TensorFlow.data.Dataset(). Enjoy! \n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2F999713370831a44bb5af627f96b42f76%2FScreen%20Shot%202020-02-28%20at%2012.09.49%20PM.png?generation=1582920604632237&alt=media\")
I have posted a starter notebook here showing how to perform CutMix and MixUp on GPU/TPU using TensorFlow.data.Dataset(). Enjoy! \n
Data augmentation has repeatedly shown to improve model accuracy. Below is a list of some current literature on data augmentation techniques:
\n\nBasic augmentation includes flipping, rotation, sheer, zoom, and shift. Starter notebook here. Next we can incorporate non-linear transformations like grid distortion and we can do color augmentation like brightness, contrast, hue, etc. See Albumentations for ideas.
\n\nPDF here. Cutout randomly removes rectangular regions of the training images. It is similar to dropout. It teaches a network to use the entire image and not depend too much on certain regions or details in the image.\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2Fa75f1ff01f51be04efc3994511c14985%2FScreen%20Shot%202020-02-28%20at%2012.03.14%20PM.png?generation=1582920211872489&alt=media\")
PDF here. Mixup blends two images and their one hot encoded labels to create a new image. This helps a network generalize to the space of images they lie \"between\" the existing training images.\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2F0fd1c6f6c0d1b2c220748e1e8517dd42%2FScreen%20Shot%202020-02-28%20at%2012.04.14%20PM.png?generation=1582920270999369&alt=media\")
PDF here. List a variety of augmentation techniques. Use reinforcement learning to explore different combinations of augmentations and find the best.\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2Fa013e63820a9f2e5751fc954bd502561%2FScreen%20Shot%202020-02-28%20at%2012.05.08%20PM.png?generation=1582920325192602&alt=media\")
PDF here. This is cutout but we fill the blank rectangular space with a portion of another image. CutMix has the benefits of cutout (described above) and it expands the the space of images by creating new images (i.e. like a dog with a cat's head).\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2Fa4d6d7b6132350ac39dbe5d3ce601b55%2FScreen%20Shot%202020-02-28%20at%2012.06.24%20PM.png?generation=1582920401462775&alt=media\")
PDF here. Randomly blend multiple images that have each been augmented themselves separately. And use a special consistency loss.\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2F2b79e1180f717408930554343b007881%2FScreen%20Shot%202020-02-28%20at%2012.07.34%20PM.png?generation=1582920469116799&alt=media\")
I have posted a starter notebook here showing how to perform CutMix and MixUp on GPU/TPU using TensorFlow.data.Dataset(). Enjoy! \n
Data augmentation has repeatedly shown to improve model accuracy. Below is a list of some current literature on data augmentation techniques:
\n\nBasic augmentation includes flipping, rotation, sheer, zoom, and shift. Starter notebook here. Next we can incorporate non-linear transformations like grid distortion and we can do color augmentation like brightness, contrast, hue, etc. See Albumentations for ideas.
\n\nPDF here. Cutout randomly removes rectangular regions of the training images. It is similar to dropout. It teaches a network to use the entire image and not depend too much on certain regions or details in the image.\n
PDF here. Mixup blends two images and their one hot encoded labels to create a new image. This helps a network generalize to the space of images they lie \"between\" the existing training images.\n
PDF here. List a variety of augmentation techniques. Use reinforcement learning to explore different combinations of augmentations and find the best.\n
PDF here. This is cutout but we fill the blank rectangular space with a portion of another image. CutMix has the benefits of cutout (described above) and it expands the the space of images by creating new images (i.e. like a dog with a cat's head).\n
PDF here. Randomly blend multiple images that have each been augmented themselves separately. And use a special consistency loss.\n
Great work!
", "rawMarkdown": "Great work!", "votes": 3 }, { "id": 765422, "postDate": "2020-03-06T16:01:37.773Z", "content": "What I always liked in Chris that he always dives deep into the thing, trying not only understand how it works, but also understand why it works.
", "rawMarkdown": "What I always liked in Chris that he always dives deep into the thing, trying not only understand **how** it works, but also understand **why** it works.", "votes": 1, "replies": [ { "id": 765473, "postDate": "2020-03-06T16:57:08.953Z", "content": "Thanks Roman
", "rawMarkdown": "Thanks Roman" } ] }, { "id": 760160, "postDate": "2020-02-29T23:13:03.683Z", "content": "great dude
", "rawMarkdown": "great dude", "votes": 1 }, { "id": 759405, "postDate": "2020-02-29T02:00:55.243Z", "content": "Now, this is amazing ! Thank you.
", "rawMarkdown": "Now, this is amazing ! Thank you.", "votes": 1 }, { "id": 759286, "postDate": "2020-02-28T20:59:31.947Z", "content": "Oddsome topic (even with my poor knowledge I can appreciate it).
", "rawMarkdown": "Oddsome topic (even with my poor knowledge I can appreciate it).", "votes": 1 }, { "id": 1265698, "postDate": "2021-04-07T06:17:55.943Z", "content": "is amazing
", "rawMarkdown": " is amazing" }, { "id": 1076156, "postDate": "2020-11-12T09:45:22.700Z", "content": "Thanks for such an awesome explanation. I have a query.. In medical data like chest Xray data, is it an good idea to add cutmix or augmix?
", "rawMarkdown": "Thanks for such an awesome explanation. I have a query.. In medical data like chest Xray data, is it an good idea to add cutmix or augmix? " }, { "id": 843805, "postDate": "2020-05-12T09:18:39.930Z", "content": "Thanks for sharing this! This was one of the highlights of this competition for me, I didn't know about those augmentation methods.
", "rawMarkdown": "Thanks for sharing this! This was one of the highlights of this competition for me, I didn't know about those augmentation methods. " }, { "id": 777781, "postDate": "2020-03-17T23:46:20.387Z", "content": "I tried using your cutmix implementation but instead train a network once and not in a cross validation way but I get memory error. What would be the problem in this situation ?
", "rawMarkdown": "I tried using your cutmix implementation but instead train a network once and not in a cross validation way but I get memory error. What would be the problem in this situation ?", "replies": [ { "id": 777809, "postDate": "2020-03-18T00:45:39.610Z", "content": "Are you using TPU or GPU? If GPU you can change the variable in code cell 3. Reduce the AUG_BATCH by 2. That will use less CPU RAM. If TPU, i'm surprised you have memory error.
AUG_BATCH = BATCH_SIZE//2\nAlso consider lower your BATCH_SIZE
No, its tpu. I am using the same batch size and aug batch which is 128.
", "rawMarkdown": "No, its tpu. I am using the same batch size and aug batch which is 128." }, { "id": 777822, "postDate": "2020-03-18T01:10:45.207Z", "content": "Ok interesting. Still try my suggestions above. Try reducing AUG_BATCH and/or BATCH_SIZE. AUG_BATCH is batch size for CutMix and BATCH_SIZE is batch size for CNN training. Augmentation occurs in TPU's CPU RAM. When i get time, i'll run the code and experiment more.
The problem was fixed after applying mixed precision training, i was able to train using 128 batch size using tpu. I was wondering though what is AUG_BATCH really for and what effect If i set it to be half the size of BATCH_SIZE ?
Now, this is amazing ! Thank you.
", "rawMarkdown": "Now, this is amazing ! Thank you.\n\n", "votes": 1 }, { "id": 1126774, "postDate": "2020-12-25T23:19:39.070Z", "content": "Thanks for the Great Explanation
", "rawMarkdown": "Thanks for the Great Explanation " } ], "comments": [ { "id": 772436, "author_name": "podsyp", "author_url": "", "post_date": "2020-03-15T13:42:11.250000", "content": "Great work!
", "votes": 3, "replies": [] }, { "id": 765422, "author_name": "Roman", "author_url": "", "post_date": "2020-03-06T16:01:37.773000", "content": "What I always liked in Chris that he always dives deep into the thing, trying not only understand how it works, but also understand why it works.
", "votes": 1, "replies": [ { "id": 765473, "author_name": "Chris Deotte", "author_url": "", "post_date": "2020-03-06T16:57:08.953000", "content": "Thanks Roman
", "votes": 0, "replies": [] } ] }, { "id": 760160, "author_name": "AmoP", "author_url": "", "post_date": "2020-02-29T23:13:03.683000", "content": "great dude
", "votes": 1, "replies": [] }, { "id": 759405, "author_name": "Martin Görner", "author_url": "", "post_date": "2020-02-29T02:00:55.243000", "content": "Now, this is amazing ! Thank you.
", "votes": 1, "replies": [] }, { "id": 759286, "author_name": "Marília Prata", "author_url": "", "post_date": "2020-02-28T20:59:31.947000", "content": "Oddsome topic (even with my poor knowledge I can appreciate it).
", "votes": 1, "replies": [] }, { "id": 1265698, "author_name": "dayoneistoday", "author_url": "", "post_date": "2021-04-07T06:17:55.943000", "content": "is amazing
", "votes": 0, "replies": [] }, { "id": 1076156, "author_name": "Ritacheta Das", "author_url": "", "post_date": "2020-11-12T09:45:22.700000", "content": "Thanks for such an awesome explanation. I have a query.. In medical data like chest Xray data, is it an good idea to add cutmix or augmix?
", "votes": 0, "replies": [] }, { "id": 843805, "author_name": "Nicolas Petitclerc", "author_url": "", "post_date": "2020-05-12T09:18:39.930000", "content": "Thanks for sharing this! This was one of the highlights of this competition for me, I didn't know about those augmentation methods.
", "votes": 0, "replies": [] }, { "id": 777781, "author_name": "Ibrahim Sherif", "author_url": "", "post_date": "2020-03-17T23:46:20.387000", "content": "I tried using your cutmix implementation but instead train a network once and not in a cross validation way but I get memory error. What would be the problem in this situation ?
", "votes": 0, "replies": [ { "id": 777809, "author_name": "Chris Deotte", "author_url": "", "post_date": "2020-03-18T00:45:39.610000", "content": "Are you using TPU or GPU? If GPU you can change the variable in code cell 3. Reduce the AUG_BATCH by 2. That will use less CPU RAM. If TPU, i'm surprised you have memory error.
AUG_BATCH = BATCH_SIZE//2\nAlso consider lower your BATCH_SIZE
No, its tpu. I am using the same batch size and aug batch which is 128.
", "votes": 0, "replies": [] }, { "id": 777822, "author_name": "Chris Deotte", "author_url": "", "post_date": "2020-03-18T01:10:45.207000", "content": "Ok interesting. Still try my suggestions above. Try reducing AUG_BATCH and/or BATCH_SIZE. AUG_BATCH is batch size for CutMix and BATCH_SIZE is batch size for CNN training. Augmentation occurs in TPU's CPU RAM. When i get time, i'll run the code and experiment more.
The problem was fixed after applying mixed precision training, i was able to train using 128 batch size using tpu. I was wondering though what is AUG_BATCH really for and what effect If i set it to be half the size of BATCH_SIZE ?
Now, this is amazing ! Thank you.
", "votes": 1, "replies": [] }, { "id": 1126774, "author_name": "Omar Zaghlol", "author_url": "", "post_date": "2020-12-25T23:19:39.070000", "content": "Thanks for the Great Explanation
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "759268": "# CutMix and MixUp Augmentation\nI have posted a starter notebook [here][1] showing how to perform CutMix and MixUp on GPU/TPU using `TensorFlow.data.Dataset()`. Enjoy! \n\n\n\n# Data Augmentation\nData augmentation has repeatedly shown to improve model accuracy. Below is a list of some current literature on data augmentation techniques:\n## Basic - Flip, Rotation, Sheer, Zoom, Shift\nBasic augmentation includes flipping, rotation, sheer, zoom, and shift. Starter notebook [here][2]. Next we can incorporate non-linear transformations like grid distortion and we can do color augmentation like brightness, contrast, hue, etc. See [Albumentations][8] for ideas. \n## Cutout (Nov 2017)\nPDF [here][3]. Cutout randomly removes rectangular regions of the training images. It is similar to dropout. It teaches a network to use the entire image and not depend too much on certain regions or details in the image.\n\n\n\n## Mixup (Apr 2018)\nPDF [here][4]. Mixup blends two images and their one hot encoded labels to create a new image. This helps a network generalize to the space of images they lie \"between\" the existing training images.\n\n\n\n## AutoAugment (Apr 2019)\nPDF [here][5]. List a variety of augmentation techniques. Use reinforcement learning to explore different combinations of augmentations and find the best.\n\n\n\n## CutMix (Aug 2019)\nPDF [here][6]. This is cutout but we fill the blank rectangular space with a portion of another image. CutMix has the benefits of cutout (described above) and it expands the the space of images by creating new images (i.e. like a dog with a cat's head).\n\n\n\n## AugMix (2020)\nPDF [here][7]. Randomly blend multiple images that have each been augmented themselves separately. And use a special consistency loss.\n\n\n\n[1]: https://www.kaggle.com/cdeotte/cutmix-and-mixup-on-gpu-tpu\n[2]: https://www.kaggle.com/cdeotte/rotation-augmentation-gpu-tpu-0-96\n[3]: https://arxiv.org/abs/1708.04552\n[4]: https://arxiv.org/abs/1710.09412\n[5]: https://arxiv.org/abs/1805.09501\n[6]: https://arxiv.org/abs/1905.04899\n[7]: https://arxiv.org/abs/1912.02781\n[8]: https://albumentations.readthedocs.io/en/latest/", "772436": "Great work!", "765422": "What I always liked in Chris that he always dives deep into the thing, trying not only understand **how** it works, but also understand **why** it works.", "760160": "great dude", "759405": "Now, this is amazing ! Thank you.", "759286": "Oddsome topic (even with my poor knowledge I can appreciate it).", "1265698": " is amazing", "1076156": "Thanks for such an awesome explanation. I have a query.. In medical data like chest Xray data, is it an good idea to add cutmix or augmix? ", "843805": "Thanks for sharing this! This was one of the highlights of this competition for me, I didn't know about those augmentation methods. ", "777781": "I tried using your cutmix implementation but instead train a network once and not in a cross validation way but I get memory error. What would be the problem in this situation ?", "829896": "", "809115": "", "759803": "Now, this is amazing ! Thank you.\n\n", "1126774": "Thanks for the Great Explanation " } }