I have tried some types of data augmentations like time/frequency mask, add noise, random flip and so on. But it seems strange that almost of them doesn't work. However I find data augmentations ususally take an important role in old competitions.
\nSo do you get score improved by taking use of data augmentations?
", "messages": [ { "id": "2656811", "postDate": "02/18/2024 03:31:57", "content": "I have tried some types of data augmentations like time/frequency mask, add noise, random flip and so on. But it seems strange that almost of them doesn't work. However I find data augmentations ususally take an important role in old competitions.
\nSo do you get score improved by taking use of data augmentations?
", "rawMarkdown": "I have tried some types of data augmentations like time/frequency mask, add noise, random flip and so on. But it seems strange that almost of them doesn't work. However I find data augmentations ususally take an important role in old competitions.\n\nSo do you get score improved by taking use of data augmentations?", "votes": null }, { "id": "2656895", "postDate": "02/18/2024 05:55:40", "content": "In my experiments, very limited (maybe ±0.01), even worse
", "rawMarkdown": "In my experiments, very limited (maybe ±0.01), even worse", "votes": null }, { "id": "2656911", "postDate": "02/18/2024 06:17:50", "content": "Yes, they work really good. I got 0.05-0.1 improvement over the baseline without any augmentations.
", "rawMarkdown": "Yes, they work really good. I got 0.05-0.1 improvement over the baseline without any augmentations.", "votes": null }, { "id": "2656940", "postDate": "02/18/2024 06:53:10", "content": "I got wild results for this myself. I had one notebook that got slightly worse (both in CV and LB) and then another notebook that improved significantly (0.06). I have no real assumption for why this may be the case but even augmentations such as a horizontal flip had both effects on the notebooks respectively.
", "rawMarkdown": "I got wild results for this myself. I had one notebook that got slightly worse (both in CV and LB) and then another notebook that improved significantly (0.06). I have no real assumption for why this may be the case but even augmentations such as a horizontal flip had both effects on the notebooks respectively.", "votes": null }, { "id": "2656942", "postDate": "02/18/2024 06:54:09", "content": "I'm sorry but I'm a little confused. You mean you get 0.05~0.1 improvement not using augmentations but I ask if there is improvement using augmentations ?
", "rawMarkdown": "I'm sorry but I'm a little confused. You mean you get 0.05~0.1 improvement not using augmentations but I ask if there is improvement using augmentations ?", "votes": null }, { "id": "2656943", "postDate": "02/18/2024 06:55:40", "content": "That's similar to mine.
", "rawMarkdown": "That's similar to mine.", "votes": null }, { "id": "2656949", "postDate": "02/18/2024 07:02:24", "content": "🤣That's so strange. May I ask whether your best single model use augmentations?
", "rawMarkdown": "🤣That's so strange. May I ask whether your best single model use augmentations?", "votes": null }, { "id": "2656956", "postDate": "02/18/2024 07:05:11", "content": "Sorry I wrote it in a confusing way. My baseline was without any augmentations. When I added augmentations, I got improvement between 0.05 and 0.1.
", "rawMarkdown": "Sorry I wrote it in a confusing way. My baseline was without any augmentations. When I added augmentations, I got improvement between 0.05 and 0.1.", "votes": null }, { "id": "2656959", "postDate": "02/18/2024 07:09:07", "content": "Currently, my best model is using augmentations
", "rawMarkdown": "Currently, my best model **is** using augmentations", "votes": null }, { "id": "2656961", "postDate": "02/18/2024 07:10:53", "content": "I understand, thanks for your reply!
", "rawMarkdown": "I understand, thanks for your reply!", "votes": null }, { "id": "2656963", "postDate": "02/18/2024 07:11:37", "content": "👌, thanks!
", "rawMarkdown": "👌, thanks!", "votes": null }, { "id": "2656982", "postDate": "02/18/2024 07:24:43", "content": "I got worse CV when using Gauss noise.
", "rawMarkdown": "I got worse CV when using Gauss noise.", "votes": null }, { "id": "2656986", "postDate": "02/18/2024 07:29:48", "content": "wow, that's incredible.
", "rawMarkdown": "wow, that's incredible.", "votes": null }, { "id": "2657075", "postDate": "02/18/2024 09:27:08", "content": "I've got around 0.06 CV improvement from augmentations in 1D eegs, however only mixup worked for 2D specs.
", "rawMarkdown": "I've got around 0.06 CV improvement from augmentations in 1D eegs, however only mixup worked for 2D specs.", "votes": null }, { "id": "2657156", "postDate": "02/18/2024 10:53:54", "content": "Mixup didn't work for me on 2D efficientnet. 🤣
", "rawMarkdown": "Mixup didn't work for me on 2D efficientnet. 🤣", "votes": null }, { "id": "2657292", "postDate": "02/18/2024 12:33:32", "content": "just to confirm @gunesevitan 0.1 improvement in KL div? that means 0.4x to 0.3x (approx)?
", "rawMarkdown": "just to confirm @gunesevitan 0.1 improvement in KL div? that means 0.4x to 0.3x (approx)?", "votes": null }, { "id": "2657296", "postDate": "02/18/2024 12:40:43", "content": "Yes augs help a lot.
", "rawMarkdown": "Yes augs help a lot.", "votes": null } ], "comments": [ { "id": 2656895, "author_name": "zijiangyang1116", "author_url": "", "post_date": "02/18/2024 05:55:40", "content": "In my experiments, very limited (maybe ±0.01), even worse
", "votes": null, "replies": [ { "id": 2656943, "author_name": "zui0711", "author_url": "", "post_date": "02/18/2024 06:55:40", "content": "That's similar to mine.
", "votes": null, "replies": [] } ] }, { "id": 2656911, "author_name": "gunesevitan", "author_url": "", "post_date": "02/18/2024 06:17:50", "content": "Yes, they work really good. I got 0.05-0.1 improvement over the baseline without any augmentations.
", "votes": null, "replies": [ { "id": 2656942, "author_name": "zui0711", "author_url": "", "post_date": "02/18/2024 06:54:09", "content": "I'm sorry but I'm a little confused. You mean you get 0.05~0.1 improvement not using augmentations but I ask if there is improvement using augmentations ?
", "votes": null, "replies": [ { "id": 2656956, "author_name": "gunesevitan", "author_url": "", "post_date": "02/18/2024 07:05:11", "content": "Sorry I wrote it in a confusing way. My baseline was without any augmentations. When I added augmentations, I got improvement between 0.05 and 0.1.
", "votes": null, "replies": [ { "id": 2656961, "author_name": "zui0711", "author_url": "", "post_date": "02/18/2024 07:10:53", "content": "I understand, thanks for your reply!
", "votes": null, "replies": [] } ] } ] }, { "id": 2656986, "author_name": "sweetyheehee", "author_url": "", "post_date": "02/18/2024 07:29:48", "content": "wow, that's incredible.
", "votes": null, "replies": [ { "id": 2657292, "author_name": "chaudharypriyanshu", "author_url": "", "post_date": "02/18/2024 12:33:32", "content": "just to confirm @gunesevitan 0.1 improvement in KL div? that means 0.4x to 0.3x (approx)?
", "votes": null, "replies": [ { "id": 2657296, "author_name": "gunesevitan", "author_url": "", "post_date": "02/18/2024 12:40:43", "content": "Yes augs help a lot.
", "votes": null, "replies": [] } ] } ] } ] }, { "id": 2656940, "author_name": "cody11null", "author_url": "", "post_date": "02/18/2024 06:53:10", "content": "I got wild results for this myself. I had one notebook that got slightly worse (both in CV and LB) and then another notebook that improved significantly (0.06). I have no real assumption for why this may be the case but even augmentations such as a horizontal flip had both effects on the notebooks respectively.
", "votes": null, "replies": [ { "id": 2656949, "author_name": "zui0711", "author_url": "", "post_date": "02/18/2024 07:02:24", "content": "🤣That's so strange. May I ask whether your best single model use augmentations?
", "votes": null, "replies": [ { "id": 2656959, "author_name": "cody11null", "author_url": "", "post_date": "02/18/2024 07:09:07", "content": "Currently, my best model is using augmentations
", "votes": null, "replies": [ { "id": 2656963, "author_name": "zui0711", "author_url": "", "post_date": "02/18/2024 07:11:37", "content": "👌, thanks!
", "votes": null, "replies": [] } ] } ] } ] }, { "id": 2656982, "author_name": "sweetyheehee", "author_url": "", "post_date": "02/18/2024 07:24:43", "content": "I got worse CV when using Gauss noise.
", "votes": null, "replies": [] }, { "id": 2657075, "author_name": "fahmiayari", "author_url": "", "post_date": "02/18/2024 09:27:08", "content": "I've got around 0.06 CV improvement from augmentations in 1D eegs, however only mixup worked for 2D specs.
", "votes": null, "replies": [ { "id": 2657156, "author_name": "zui0711", "author_url": "", "post_date": "02/18/2024 10:53:54", "content": "Mixup didn't work for me on 2D efficientnet. 🤣
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "2656811": "I have tried some types of data augmentations like time/frequency mask, add noise, random flip and so on. But it seems strange that almost of them doesn't work. However I find data augmentations ususally take an important role in old competitions.\n\nSo do you get score improved by taking use of data augmentations?", "2656895": "In my experiments, very limited (maybe ±0.01), even worse", "2656911": "Yes, they work really good. I got 0.05-0.1 improvement over the baseline without any augmentations.", "2656940": "I got wild results for this myself. I had one notebook that got slightly worse (both in CV and LB) and then another notebook that improved significantly (0.06). I have no real assumption for why this may be the case but even augmentations such as a horizontal flip had both effects on the notebooks respectively.", "2656942": "I'm sorry but I'm a little confused. You mean you get 0.05~0.1 improvement not using augmentations but I ask if there is improvement using augmentations ?", "2656943": "That's similar to mine.", "2656949": "🤣That's so strange. May I ask whether your best single model use augmentations?", "2656956": "Sorry I wrote it in a confusing way. My baseline was without any augmentations. When I added augmentations, I got improvement between 0.05 and 0.1.", "2656959": "Currently, my best model **is** using augmentations", "2656961": "I understand, thanks for your reply!", "2656963": "👌, thanks!", "2656982": "I got worse CV when using Gauss noise.", "2656986": "wow, that's incredible.", "2657075": "I've got around 0.06 CV improvement from augmentations in 1D eegs, however only mixup worked for 2D specs.", "2657156": "Mixup didn't work for me on 2D efficientnet. 🤣", "2657292": "just to confirm @gunesevitan 0.1 improvement in KL div? that means 0.4x to 0.3x (approx)?", "2657296": "Yes augs help a lot." }, "source": "meta" }