Last summer (April-May) I participated in the same competition with much less data. That motivated me to try different things to solve the problem.
\nThe main challenges main:
\nI tried to divide the challenges and combine their solutions, which led to BiLinear EfficientNet Focal Loss+ Label Smoothing.
\ni) BiLinear EfficientNet for Fine grain classification: Effiecient Net was state of the art last year, you can use the latest CNN backbone too.
\nii) Focal Loss+ Label Smoothing for Class imbalance: Tried various loss functions to improve the training where we have less samples of target class.
The code, results and other details of the solution can be found in this Kaggle Notebook: https://www.kaggle.com/jimitshah777/bilinear-efficientnet-focal-loss-label-smoothing
\nBest of luck!
", "messages": [ { "id": "1293271", "postDate": "05/04/2021 17:49:59", "content": "Last summer (April-May) I participated in the same competition with much less data. That motivated me to try different things to solve the problem.
\nThe main challenges main:
\nI tried to divide the challenges and combine their solutions, which led to BiLinear EfficientNet Focal Loss+ Label Smoothing.
\ni) BiLinear EfficientNet for Fine grain classification: Effiecient Net was state of the art last year, you can use the latest CNN backbone too.
\nii) Focal Loss+ Label Smoothing for Class imbalance: Tried various loss functions to improve the training where we have less samples of target class.
The code, results and other details of the solution can be found in this Kaggle Notebook: https://www.kaggle.com/jimitshah777/bilinear-efficientnet-focal-loss-label-smoothing
\nBest of luck!
", "rawMarkdown": "Last summer (April-May) I participated in the same competition with much less data. That motivated me to try different things to solve the problem. \n\nThe main challenges main:\n1. Fine grain details\n2. Class imbalance \n\nI tried to divide the challenges and combine their solutions, which led to [BiLinear EfficientNet Focal Loss+ Label Smoothing](https://www.kaggle.com/jimitshah777/bilinear-efficientnet-focal-loss-label-smoothing). \n\ni) BiLinear EfficientNet for Fine grain classification: Effiecient Net was state of the art last year, you can use the latest CNN backbone too.\nii) Focal Loss+ Label Smoothing for Class imbalance: Tried various loss functions to improve the training where we have less samples of target class.\n\nThe code, results and other details of the solution can be found in this Kaggle Notebook: https://www.kaggle.com/jimitshah777/bilinear-efficientnet-focal-loss-label-smoothing\n\nBest of luck!", "votes": null }, { "id": "1293369", "postDate": "05/04/2021 19:11:13", "content": "Looks like nice idea, I'll try it. Thanks for sharing this.
", "rawMarkdown": "Looks like nice idea, I'll try it. Thanks for sharing this.", "votes": null }, { "id": "1293718", "postDate": "05/05/2021 05:51:51", "content": "Hello. Thanks to share this awesome information to us :)
\nI have a question. I think Label Smoothing can be applied about multi-class classification problem.
\nBut, is it okay to use Label Smoothing method to multi-label classification problem?
Yes you can use it in this problem too. What label smoothing basically does is regularization. Due to this, the model doesn't learn too confidently. Hence it explores more paths, thus leading to generalization.
", "rawMarkdown": "Yes you can use it in this problem too. What label smoothing basically does is regularization. Due to this, the model doesn't learn too confidently. Hence it explores more paths, thus leading to generalization.", "votes": null }, { "id": "1294471", "postDate": "05/05/2021 17:14:53", "content": "Thank you for your reply :) Yes, I agree with your opinion. Because the dataset from competition wouldn't show the distribution of real data. By the way in this problem, the label 'healthy' cannot be existed with other labels as far as I know. Do you think label smoothing still can be used?
", "rawMarkdown": "Thank you for your reply :) Yes, I agree with your opinion. Because the dataset from competition wouldn't show the distribution of real data. By the way in this problem, the label 'healthy' cannot be existed with other labels as far as I know. Do you think label smoothing still can be used?", "votes": null }, { "id": "1297095", "postDate": "05/07/2021 18:25:28", "content": "I am not sure about the problem statement and how this particular competition works.
\nBut when it comes to pathology of any kind (in plants or in humans), I have observed that there is always a class imbalance, because finding the diseased class is rare than that of healthy.
Looks like nice idea, I'll try it. Thanks for sharing this.
", "votes": null, "replies": [] }, { "id": 1293718, "author_name": "ljh0128", "author_url": "", "post_date": "05/05/2021 05:51:51", "content": "Hello. Thanks to share this awesome information to us :)
\nI have a question. I think Label Smoothing can be applied about multi-class classification problem.
\nBut, is it okay to use Label Smoothing method to multi-label classification problem?
Yes you can use it in this problem too. What label smoothing basically does is regularization. Due to this, the model doesn't learn too confidently. Hence it explores more paths, thus leading to generalization.
", "votes": null, "replies": [] }, { "id": 1294471, "author_name": "ljh0128", "author_url": "", "post_date": "05/05/2021 17:14:53", "content": "Thank you for your reply :) Yes, I agree with your opinion. Because the dataset from competition wouldn't show the distribution of real data. By the way in this problem, the label 'healthy' cannot be existed with other labels as far as I know. Do you think label smoothing still can be used?
", "votes": null, "replies": [] }, { "id": 1297095, "author_name": "jimitshah777", "author_url": "", "post_date": "05/07/2021 18:25:28", "content": "I am not sure about the problem statement and how this particular competition works.
\nBut when it comes to pathology of any kind (in plants or in humans), I have observed that there is always a class imbalance, because finding the diseased class is rare than that of healthy.