{ "id": 194278, "title": "Agmentation", "url": "/competitions/lyft-motion-prediction-autonomous-vehicles/discussion/194278", "author_name": "", "post_date": "2020-10-31T19:33:59.182187600Z", "votes": 1, "comment_count": 7, "views": 0, "content": "

Hi there all,

\n

Has anyone figure out how to do augmentations or and use data_transforms ?
\nThanks!

", "messages": [ { "id": "1065808", "postDate": "10/31/2020 19:33:59", "content": "

Hi there all,

\n

Has anyone figure out how to do augmentations or and use data_transforms ?
\nThanks!

", "rawMarkdown": "Hi there all, \n\nHas anyone figure out how to do augmentations or and use data_transforms ?\nThanks!", "votes": null }, { "id": "1065861", "postDate": "10/31/2020 23:12:39", "content": "

Some people have different ideas:

\n

e.g.

\n

@ryches came up with following: https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/188368

\n

@fergusoci discussed how relevant augmentation is in this competition: https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/191042

", "rawMarkdown": "Some people have different ideas:\n\ne.g. \n\n@ryches came up with following: https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/188368\n\n@fergusoci discussed how relevant augmentation is in this competition: https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/191042", "votes": null }, { "id": "1066786", "postDate": "11/02/2020 03:16:43", "content": "

one you change the problem into the prediction of image coord (instead of world coord), you can use the same old image augmentation like flip, rotate image.

\n

you can then also apply TTA

", "rawMarkdown": "one you change the problem into the prediction of image coord (instead of world coord), you can use the same old image augmentation like flip, rotate image.\n\nyou can then also apply TTA", "votes": null }, { "id": "1066950", "postDate": "11/02/2020 08:51:02", "content": "

But we also have to adjust other parts of the dataset as mention by Dmytro Poplavskiy in here, which could be a lot more jobs to do… Doing that, can we be sure that there will not be any inconsistency? Such a \"mysterious\" way to load data really drives me crazy…

", "rawMarkdown": "But we also have to adjust other parts of the dataset as mention by Dmytro Poplavskiy in [here](https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/188368), which could be a lot more jobs to do... Doing that, can we be sure that there will not be any inconsistency? Such a \"mysterious\" way to load data really drives me crazy...", "votes": null }, { "id": "1066956", "postDate": "11/02/2020 09:00:33", "content": "

Its questionable if augmentation with such a huge train dataset is worth it, it gave @fergusoci as stated in her discussion only a improvement of 0.2 - 0.5:

\n

https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/191042

", "rawMarkdown": "Its questionable if augmentation with such a huge train dataset is worth it, it gave @fergusoci as stated in her discussion only a improvement of 0.2 - 0.5: \n\nhttps://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/191042", "votes": null }, { "id": "1066991", "postDate": "11/02/2020 09:22:20", "content": "

it is a question of \"efficiency\", i.e performance gain over the cost incurred.

\n

the advantage of augmentation may come in TTA and semi/self-supervised learning

", "rawMarkdown": "it is a question of \"efficiency\", i.e performance gain over the cost incurred.\n\nthe advantage of augmentation may come in TTA and semi/self-supervised learning", "votes": null }, { "id": "1067692", "postDate": "11/02/2020 16:24:50", "content": "

Thanks you guys, for the responses! but I that the Lyft has some mayor bugs and still very static on it's implementation, meaning not flexible. specially with the SDK with lack of documentation. Will read the different links posted here.

\n

thanks again.

", "rawMarkdown": "Thanks you guys, for the responses! but I that the Lyft has some mayor bugs and still very static on it's implementation, meaning not flexible. specially with the SDK with lack of documentation. Will read the different links posted here.\n\nthanks again.", "votes": null }, { "id": "1069313", "postDate": "11/04/2020 10:02:58", "content": "

Hey :) Have you checked the official documentation? Although it doesn't cover 100% of l5kit it's still a good place to start from if you're looking for info about the data format and other stuff. If you think something is missing feel free to open an issue in l5kit directly, we'll try to address it :)!

", "rawMarkdown": "Hey :) Have you checked the [official documentation](https://lyft.github.io/l5kit/)? Although it doesn't cover 100% of l5kit it's still a good place to start from if you're looking for info about the data format and other stuff. If you think something is missing feel free to open an issue in l5kit directly, we'll try to address it :)!", "votes": null } ], "comments": [ { "id": 1065861, "author_name": "aliabdin1", "author_url": "", "post_date": "10/31/2020 23:12:39", "content": "

Some people have different ideas:

\n

e.g.

\n

@ryches came up with following: https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/188368

\n

@fergusoci discussed how relevant augmentation is in this competition: https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/191042

", "votes": null, "replies": [] }, { "id": 1066786, "author_name": "hengck23", "author_url": "", "post_date": "11/02/2020 03:16:43", "content": "

one you change the problem into the prediction of image coord (instead of world coord), you can use the same old image augmentation like flip, rotate image.

\n

you can then also apply TTA

", "votes": null, "replies": [ { "id": 1066950, "author_name": "spicychicken38", "author_url": "", "post_date": "11/02/2020 08:51:02", "content": "

But we also have to adjust other parts of the dataset as mention by Dmytro Poplavskiy in here, which could be a lot more jobs to do… Doing that, can we be sure that there will not be any inconsistency? Such a \"mysterious\" way to load data really drives me crazy…

", "votes": null, "replies": [] }, { "id": 1066956, "author_name": "aliabdin1", "author_url": "", "post_date": "11/02/2020 09:00:33", "content": "

Its questionable if augmentation with such a huge train dataset is worth it, it gave @fergusoci as stated in her discussion only a improvement of 0.2 - 0.5:

\n

https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/191042

", "votes": null, "replies": [] }, { "id": 1066991, "author_name": "hengck23", "author_url": "", "post_date": "11/02/2020 09:22:20", "content": "

it is a question of \"efficiency\", i.e performance gain over the cost incurred.

\n

the advantage of augmentation may come in TTA and semi/self-supervised learning

", "votes": null, "replies": [] } ] }, { "id": 1067692, "author_name": "oscarrangel", "author_url": "", "post_date": "11/02/2020 16:24:50", "content": "

Thanks you guys, for the responses! but I that the Lyft has some mayor bugs and still very static on it's implementation, meaning not flexible. specially with the SDK with lack of documentation. Will read the different links posted here.

\n

thanks again.

", "votes": null, "replies": [ { "id": 1069313, "author_name": "lucabergamini", "author_url": "", "post_date": "11/04/2020 10:02:58", "content": "

Hey :) Have you checked the official documentation? Although it doesn't cover 100% of l5kit it's still a good place to start from if you're looking for info about the data format and other stuff. If you think something is missing feel free to open an issue in l5kit directly, we'll try to address it :)!

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1065808": "Hi there all, \n\nHas anyone figure out how to do augmentations or and use data_transforms ?\nThanks!", "1065861": "Some people have different ideas:\n\ne.g. \n\n@ryches came up with following: https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/188368\n\n@fergusoci discussed how relevant augmentation is in this competition: https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/191042", "1066786": "one you change the problem into the prediction of image coord (instead of world coord), you can use the same old image augmentation like flip, rotate image.\n\nyou can then also apply TTA", "1066950": "But we also have to adjust other parts of the dataset as mention by Dmytro Poplavskiy in [here](https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/188368), which could be a lot more jobs to do... Doing that, can we be sure that there will not be any inconsistency? Such a \"mysterious\" way to load data really drives me crazy...", "1066956": "Its questionable if augmentation with such a huge train dataset is worth it, it gave @fergusoci as stated in her discussion only a improvement of 0.2 - 0.5: \n\nhttps://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/191042", "1066991": "it is a question of \"efficiency\", i.e performance gain over the cost incurred.\n\nthe advantage of augmentation may come in TTA and semi/self-supervised learning", "1067692": "Thanks you guys, for the responses! but I that the Lyft has some mayor bugs and still very static on it's implementation, meaning not flexible. specially with the SDK with lack of documentation. Will read the different links posted here.\n\nthanks again.", "1069313": "Hey :) Have you checked the [official documentation](https://lyft.github.io/l5kit/)? Although it doesn't cover 100% of l5kit it's still a good place to start from if you're looking for info about the data format and other stuff. If you think something is missing feel free to open an issue in l5kit directly, we'll try to address it :)!" }, "source": "meta" }