What augmentations do you use?I use VerticalFlip, HorizontalFlip, ShiftScaleRotate and RandomBrightness. My score ResNet 50 - 0.893
", "messages": [ { "id": "1179638", "postDate": "01/31/2021 17:38:00", "content": "What augmentations do you use?I use VerticalFlip, HorizontalFlip, ShiftScaleRotate and RandomBrightness. My score ResNet 50 - 0.893
", "rawMarkdown": "What augmentations do you use?I use VerticalFlip, HorizontalFlip, ShiftScaleRotate and RandomBrightness. My score ResNet 50 - 0.893", "votes": null }, { "id": "1179645", "postDate": "01/31/2021 17:44:48", "content": "you can try mwh : https://github.com/yuhao318/mwh
\nit gave me 0.003+ boost in cv
If I am not wrong this implementation is in PyTorch, right?
\nIs there anything for keras?
\nThis huge community for PyTorch really wanna make a switch to it, but I just learned keras. Sad life.
@mobassir you use the exact same code or the one is commented?
", "rawMarkdown": "mobassir you use the exact same code or the one is commented?", "votes": null }, { "id": "1180729", "postDate": "02/01/2021 12:46:03", "content": "dear @mohneesh7 i haven't seen any tf implementation,,i mostly use torch for any computer vision task,,so sorry about that
\ndear @projdev i use this code exactly for my experiment : https://github.com/yuhao318/mwh/blob/main/utils.py
", "rawMarkdown": "dear @mohneesh7 i haven't seen any tf implementation,,i mostly use torch for any computer vision task,,so sorry about that\n\ndear @projdev i use this code exactly for my experiment : https://github.com/yuhao318/mwh/blob/main/utils.py", "votes": null }, { "id": "1180804", "postDate": "02/01/2021 13:34:15", "content": "I found that a larger epoch (100 +) was used in the initial implementation of training from scratch. I would like to ask, in this competition, do you apply this strategy to the small epoch training (using pre-training).
\nOf course, if you don't want to reveal too many details before the end of the game, I can understand. :)
@chinesewuji it converged within 40 epoch for me but if you mix other augmentations with it then it might won't work well for you
", "rawMarkdown": "chinesewuji it converged within 40 epoch for me but if you mix other augmentations with it then it might won't work well for you", "votes": null }, { "id": "1180913", "postDate": "02/01/2021 14:32:59", "content": "Man I hate this, after competeing in this competition I came to the importance of pytorch. Will definitely try to learn pytorch next.
", "rawMarkdown": "Man I hate this, after competeing in this competition I came to the importance of pytorch. Will definitely try to learn pytorch next.", "votes": null }, { "id": "1180989", "postDate": "02/01/2021 15:34:54", "content": "Thanks for sharing the strategy！
", "rawMarkdown": "Thanks for sharing the strategy！", "votes": null }, { "id": "1182208", "postDate": "02/02/2021 11:05:10", "content": "Thanks,
\nI've tried mwh with epochs 40
\nbut from 10 epoch, Score & valid loss don't get better
\nIs it okay to use Taylor cross entropy loss with MWH?
\n(if you don't want you don have to tell about it, Thanks for great Strategy)
@deepbluebird yes i tried Taylor cross entropy loss with mwh and it worked for me though
", "rawMarkdown": "deepbluebird yes i tried Taylor cross entropy loss with mwh and it worked for me though", "votes": null }, { "id": "1182434", "postDate": "02/02/2021 13:03:06", "content": "I use VerticalFlip, HorizontalFlip, and RandomZoom. Plus I use TTA with RandomContrast. My score is EfficientNetB4 - 0.895… I tried heavy aug like dropout, score got worse in my case.
", "rawMarkdown": "I use VerticalFlip, HorizontalFlip, and RandomZoom. Plus I use TTA with RandomContrast. My score is EfficientNetB4 - 0.895... I tried heavy aug like dropout, score got worse in my case.", "votes": null }, { "id": "1182794", "postDate": "02/02/2021 15:43:13", "content": "snapmix, fmix helped me boost LB by 0.004.
", "rawMarkdown": "snapmix, fmix helped me boost LB by 0.004.", "votes": null } ], "comments": [ { "id": 1179645, "author_name": "mobassir", "author_url": "", "post_date": "01/31/2021 17:44:48", "content": "you can try mwh : https://github.com/yuhao318/mwh
\nit gave me 0.003+ boost in cv
If I am not wrong this implementation is in PyTorch, right?
\nIs there anything for keras?
\nThis huge community for PyTorch really wanna make a switch to it, but I just learned keras. Sad life.
@mobassir you use the exact same code or the one is commented?
", "votes": null, "replies": [] }, { "id": 1180729, "author_name": "mobassir", "author_url": "", "post_date": "02/01/2021 12:46:03", "content": "dear @mohneesh7 i haven't seen any tf implementation,,i mostly use torch for any computer vision task,,so sorry about that
\ndear @projdev i use this code exactly for my experiment : https://github.com/yuhao318/mwh/blob/main/utils.py
", "votes": null, "replies": [] }, { "id": 1180804, "author_name": "chinesewuji", "author_url": "", "post_date": "02/01/2021 13:34:15", "content": "I found that a larger epoch (100 +) was used in the initial implementation of training from scratch. I would like to ask, in this competition, do you apply this strategy to the small epoch training (using pre-training).
\nOf course, if you don't want to reveal too many details before the end of the game, I can understand. :)
@chinesewuji it converged within 40 epoch for me but if you mix other augmentations with it then it might won't work well for you
", "votes": null, "replies": [] }, { "id": 1180913, "author_name": "mohneesh7", "author_url": "", "post_date": "02/01/2021 14:32:59", "content": "Man I hate this, after competeing in this competition I came to the importance of pytorch. Will definitely try to learn pytorch next.
", "votes": null, "replies": [] }, { "id": 1180989, "author_name": "chinesewuji", "author_url": "", "post_date": "02/01/2021 15:34:54", "content": "Thanks for sharing the strategy！
", "votes": null, "replies": [] }, { "id": 1182208, "author_name": "deepbluebird", "author_url": "", "post_date": "02/02/2021 11:05:10", "content": "Thanks,
\nI've tried mwh with epochs 40
\nbut from 10 epoch, Score & valid loss don't get better
\nIs it okay to use Taylor cross entropy loss with MWH?
\n(if you don't want you don have to tell about it, Thanks for great Strategy)
@deepbluebird yes i tried Taylor cross entropy loss with mwh and it worked for me though
", "votes": null, "replies": [] } ] }, { "id": 1182434, "author_name": "vkehfdl1", "author_url": "", "post_date": "02/02/2021 13:03:06", "content": "I use VerticalFlip, HorizontalFlip, and RandomZoom. Plus I use TTA with RandomContrast. My score is EfficientNetB4 - 0.895… I tried heavy aug like dropout, score got worse in my case.
", "votes": null, "replies": [] }, { "id": 1182794, "author_name": "tamilselvanmoorthy", "author_url": "", "post_date": "02/02/2021 15:43:13", "content": "snapmix, fmix helped me boost LB by 0.004.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1179638": "What augmentations do you use?I use VerticalFlip, HorizontalFlip, ShiftScaleRotate and RandomBrightness. My score ResNet 50 - 0.893", "1179645": "you can try mwh : https://github.com/yuhao318/mwh\nit gave me 0.003+ boost in cv", "1180675": "If I am not wrong this implementation is in PyTorch, right?\nIs there anything for keras?\nThis huge community for PyTorch really wanna make a switch to it, but I just learned keras. Sad life.", "1180694": "mobassir you use the exact same code or the one is commented?", "1180729": "dear @mohneesh7 i haven't seen any tf implementation,,i mostly use torch for any computer vision task,,so sorry about that\n\ndear @projdev i use this code exactly for my experiment : https://github.com/yuhao318/mwh/blob/main/utils.py", "1180804": "I found that a larger epoch (100 +) was used in the initial implementation of training from scratch. I would like to ask, in this competition, do you apply this strategy to the small epoch training (using pre-training).\nOf course, if you don't want to reveal too many details before the end of the game, I can understand. :)", "1180812": "chinesewuji it converged within 40 epoch for me but if you mix other augmentations with it then it might won't work well for you", "1180913": "Man I hate this, after competeing in this competition I came to the importance of pytorch. Will definitely try to learn pytorch next.", "1180989": "Thanks for sharing the strategy！", "1182208": "Thanks, \nI've tried mwh with epochs 40\nbut from 10 epoch, Score & valid loss don't get better\nIs it okay to use Taylor cross entropy loss with MWH?\n(if you don't want you don have to tell about it, Thanks for great Strategy)", "1182211": "deepbluebird yes i tried Taylor cross entropy loss with mwh and it worked for me though", "1182434": "I use VerticalFlip, HorizontalFlip, and RandomZoom. Plus I use TTA with RandomContrast. My score is EfficientNetB4 - 0.895... I tried heavy aug like dropout, score got worse in my case.", "1182794": "snapmix, fmix helped me boost LB by 0.004." }, "source": "meta" }