. It would be much safer to run the code line by line and watch how it affects the CV score.
Improve the efficiency, I spent 20 days refactoring the code just to speed up the run time end-to-end without any submissions. I recommend the FeatureStore which is very fast to reload the part needed instead of reading the whole thing.
Training base models are painfully slow. Start early and don't give up on it. You can always study post-processing on the side. It is more likely that some good post-processing kernels (rather than good base models) are shared in the final weeks. If you have good base models, you will benefit from such sharing instead of getting hurt.
There are always randomness and luck. Be prepared that the results could be disappointing despite every effort we put in.
Last but not least, I would like to thank the top teams for sending me invitations. Very flattered. 😊
", "rawMarkdown": "I really like this competition and have been in it since the beginning. The dataset is collected from Hangzhou, the city where I spent my best years and met my wife. I saw the malls I've been to and my favorite restaurants when processing the data. It's a wonderful feeling. Hope the pandemic could end soon and people around the world could travel freely. Hangzhou is amazing. I recommend it to everyone.\n\nI would like to give some tips if you are new to kaggle. Hope you find them useful.\n\n1. Do not tune the models based on the leaderboard score. The public leaderboard has 15% test data and that's less than 1% of the training data. The cross-validation is much more representative of the model's performance.\n\n2. Use the leaderboard as a way of debugging. If the CV score improves and the LB score worsens, do not jump to the conclusion \"just trust the CV\". There are so many steps in the post-processing pipeline, maybe something goes wrong with test data. \n\n3. Study the public kernels but don't use the results as submissions. The kernels are wonderful and have huge potentials. But unfortunately, the endless dependencies are out of control. It would be much safer to run the code line by line and watch how it affects the CV score.\n\n4. Improve the efficiency, I spent 20 days refactoring the code just to speed up the run time end-to-end without any submissions. I recommend the [FeatureStore] (https://www.kaggle.com/kenmatsu4/feature-store-for-indoor-location-navigation) which is very fast to reload the part needed instead of reading the whole thing.\n\n5. Training base models are painfully slow. Start early and don't give up on it. You can always study post-processing on the side. It is more likely that some good post-processing kernels (rather than good base models) are shared in the final weeks. If you have good base models, you will benefit from such sharing instead of getting hurt.\n\n6. There are always randomness and luck. Be prepared that the results could be disappointing despite every effort we put in. \n\nLast but not least, I would like to thank the top teams for sending me invitations. Very flattered. 😊", "votes": 76 }, { "id": 1277998, "postDate": "2021-04-19T13:22:33.687Z", "content": "5 <- agree my model takes more than 1month to train 5fold, still under training :<
", "rawMarkdown": "5 <- agree my model takes more than 1month to train 5fold, still under training :<", "votes": 4 }, { "id": 1278646, "postDate": "2021-04-20T06:21:13.750Z", "content": "You are always the shining beacon of rookie kagglers
", "rawMarkdown": "You are always the shining beacon of rookie kagglers", "votes": 1 }, { "id": 1277821, "postDate": "2021-04-19T09:02:22.253Z", "content": "Thank you. You are very smart, focused, hard working and . Learn from you!
", "rawMarkdown": "Thank you. You are very smart, focused, hard working and . Learn from you!", "votes": 2 }, { "id": 1281108, "postDate": "2021-04-22T16:23:48.577Z", "content": "Thanks for sharing! Would you mind to share your base models' LB score without any post-processing?
", "rawMarkdown": "Thanks for sharing! Would you mind to share your base models' LB score without any post-processing?", "replies": [ { "id": 1281412, "postDate": "2021-04-22T23:12:42.233Z", "content": "I replied here. https://www.kaggle.com/c/indoor-location-navigation/discussion/232995#1281411
", "rawMarkdown": "I replied here. https://www.kaggle.com/c/indoor-location-navigation/discussion/232995#1281411", "votes": 1 }, { "id": 1286280, "postDate": "2021-04-27T18:53:27.767Z", "rawMarkdown": "", "votes": 1, "isDeleted": true } ] }, { "id": 1278917, "postDate": "2021-04-20T12:16:19.580Z", "content": "Thanks! friend
", "rawMarkdown": "Thanks! friend", "replies": [ { "id": 1286281, "postDate": "2021-04-27T18:53:59.250Z", "rawMarkdown": "", "votes": -2, "isDeleted": true }, { "id": 1286283, "postDate": "2021-04-27T18:54:11.113Z", "rawMarkdown": "", "votes": -2, "isDeleted": true } ] }, { "id": 1284604, "postDate": "2021-04-26T06:11:20.323Z", "rawMarkdown": "", "isDeleted": true }, { "id": 1303201, "postDate": "2021-05-12T00:41:07.130Z", "content": "Thank you. It's quite insightful.
", "rawMarkdown": "Thank you. It's quite insightful. ", "votes": 1 }, { "id": 1278619, "postDate": "2021-04-20T05:44:08.103Z", "content": "thanks for sharing tips !
", "rawMarkdown": "thanks for sharing tips !", "votes": 1 }, { "id": 1278076, "postDate": "2021-04-19T14:41:54.390Z", "content": "Thanks. I will follow your guideline.
", "rawMarkdown": "Thanks. I will follow your guideline.", "votes": 1 } ], "comments": [ { "id": 1277998, "author_name": "Ryunosuke Ishizaki", "author_url": "", "post_date": "2021-04-19T13:22:33.687000", "content": "5 <- agree my model takes more than 1month to train 5fold, still under training :<
", "votes": 4, "replies": [] }, { "id": 1278646, "author_name": "Medivh", "author_url": "", "post_date": "2021-04-20T06:21:13.750000", "content": "You are always the shining beacon of rookie kagglers
", "votes": 1, "replies": [] }, { "id": 1277821, "author_name": "Max2020", "author_url": "", "post_date": "2021-04-19T09:02:22.253000", "content": "Thank you. You are very smart, focused, hard working and . Learn from you!
", "votes": 2, "replies": [] }, { "id": 1281108, "author_name": "Ethan", "author_url": "", "post_date": "2021-04-22T16:23:48.577000", "content": "Thanks for sharing! Would you mind to share your base models' LB score without any post-processing?
", "votes": 0, "replies": [ { "id": 1281412, "author_name": "Jiwei Liu", "author_url": "", "post_date": "2021-04-22T23:12:42.233000", "content": "I replied here. https://www.kaggle.com/c/indoor-location-navigation/discussion/232995#1281411
", "votes": 1, "replies": [] }, { "id": 1286280, "author_name": "", "author_url": "", "post_date": "2021-04-27T18:53:27.767000", "content": "", "votes": 1, "replies": [] } ] }, { "id": 1278917, "author_name": "Santosh kumar", "author_url": "", "post_date": "2021-04-20T12:16:19.580000", "content": "Thanks! friend
", "votes": 0, "replies": [ { "id": 1286281, "author_name": "", "author_url": "", "post_date": "2021-04-27T18:53:59.250000", "content": "", "votes": -2, "replies": [] }, { "id": 1286283, "author_name": "", "author_url": "", "post_date": "2021-04-27T18:54:11.113000", "content": "", "votes": -2, "replies": [] } ] }, { "id": 1284604, "author_name": "", "author_url": "", "post_date": "2021-04-26T06:11:20.323000", "content": "", "votes": 0, "replies": [] }, { "id": 1303201, "author_name": "jongyoon", "author_url": "", "post_date": "2021-05-12T00:41:07.130000", "content": "Thank you. It's quite insightful.
", "votes": 1, "replies": [] }, { "id": 1278619, "author_name": "Sidney Ng", "author_url": "", "post_date": "2021-04-20T05:44:08.103000", "content": "thanks for sharing tips !
", "votes": 1, "replies": [] }, { "id": 1278076, "author_name": "mamas", "author_url": "", "post_date": "2021-04-19T14:41:54.390000", "content": "Thanks. I will follow your guideline.
", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "1277803": "I really like this competition and have been in it since the beginning. The dataset is collected from Hangzhou, the city where I spent my best years and met my wife. I saw the malls I've been to and my favorite restaurants when processing the data. It's a wonderful feeling. Hope the pandemic could end soon and people around the world could travel freely. Hangzhou is amazing. I recommend it to everyone.\n\nI would like to give some tips if you are new to kaggle. Hope you find them useful.\n\n1. Do not tune the models based on the leaderboard score. The public leaderboard has 15% test data and that's less than 1% of the training data. The cross-validation is much more representative of the model's performance.\n\n2. Use the leaderboard as a way of debugging. If the CV score improves and the LB score worsens, do not jump to the conclusion \"just trust the CV\". There are so many steps in the post-processing pipeline, maybe something goes wrong with test data. \n\n3. Study the public kernels but don't use the results as submissions. The kernels are wonderful and have huge potentials. But unfortunately, the endless dependencies are out of control. It would be much safer to run the code line by line and watch how it affects the CV score.\n\n4. Improve the efficiency, I spent 20 days refactoring the code just to speed up the run time end-to-end without any submissions. I recommend the [FeatureStore] (https://www.kaggle.com/kenmatsu4/feature-store-for-indoor-location-navigation) which is very fast to reload the part needed instead of reading the whole thing.\n\n5. Training base models are painfully slow. Start early and don't give up on it. You can always study post-processing on the side. It is more likely that some good post-processing kernels (rather than good base models) are shared in the final weeks. If you have good base models, you will benefit from such sharing instead of getting hurt.\n\n6. There are always randomness and luck. Be prepared that the results could be disappointing despite every effort we put in. \n\nLast but not least, I would like to thank the top teams for sending me invitations. Very flattered. 😊", "1277998": "5 <- agree my model takes more than 1month to train 5fold, still under training :<", "1278646": "You are always the shining beacon of rookie kagglers", "1277821": "Thank you. You are very smart, focused, hard working and . Learn from you!", "1281108": "Thanks for sharing! Would you mind to share your base models' LB score without any post-processing?", "1278917": "Thanks! friend", "1284604": "", "1303201": "Thank you. It's quite insightful. ", "1278619": "thanks for sharing tips !", "1278076": "Thanks. I will follow your guideline." } }