After reading many top solutions from last year's Plant Pathology competition. Here are some insights:
\nSame images with different labels. Some images in the training dataset were generated by one image, but they had different labels. Many top solutions used techniques like Knowledge Distillation, Label Smoothing, and Pseudo-Labelling.
\nFew people used upsampling to balance the dataset.
\nAdvice: A wrong classification of a sample with multiple disease categories will have significant impact on the final results. So, don’t trust the public leaderboard, trust your CV.
\nPeople used different many different model but EfficientNet B(0-7) appeared most frequently. Model pre-trained on noisy-student dataset tend to perform better. And finally, larger images size (600+) tend to give better results.
\nI will be testing all these ideas in the current competition. Stay tuned !
", "messages": [ { "id": "1241412", "postDate": "03/17/2021 03:12:15", "content": "After reading many top solutions from last year's Plant Pathology competition. Here are some insights:
\nSame images with different labels. Some images in the training dataset were generated by one image, but they had different labels. Many top solutions used techniques like Knowledge Distillation, Label Smoothing, and Pseudo-Labelling.
\nFew people used upsampling to balance the dataset.
\nAdvice: A wrong classification of a sample with multiple disease categories will have significant impact on the final results. So, don’t trust the public leaderboard, trust your CV.
\nPeople used different many different model but EfficientNet B(0-7) appeared most frequently. Model pre-trained on noisy-student dataset tend to perform better. And finally, larger images size (600+) tend to give better results.
\nI will be testing all these ideas in the current competition. Stay tuned !
", "rawMarkdown": "After reading many top solutions from last year's [Plant Pathology competition](https://www.kaggle.com/c/plant-pathology-2020-fgvc7). Here are some insights:\n\n#### 1. Noisy Dataset\n\nSame images with different labels. Some images in the training dataset were generated by one image, but they had different labels. Many top solutions used techniques like **Knowledge Distillation**, **Label Smoothing**, and **Pseudo-Labelling**.\n\n#### 2. Imbalance Dataset\n\nFew people used **upsampling** to balance the dataset.\n\n**Advice:** A wrong classification of a sample with multiple disease categories will have significant impact on the final results. So, don’t trust the public leaderboard, trust your CV.\n\n#### 3. Model architectures\n\nPeople used different many different model but EfficientNet B(0-7) appeared most frequently. Model pre-trained on noisy-student dataset tend to perform better. And finally, larger images size (600+) tend to give better results. \n\nI will be testing all these ideas in the current competition. Stay tuned !", "votes": null }, { "id": "1243322", "postDate": "03/18/2021 06:28:51", "content": "@ankursingh12 Thanks so much for the curated insights Ankur . Very Helpful
", "rawMarkdown": "ankursingh12 Thanks so much for the curated insights Ankur . Very Helpful", "votes": null }, { "id": "1244090", "postDate": "03/18/2021 17:33:18", "content": "Trying my best to learn! Between, nice name!
", "rawMarkdown": "Trying my best to learn! Between, nice name!", "votes": null }, { "id": "1248707", "postDate": "03/22/2021 19:12:52", "content": "Thanks @ankursingh12, also believe last year we had a category for multiple diseases and single disease only, now we have each combination of disease which makes it a bit more challenging.
", "rawMarkdown": "Thanks @ankursingh12, also believe last year we had a category for multiple diseases and single disease only, now we have each combination of disease which makes it a bit more challenging.", "votes": null }, { "id": "1250307", "postDate": "03/23/2021 23:43:12", "content": "Agreed, @saurabhbagchi !
\nIt's actually, pretty complex. I have seen a lot of people approaching it as multi-class (12 classes ) classification problem. Technically, its a multi-label problem. But because of class imbalance, its much difficult to get a good score.
\nI think, multi-class approach is good for now. But as the competition progresses, it will start following short. What is your take?
", "rawMarkdown": "Agreed, @saurabhbagchi ! \n\nIt's actually, pretty complex. I have seen a lot of people approaching it as multi-class (12 classes ) classification problem. Technically, its a multi-label problem. But because of class imbalance, its much difficult to get a good score. \n\nI think, multi-class approach is good for now. But as the competition progresses, it will start following short. What is your take?", "votes": null }, { "id": "1250494", "postDate": "03/24/2021 04:53:57", "content": "Yes class imbalance is difficult to tackle specially on unseen data, whoever handles it better through the code will prevail.
", "rawMarkdown": "Yes class imbalance is difficult to tackle specially on unseen data, whoever handles it better through the code will prevail.", "votes": null } ], "comments": [ { "id": 1243322, "author_name": "usharengaraju", "author_url": "", "post_date": "03/18/2021 06:28:51", "content": "@ankursingh12 Thanks so much for the curated insights Ankur . Very Helpful
", "votes": null, "replies": [ { "id": 1244090, "author_name": "ankursingh12", "author_url": "", "post_date": "03/18/2021 17:33:18", "content": "Trying my best to learn! Between, nice name!
", "votes": null, "replies": [] } ] }, { "id": 1248707, "author_name": "saurabhbagchi", "author_url": "", "post_date": "03/22/2021 19:12:52", "content": "Thanks @ankursingh12, also believe last year we had a category for multiple diseases and single disease only, now we have each combination of disease which makes it a bit more challenging.
", "votes": null, "replies": [ { "id": 1250307, "author_name": "ankursingh12", "author_url": "", "post_date": "03/23/2021 23:43:12", "content": "Agreed, @saurabhbagchi !
\nIt's actually, pretty complex. I have seen a lot of people approaching it as multi-class (12 classes ) classification problem. Technically, its a multi-label problem. But because of class imbalance, its much difficult to get a good score.
\nI think, multi-class approach is good for now. But as the competition progresses, it will start following short. What is your take?
", "votes": null, "replies": [] }, { "id": 1250494, "author_name": "saurabhbagchi", "author_url": "", "post_date": "03/24/2021 04:53:57", "content": "Yes class imbalance is difficult to tackle specially on unseen data, whoever handles it better through the code will prevail.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1241412": "After reading many top solutions from last year's [Plant Pathology competition](https://www.kaggle.com/c/plant-pathology-2020-fgvc7). Here are some insights:\n\n#### 1. Noisy Dataset\n\nSame images with different labels. Some images in the training dataset were generated by one image, but they had different labels. Many top solutions used techniques like **Knowledge Distillation**, **Label Smoothing**, and **Pseudo-Labelling**.\n\n#### 2. Imbalance Dataset\n\nFew people used **upsampling** to balance the dataset.\n\n**Advice:** A wrong classification of a sample with multiple disease categories will have significant impact on the final results. So, don’t trust the public leaderboard, trust your CV.\n\n#### 3. Model architectures\n\nPeople used different many different model but EfficientNet B(0-7) appeared most frequently. Model pre-trained on noisy-student dataset tend to perform better. And finally, larger images size (600+) tend to give better results. \n\nI will be testing all these ideas in the current competition. Stay tuned !", "1243322": "ankursingh12 Thanks so much for the curated insights Ankur . Very Helpful", "1244090": "Trying my best to learn! Between, nice name!", "1248707": "Thanks @ankursingh12, also believe last year we had a category for multiple diseases and single disease only, now we have each combination of disease which makes it a bit more challenging.", "1250307": "Agreed, @saurabhbagchi ! \n\nIt's actually, pretty complex. I have seen a lot of people approaching it as multi-class (12 classes ) classification problem. Technically, its a multi-label problem. But because of class imbalance, its much difficult to get a good score. \n\nI think, multi-class approach is good for now. But as the competition progresses, it will start following short. What is your take?", "1250494": "Yes class imbalance is difficult to tackle specially on unseen data, whoever handles it better through the code will prevail." }, "source": "meta" }