I've been wondering about the augmentation part for a while as in the end these are co-ordinates has augmentation proven to be useful in the final submission.
", "messages": [ { "id": 405292, "postDate": "2018-10-17T07:38:41.297Z", "content": "but dont you think the data we have already contains lots of variations for each class?
", "rawMarkdown": "but dont you think the data we have already contains lots of variations for each class?", "votes": 1, "replies": [ { "id": 405543, "postDate": "2018-10-17T17:17:03.850Z", "content": "For neural networks and ML in general there is on such thing as too much data and variation, you just have to make sure that all the classes have about the same part in the variety
", "rawMarkdown": "For neural networks and ML in general there is on such thing as too much data and variation, you just have to make sure that all the classes have about the same part in the variety " } ] }, { "id": 405187, "postDate": "2018-10-17T03:11:15.547Z", "content": "I have been thinking about this, and although rotating and flipping images might work we have to think about the data as doodles. One person's doodle might be, or dare i say, WILL BE distorted, bend, have less lines, or even more (although this data augmentation might be too hard or even impossible to predict).
\n\nMain point is if we want to do data augmentation we should apply doodle properties, which on the top of my head are, doodle bent, flip, distord, zoom?, rotate. I will try to apply this on my new kernel after i get the current results.\nThere are some comments that talk about it here
", "rawMarkdown": "I have been thinking about this, and although rotating and flipping images might work we have to think about the data as doodles. One person's doodle might be, or dare i say, WILL BE distorted, bend, have less lines, or even more (although this data augmentation might be too hard or even impossible to predict). \n\n\n\nMain point is if we want to do data augmentation we should apply doodle properties, which on the top of my head are, doodle bent, flip, distord, zoom?, rotate. I will try to apply this on my new kernel after i get the current results.\nThere are some comments that talk about it [here][1]\n\n [1]: https://www.kaggle.com/c/quickdraw-doodle-recognition/discussion/68006", "votes": 1 }, { "id": 403065, "postDate": "2018-10-12T20:09:24.293Z", "content": "I've been wondering about the augmentation part for a while as in the end these are co-ordinates has augmentation proven to be useful in the final submission.
", "rawMarkdown": "I've been wondering about the augmentation part for a while as in the end these are co-ordinates has augmentation proven to be useful in the final submission." } ], "comments": [ { "id": 405292, "author_name": "Heisenberg", "author_url": "", "post_date": "2018-10-17T07:38:41.297000", "content": "but dont you think the data we have already contains lots of variations for each class?
", "votes": 1, "replies": [ { "id": 405543, "author_name": "André Neves", "author_url": "", "post_date": "2018-10-17T17:17:03.850000", "content": "For neural networks and ML in general there is on such thing as too much data and variation, you just have to make sure that all the classes have about the same part in the variety
", "votes": 0, "replies": [] } ] }, { "id": 405187, "author_name": "André Neves", "author_url": "", "post_date": "2018-10-17T03:11:15.547000", "content": "I have been thinking about this, and although rotating and flipping images might work we have to think about the data as doodles. One person's doodle might be, or dare i say, WILL BE distorted, bend, have less lines, or even more (although this data augmentation might be too hard or even impossible to predict).
\n\nMain point is if we want to do data augmentation we should apply doodle properties, which on the top of my head are, doodle bent, flip, distord, zoom?, rotate. I will try to apply this on my new kernel after i get the current results.\nThere are some comments that talk about it here
", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "405292": "but dont you think the data we have already contains lots of variations for each class?", "405187": "I have been thinking about this, and although rotating and flipping images might work we have to think about the data as doodles. One person's doodle might be, or dare i say, WILL BE distorted, bend, have less lines, or even more (although this data augmentation might be too hard or even impossible to predict). \n\n\n\nMain point is if we want to do data augmentation we should apply doodle properties, which on the top of my head are, doodle bent, flip, distord, zoom?, rotate. I will try to apply this on my new kernel after i get the current results.\nThere are some comments that talk about it [here][1]\n\n [1]: https://www.kaggle.com/c/quickdraw-doodle-recognition/discussion/68006", "403065": "I've been wondering about the augmentation part for a while as in the end these are co-ordinates has augmentation proven to be useful in the final submission." } }