Hi all,\nThank you all for sharing your solutions!
\n\nI feel one of the key point of this competition was finding best augmentations and I couldn't manage to find it. I tried random combinations or copied from similar past competitions, but none of them worked for me.
\n\nCould you someone share how you found your best augmentations?
\n\nthanks!
", "messages": [ { "id": "621262", "postDate": "09/08/2019 11:36:19", "content": "Hi all,\nThank you all for sharing your solutions!
\n\nI feel one of the key point of this competition was finding best augmentations and I couldn't manage to find it. I tried random combinations or copied from similar past competitions, but none of them worked for me.
\n\nCould you someone share how you found your best augmentations?
\n\nthanks!
", "rawMarkdown": "Hi all,\nThank you all for sharing your solutions!\n\nI feel one of the key point of this competition was finding best augmentations and I couldn't manage to find it. I tried random combinations or copied from similar past competitions, but none of them worked for me.\n\nCould you someone share how you found your best augmentations?\n\nthanks!", "votes": null }, { "id": "621297", "postDate": "09/08/2019 12:17:32", "content": "This was my first computer vision competition where I really sat down and studied everything I could get my hands on. My first step in any competition is to read solutions from past competitions and papers to get an idea of \"tried and true\" approaches. In this case, I looked at what were popular augmentation strategies (past DR comp, CV comps this year, and papers on DR published in arvix) and made a list of things to try. I've always relied on a local CV for all competitions since Mercedes - I try my best not to get caught up with the public LB score, and it's helped me with some of my recent competitions such as Quora, etc. In this case, I iterated through the list of augmentations - tried one, then a combination of a couple, just to get a \"feel\" of what might work \"better\". This slow and manual approach was motivated by the auto augmentation paper.
", "rawMarkdown": "This was my first computer vision competition where I really sat down and studied everything I could get my hands on. My first step in any competition is to read solutions from past competitions and papers to get an idea of \"tried and true\" approaches. In this case, I looked at what were popular augmentation strategies (past DR comp, CV comps this year, and papers on DR published in arvix) and made a list of things to try. I've always relied on a local CV for all competitions since Mercedes - I try my best not to get caught up with the public LB score, and it's helped me with some of my recent competitions such as Quora, etc. In this case, I iterated through the list of augmentations - tried one, then a combination of a couple, just to get a \"feel\" of what might work \"better\". This slow and manual approach was motivated by the auto augmentation paper.", "votes": null }, { "id": "623018", "postDate": "09/10/2019 11:35:05", "content": "Thank you Thomas for sharing your method.
\n\n\n\n\nI iterated through the list of augmentations - tried one, then a combination of a couple, just to get a \"feel\" of what might work \"better
\n
The \"feel\" part was hard for me. I was not able to get a sense of \"what might work\".
", "rawMarkdown": "Thank you Thomas for sharing your method.\n\n> I iterated through the list of augmentations - tried one, then a combination of a couple, just to get a \"feel\" of what might work \"better\n\nThe \"feel\" part was hard for me. I was not able to get a sense of \"what might work\".", "votes": null }, { "id": "623039", "postDate": "09/10/2019 11:56:40", "content": "Hi @higepon , did you see my kernel on this topic : \nhttps://www.kaggle.com/ratthachat/aptos-updated-albumentation-meets-grad-cam
\n\nFirst, I also used intuition, but the most important step is to confirm our intuition with visualization / performance improvement.
\n\nBTW, I already PM you but perhaps you didn't see it yet :)
", "rawMarkdown": "Hi @higepon , did you see my kernel on this topic : \nhttps://www.kaggle.com/ratthachat/aptos-updated-albumentation-meets-grad-cam\n\nFirst, I also used intuition, but the most important step is to confirm our intuition with visualization / performance improvement.\n\nBTW, I already PM you but perhaps you didn't see it yet :)", "votes": null }, { "id": "623131", "postDate": "09/10/2019 13:59:52", "content": "Ah. I quickly checked your kennel during the competition and never revisited it. I really should have done it.
\n\n\n\n\nBTW, I already PM you but perhaps you didn't see it yet :)
\n
Thank you! Unfortunately I wasn't able to find it. Which channel did you use to contact me :)?
", "rawMarkdown": "Ah. I quickly checked your kennel during the competition and never revisited it. I really should have done it.\n\n> BTW, I already PM you but perhaps you didn't see it yet :)\n\nThank you! Unfortunately I wasn't able to find it. Which channel did you use to contact me :)?", "votes": null }, { "id": "623221", "postDate": "09/10/2019 16:00:23", "content": "There is “contact user” in the profile page. So I use this channel. This should correspond to the registered email in Kaggle :)
", "rawMarkdown": "There is “contact user” in the profile page. So I use this channel. This should correspond to the registered email in Kaggle :)", "votes": null }, { "id": "623387", "postDate": "09/10/2019 21:05:16", "content": "Found it. Thanks! Replying.
", "rawMarkdown": "Found it. Thanks! Replying.", "votes": null } ], "comments": [ { "id": 621297, "author_name": "learnmower", "author_url": "", "post_date": "09/08/2019 12:17:32", "content": "This was my first computer vision competition where I really sat down and studied everything I could get my hands on. My first step in any competition is to read solutions from past competitions and papers to get an idea of \"tried and true\" approaches. In this case, I looked at what were popular augmentation strategies (past DR comp, CV comps this year, and papers on DR published in arvix) and made a list of things to try. I've always relied on a local CV for all competitions since Mercedes - I try my best not to get caught up with the public LB score, and it's helped me with some of my recent competitions such as Quora, etc. In this case, I iterated through the list of augmentations - tried one, then a combination of a couple, just to get a \"feel\" of what might work \"better\". This slow and manual approach was motivated by the auto augmentation paper.
", "votes": null, "replies": [ { "id": 623018, "author_name": "higepon", "author_url": "", "post_date": "09/10/2019 11:35:05", "content": "Thank you Thomas for sharing your method.
\n\n\n\n\nI iterated through the list of augmentations - tried one, then a combination of a couple, just to get a \"feel\" of what might work \"better
\n
The \"feel\" part was hard for me. I was not able to get a sense of \"what might work\".
", "votes": null, "replies": [] }, { "id": 623039, "author_name": "ratthachat", "author_url": "", "post_date": "09/10/2019 11:56:40", "content": "Hi @higepon , did you see my kernel on this topic : \nhttps://www.kaggle.com/ratthachat/aptos-updated-albumentation-meets-grad-cam
\n\nFirst, I also used intuition, but the most important step is to confirm our intuition with visualization / performance improvement.
\n\nBTW, I already PM you but perhaps you didn't see it yet :)
", "votes": null, "replies": [] }, { "id": 623131, "author_name": "higepon", "author_url": "", "post_date": "09/10/2019 13:59:52", "content": "Ah. I quickly checked your kennel during the competition and never revisited it. I really should have done it.
\n\n\n\n\nBTW, I already PM you but perhaps you didn't see it yet :)
\n
Thank you! Unfortunately I wasn't able to find it. Which channel did you use to contact me :)?
", "votes": null, "replies": [] }, { "id": 623221, "author_name": "ratthachat", "author_url": "", "post_date": "09/10/2019 16:00:23", "content": "There is “contact user” in the profile page. So I use this channel. This should correspond to the registered email in Kaggle :)
", "votes": null, "replies": [] }, { "id": 623387, "author_name": "higepon", "author_url": "", "post_date": "09/10/2019 21:05:16", "content": "Found it. Thanks! Replying.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "621262": "Hi all,\nThank you all for sharing your solutions!\n\nI feel one of the key point of this competition was finding best augmentations and I couldn't manage to find it. I tried random combinations or copied from similar past competitions, but none of them worked for me.\n\nCould you someone share how you found your best augmentations?\n\nthanks!", "621297": "This was my first computer vision competition where I really sat down and studied everything I could get my hands on. My first step in any competition is to read solutions from past competitions and papers to get an idea of \"tried and true\" approaches. In this case, I looked at what were popular augmentation strategies (past DR comp, CV comps this year, and papers on DR published in arvix) and made a list of things to try. I've always relied on a local CV for all competitions since Mercedes - I try my best not to get caught up with the public LB score, and it's helped me with some of my recent competitions such as Quora, etc. In this case, I iterated through the list of augmentations - tried one, then a combination of a couple, just to get a \"feel\" of what might work \"better\". This slow and manual approach was motivated by the auto augmentation paper.", "623018": "Thank you Thomas for sharing your method.\n\n> I iterated through the list of augmentations - tried one, then a combination of a couple, just to get a \"feel\" of what might work \"better\n\nThe \"feel\" part was hard for me. I was not able to get a sense of \"what might work\".", "623039": "Hi @higepon , did you see my kernel on this topic : \nhttps://www.kaggle.com/ratthachat/aptos-updated-albumentation-meets-grad-cam\n\nFirst, I also used intuition, but the most important step is to confirm our intuition with visualization / performance improvement.\n\nBTW, I already PM you but perhaps you didn't see it yet :)", "623131": "Ah. I quickly checked your kennel during the competition and never revisited it. I really should have done it.\n\n> BTW, I already PM you but perhaps you didn't see it yet :)\n\nThank you! Unfortunately I wasn't able to find it. Which channel did you use to contact me :)?", "623221": "There is “contact user” in the profile page. So I use this channel. This should correspond to the registered email in Kaggle :)", "623387": "Found it. Thanks! Replying." }, "source": "meta" }