{ "id": 243042, "title": "1st place solution", "url": "/competitions/plant-pathology-2021-fgvc8/discussion/243042", "author_name": "wanghao", "post_date": "2021-06-01T03:17:03.378000", "votes": 41, "comment_count": 4, "views": 0, "content": "

 

\n", "messages": [ { "id": 1330642, "postDate": "2021-06-01T03:17:03.380Z", "content": "

 

\n", "rawMarkdown": " \n+ **soft label
**\n       train 5-folds efficientnetv2
\n       label==1 & pred_score < 0.7 ==> soft_label=0.3(for train folds)\n
\n+ **multi-label augment
**\n + **cutmix
**\n       different from norm cutmix, I didn't use:
\n              loss = lam * loss(outputs, labels) + (1-lam) * loss(outputs, mix_labels)
\n       because I think 'healthy' + 'scab' == 'scab'

\n'scab'+'frog_eye_leaf_spot' ==> 'scab frog_eye_leaf_spot'(There are also other different combinations)\n![1.png](https://i.loli.net/2021/06/01/tAIaBcgWirRuwM9.png)\n\n + **mosaic
**\n'scab' + 'scab' + 'scab frog_eye_leaf_spot' + 'complex' ==>'scab frog_eye_leaf_spot complex' (There are also other different combinations)
\n![2.png](https://i.loli.net/2021/06/01/FDCuve7csdXROUJ.png)\n
\n+ **augment
** \nCrop HFlip VFlip brightness contrast ShiftScaleRotate OpticalDistortion GridDistortion IAAPiecewiseAffine Cutout CoarseDropout\n
\n+ **other
**\n       resnet50-5folds+resnext50_32x4d-5folds
\n       TTA(flip, centerCrop)
\n       class balance(last three epoch)
\n       label_smooth
\n       384x576
\n       warmup
\n       CosineAnnealingLR
\n       AdamW
\n+ **unfinished LeafGAN
**\nDue to time and computing resource constraints, I did not finish it, but I think it's the best way for multi-label
\nwe can train LeafGAN generate healthy->disease_A and healthy->disease_B, than we can get multi-label('disease_A disease_B') by mixup('disease_A', 'disease_B')
\nhttps://github.com/IyatomiLab/LeafGAN\n![3.png](https://i.loli.net/2021/06/01/LQB3dEtp8IrGfg2.png)", "votes": 41 }, { "id": 1360020, "postDate": "2021-06-21T18:22:46.817Z", "content": "

Congrats on 1st place!

", "rawMarkdown": "Congrats on 1st place!" }, { "id": 1332907, "postDate": "2021-06-02T11:31:34.727Z", "content": "

Thx, very helpful for me.))

", "rawMarkdown": "Thx, very helpful for me.))" }, { "id": 1331233, "postDate": "2021-06-01T10:51:29.877Z", "content": "

Thank you for sharing your solution. I have a few questions regarding your approach

\n\n

Sorry if any of these questions seem trivial, I am still learning a lot of stuff, so I am trying to gain as much insight into the decision processes as possible.

\n

Thank you

", "rawMarkdown": "Thank you for sharing your solution. I have a few questions regarding your approach\n- Did you use any ensambling techniques or just simply used average?\n- What is > lam> in the loss you have calculated?\n- Also how did you apply the class balance in the last three epochs.\n\nSorry if any of these questions seem trivial, I am still learning a lot of stuff, so I am trying to gain as much insight into the decision processes as possible.\n\nThank you\n" }, { "id": 1331319, "postDate": "2021-06-01T11:51:01.693Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 1360020, "author_name": "Okyanus", "author_url": "", "post_date": "2021-06-21T18:22:46.817000", "content": "

Congrats on 1st place!

", "votes": 0, "replies": [] }, { "id": 1332907, "author_name": "Oknedis", "author_url": "", "post_date": "2021-06-02T11:31:34.727000", "content": "

Thx, very helpful for me.))

", "votes": 0, "replies": [] }, { "id": 1331233, "author_name": "Gur Amrit Singh", "author_url": "", "post_date": "2021-06-01T10:51:29.877000", "content": "

Thank you for sharing your solution. I have a few questions regarding your approach

\n\n

Sorry if any of these questions seem trivial, I am still learning a lot of stuff, so I am trying to gain as much insight into the decision processes as possible.

\n

Thank you

", "votes": 0, "replies": [] }, { "id": 1331319, "author_name": "", "author_url": "", "post_date": "2021-06-01T11:51:01.693000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1330642": " \n+ **soft label
**\n       train 5-folds efficientnetv2
\n       label==1 & pred_score < 0.7 ==> soft_label=0.3(for train folds)\n
\n+ **multi-label augment
**\n + **cutmix
**\n       different from norm cutmix, I didn't use:
\n              loss = lam * loss(outputs, labels) + (1-lam) * loss(outputs, mix_labels)
\n       because I think 'healthy' + 'scab' == 'scab'

\n'scab'+'frog_eye_leaf_spot' ==> 'scab frog_eye_leaf_spot'(There are also other different combinations)\n![1.png](https://i.loli.net/2021/06/01/tAIaBcgWirRuwM9.png)\n\n + **mosaic
**\n'scab' + 'scab' + 'scab frog_eye_leaf_spot' + 'complex' ==>'scab frog_eye_leaf_spot complex' (There are also other different combinations)
\n![2.png](https://i.loli.net/2021/06/01/FDCuve7csdXROUJ.png)\n
\n+ **augment
** \nCrop HFlip VFlip brightness contrast ShiftScaleRotate OpticalDistortion GridDistortion IAAPiecewiseAffine Cutout CoarseDropout\n
\n+ **other
**\n       resnet50-5folds+resnext50_32x4d-5folds
\n       TTA(flip, centerCrop)
\n       class balance(last three epoch)
\n       label_smooth
\n       384x576
\n       warmup
\n       CosineAnnealingLR
\n       AdamW
\n+ **unfinished LeafGAN
**\nDue to time and computing resource constraints, I did not finish it, but I think it's the best way for multi-label
\nwe can train LeafGAN generate healthy->disease_A and healthy->disease_B, than we can get multi-label('disease_A disease_B') by mixup('disease_A', 'disease_B')
\nhttps://github.com/IyatomiLab/LeafGAN\n![3.png](https://i.loli.net/2021/06/01/LQB3dEtp8IrGfg2.png)", "1360020": "Congrats on 1st place!", "1332907": "Thx, very helpful for me.))", "1331233": "Thank you for sharing your solution. I have a few questions regarding your approach\n- Did you use any ensambling techniques or just simply used average?\n- What is > lam> in the loss you have calculated?\n- Also how did you apply the class balance in the last three epochs.\n\nSorry if any of these questions seem trivial, I am still learning a lot of stuff, so I am trying to gain as much insight into the decision processes as possible.\n\nThank you\n", "1331319": "" } }