Can you share the detail of data Augmentation boosted your score\nand How to use the depth pretrained network\nBecause I use the resnet50 of keras but the smallest size of input image of the network is over the (64,64,1)
", "messages": [ { "id": 421535, "postDate": "2018-11-15T05:27:44.947Z", "content": "Can you share the detail of data Augmentation boosted your score\nand How to use the depth pretrained network\nBecause I use the resnet50 of keras but the smallest size of input image of the network is over the (64,64,1)
", "rawMarkdown": "Can you share the detail of data Augmentation boosted your score\nand How to use the depth pretrained network\nBecause I use the resnet50 of keras but the smallest size of input image of the network is over the (64,64,1)", "votes": 3 }, { "id": 421983, "postDate": "2018-11-15T16:30:27.873Z", "content": "I tried flip and a small degree of rotation, but I didn't notice any real improvement.
", "rawMarkdown": "I tried flip and a small degree of rotation, but I didn't notice any real improvement.", "votes": 1 }, { "id": 429882, "postDate": "2018-11-29T13:51:34.553Z", "content": "I did a bit of padding the images and shift up down, made it worse T.T
", "rawMarkdown": "I did a bit of padding the images and shift up down, made it worse T.T", "replies": [ { "id": 429937, "postDate": "2018-11-29T15:11:29.480Z", "content": "@HuyenNguyen\ndid you try augmentation as @Appian mentioned 80% of strokes ?
for me, draw and rotate each stroke then combine them.but it need large compute power so，it take lot of time\ndata augmentation by stroke is good way.\nand another way try change stroke order, but i have no more time to try.
\n贫穷限制了我的发展，这句话你应该可能看懂的 :)
Augmentation is usually better for small training dataset on relatively large model.\nFor this competition, it is not needed. \nBut you may try flip augmentation which won't hurt.
", "rawMarkdown": "Augmentation is usually better for small training dataset on relatively large model.\nFor this competition, it is not needed. \nBut you may try flip augmentation which won't hurt.", "replies": [ { "id": 422011, "postDate": "2018-11-15T17:06:31.070Z", "content": "\"Augmentation is usually better for small training dataset ...\"
\n\nNot true.
\n\nFor this challenge, you should look at the train and test images and think of how to make the train more similar to the test?
\n\nEspecially, check those mis-classified validation images. why are them different from the training set?
\n\ne.g.
\n\nincomplete drawing : hence drop stroke can be a good augmentation
\n\ninclusion of text and arrow : hence add text or arrow can be a good augmentation
", "rawMarkdown": "\"Augmentation is usually better for small training dataset ...\" \n\nNot true.\n\nFor this challenge, you should look at the train and test images and think of how to make the train more similar to the test?\n\nEspecially, check those mis-classified validation images. why are them different from the training set?\n\ne.g. \n\n incomplete drawing : hence drop stroke can be a good augmentation\n\ninclusion of text and arrow : hence add text or arrow can be a good augmentation\n \n", "votes": 8 }, { "id": 422241, "postDate": "2018-11-16T00:24:40.060Z", "content": "Hi, Heng. How did you adjust the learning_rate?
", "rawMarkdown": "Hi, Heng. How did you adjust the learning_rate?" }, { "id": 427312, "postDate": "2018-11-25T06:58:40.633Z", "content": "Hi Heng, did you try to drop stokes randomly and did it work?
", "rawMarkdown": "Hi Heng, did you try to drop stokes randomly and did it work?" } ] }, { "id": 421806, "postDate": "2018-11-15T12:53:36.790Z", "content": "So far, I haven't used any data augmentation, I don't know if it works. you can modify the variable of min_size from imagenet_utils.py for keras, and then you can use 64*64.
", "rawMarkdown": "So far, I haven't used any data augmentation, I don't know if it works. you can modify the variable of min_size from imagenet_utils.py for keras, and then you can use 64*64. " } ], "comments": [ { "id": 421983, "author_name": "Allen", "author_url": "", "post_date": "2018-11-15T16:30:27.873000", "content": "I tried flip and a small degree of rotation, but I didn't notice any real improvement.
", "votes": 1, "replies": [] }, { "id": 429882, "author_name": "HuyenNguyen", "author_url": "", "post_date": "2018-11-29T13:51:34.553000", "content": "I did a bit of padding the images and shift up down, made it worse T.T
", "votes": 0, "replies": [ { "id": 429937, "author_name": "remidi", "author_url": "", "post_date": "2018-11-29T15:11:29.480000", "content": "@HuyenNguyen\ndid you try augmentation as @Appian mentioned 80% of strokes ?
for me, draw and rotate each stroke then combine them.but it need large compute power so，it take lot of time\ndata augmentation by stroke is good way.\nand another way try change stroke order, but i have no more time to try.
\n贫穷限制了我的发展，这句话你应该可能看懂的 :)
Augmentation is usually better for small training dataset on relatively large model.\nFor this competition, it is not needed. \nBut you may try flip augmentation which won't hurt.
", "votes": 0, "replies": [ { "id": 422011, "author_name": "hengck23", "author_url": "", "post_date": "2018-11-15T17:06:31.070000", "content": "\"Augmentation is usually better for small training dataset ...\"
\n\nNot true.
\n\nFor this challenge, you should look at the train and test images and think of how to make the train more similar to the test?
\n\nEspecially, check those mis-classified validation images. why are them different from the training set?
\n\ne.g.
\n\nincomplete drawing : hence drop stroke can be a good augmentation
\n\ninclusion of text and arrow : hence add text or arrow can be a good augmentation
", "votes": 8, "replies": [] }, { "id": 422241, "author_name": "Gary", "author_url": "", "post_date": "2018-11-16T00:24:40.060000", "content": "Hi, Heng. How did you adjust the learning_rate?
", "votes": 0, "replies": [] }, { "id": 427312, "author_name": "[he.ai]soulmachine", "author_url": "", "post_date": "2018-11-25T06:58:40.633000", "content": "Hi Heng, did you try to drop stokes randomly and did it work?
", "votes": 0, "replies": [] } ] }, { "id": 421806, "author_name": "Gary", "author_url": "", "post_date": "2018-11-15T12:53:36.790000", "content": "So far, I haven't used any data augmentation, I don't know if it works. you can modify the variable of min_size from imagenet_utils.py for keras, and then you can use 64*64.
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "421535": "Can you share the detail of data Augmentation boosted your score\nand How to use the depth pretrained network\nBecause I use the resnet50 of keras but the smallest size of input image of the network is over the (64,64,1)", "421983": "I tried flip and a small degree of rotation, but I didn't notice any real improvement.", "429882": "I did a bit of padding the images and shift up down, made it worse T.T", "428283": "for me, draw and rotate each stroke then combine them.but it need large compute power so，it take lot of time\ndata augmentation by stroke is good way.\nand another way try change stroke order, but i have no more time to try. \n贫穷限制了我的发展，这句话你应该可能看懂的 :)", "421929": "Augmentation is usually better for small training dataset on relatively large model.\nFor this competition, it is not needed. \nBut you may try flip augmentation which won't hurt.", "421806": "So far, I haven't used any data augmentation, I don't know if it works. you can modify the variable of min_size from imagenet_utils.py for keras, and then you can use 64*64. " } }