I applied pseudo-labeling and it was good on the validation dataset that comes from labeled data. However, it gave me a slightly lower score in the public leaderboard. Does it mean that it can be good in the private leaderboard?\nAre there any tricks while applying pseudo-labeling?\nThanks
", "messages": [ { "id": "617593", "postDate": "09/04/2019 09:38:34", "content": "I applied pseudo-labeling and it was good on the validation dataset that comes from labeled data. However, it gave me a slightly lower score in the public leaderboard. Does it mean that it can be good in the private leaderboard?\nAre there any tricks while applying pseudo-labeling?\nThanks
", "rawMarkdown": "I applied pseudo-labeling and it was good on the validation dataset that comes from labeled data. However, it gave me a slightly lower score in the public leaderboard. Does it mean that it can be good in the private leaderboard?\nAre there any tricks while applying pseudo-labeling?\nThanks", "votes": null }, { "id": "617603", "postDate": "09/04/2019 09:56:32", "content": "it can be good for private leaderboard
", "rawMarkdown": "it can be good for private leaderboard", "votes": null }, { "id": "617621", "postDate": "09/04/2019 10:18:17", "content": "So , do you usually submit two final versions, one with pseudo-labeling and one without it?
", "rawMarkdown": "So , do you usually submit two final versions, one with pseudo-labeling and one without it?", "votes": null }, { "id": "617765", "postDate": "09/04/2019 12:55:56", "content": "How do you apply pseudo-label in this kind of image competition? Thanks.
", "rawMarkdown": "How do you apply pseudo-label in this kind of image competition? Thanks.", "votes": null }, { "id": "617796", "postDate": "09/04/2019 13:46:19", "content": "@mariammohamed yeah after reading this discussion post,i am thinking about it a lot
", "rawMarkdown": "mariammohamed yeah after reading this[ discussion post](https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/105930#latest-615302),i am thinking about it a lot", "votes": null }, { "id": "617848", "postDate": "09/04/2019 14:33:48", "content": "I am not sure if my approach is the best, \nbut I split the test into chunks, \n1. predict labels for a chunk and merge it with the training data\n2. train the model\n3. validate on labeled data\n3. repeat 1-2-3 steps
", "rawMarkdown": "I am not sure if my approach is the best, \nbut I split the test into chunks, \n1. predict labels for a chunk and merge it with the training data\n2. train the model\n3. validate on labeled data\n3. repeat 1-2-3 steps", "votes": null }, { "id": "617850", "postDate": "09/04/2019 14:36:11", "content": "Also, have a look at this discussion post, suggested by @mobassir. It is more useful
", "rawMarkdown": "Also, have a look at this [discussion post](https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/105930#latest-615302), suggested by @mobassir. It is more useful", "votes": null }, { "id": "617917", "postDate": "09/04/2019 16:04:01", "content": "Thanks for the explanation!
", "rawMarkdown": "Thanks for the explanation!", "votes": null }, { "id": "618025", "postDate": "09/04/2019 18:30:46", "content": "You are welcome
", "rawMarkdown": "You are welcome", "votes": null }, { "id": "618871", "postDate": "09/05/2019 15:19:58", "content": "I've had varying success with pseudo-labelling in the past. Often it comes down to what is an acceptable label. I've turned to using a bit more active learning in practice. I think that the discussion posting that @mobassir posted is very enlightening as well. There is a lot of good ideas there I think.
\n\nAlso, it is early and never hurts to try different things. Rely on your CV and not the public leaderboard though, it is much more likely that will help you on the private side.
", "rawMarkdown": "I've had varying success with pseudo-labelling in the past. Often it comes down to what is an acceptable label. I've turned to using a bit more active learning in practice. I think that the discussion posting that @mobassir posted is very enlightening as well. There is a lot of good ideas there I think. \n\nAlso, it is early and never hurts to try different things. Rely on your CV and not the public leaderboard though, it is much more likely that will help you on the private side.", "votes": null } ], "comments": [ { "id": 617603, "author_name": "mobassir", "author_url": "", "post_date": "09/04/2019 09:56:32", "content": "it can be good for private leaderboard
", "votes": null, "replies": [ { "id": 617621, "author_name": "mariammohamed", "author_url": "", "post_date": "09/04/2019 10:18:17", "content": "So , do you usually submit two final versions, one with pseudo-labeling and one without it?
", "votes": null, "replies": [] }, { "id": 617796, "author_name": "mobassir", "author_url": "", "post_date": "09/04/2019 13:46:19", "content": "@mariammohamed yeah after reading this discussion post,i am thinking about it a lot
", "votes": null, "replies": [] } ] }, { "id": 617765, "author_name": "gogo827jz", "author_url": "", "post_date": "09/04/2019 12:55:56", "content": "How do you apply pseudo-label in this kind of image competition? Thanks.
", "votes": null, "replies": [ { "id": 617848, "author_name": "mariammohamed", "author_url": "", "post_date": "09/04/2019 14:33:48", "content": "I am not sure if my approach is the best, \nbut I split the test into chunks, \n1. predict labels for a chunk and merge it with the training data\n2. train the model\n3. validate on labeled data\n3. repeat 1-2-3 steps
", "votes": null, "replies": [] }, { "id": 617850, "author_name": "mariammohamed", "author_url": "", "post_date": "09/04/2019 14:36:11", "content": "Also, have a look at this discussion post, suggested by @mobassir. It is more useful
", "votes": null, "replies": [] }, { "id": 617917, "author_name": "gogo827jz", "author_url": "", "post_date": "09/04/2019 16:04:01", "content": "Thanks for the explanation!
", "votes": null, "replies": [] }, { "id": 618025, "author_name": "mariammohamed", "author_url": "", "post_date": "09/04/2019 18:30:46", "content": "You are welcome
", "votes": null, "replies": [] } ] }, { "id": 618871, "author_name": "cbryant", "author_url": "", "post_date": "09/05/2019 15:19:58", "content": "I've had varying success with pseudo-labelling in the past. Often it comes down to what is an acceptable label. I've turned to using a bit more active learning in practice. I think that the discussion posting that @mobassir posted is very enlightening as well. There is a lot of good ideas there I think.
\n\nAlso, it is early and never hurts to try different things. Rely on your CV and not the public leaderboard though, it is much more likely that will help you on the private side.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "617593": "I applied pseudo-labeling and it was good on the validation dataset that comes from labeled data. However, it gave me a slightly lower score in the public leaderboard. Does it mean that it can be good in the private leaderboard?\nAre there any tricks while applying pseudo-labeling?\nThanks", "617603": "it can be good for private leaderboard", "617621": "So , do you usually submit two final versions, one with pseudo-labeling and one without it?", "617765": "How do you apply pseudo-label in this kind of image competition? Thanks.", "617796": "mariammohamed yeah after reading this[ discussion post](https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/105930#latest-615302),i am thinking about it a lot", "617848": "I am not sure if my approach is the best, \nbut I split the test into chunks, \n1. predict labels for a chunk and merge it with the training data\n2. train the model\n3. validate on labeled data\n3. repeat 1-2-3 steps", "617850": "Also, have a look at this [discussion post](https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/105930#latest-615302), suggested by @mobassir. It is more useful", "617917": "Thanks for the explanation!", "618025": "You are welcome", "618871": "I've had varying success with pseudo-labelling in the past. Often it comes down to what is an acceptable label. I've turned to using a bit more active learning in practice. I think that the discussion posting that @mobassir posted is very enlightening as well. There is a lot of good ideas there I think. \n\nAlso, it is early and never hurts to try different things. Rely on your CV and not the public leaderboard though, it is much more likely that will help you on the private side." }, "source": "meta" }