Hi everyone , I know this is a silly question to ask when only 3 weeks are left for the deadline but unfortunately I cannot get an accuracy greater than 65%. I have tried to follow other notebooks and I have also taken into account some of the concepts that has been discussed in the forum.
\nHi everyone , I know this is a silly question to ask when only 3 weeks are left for the deadline but unfortunately I cannot get an accuracy greater than 65%. I have tried to follow other notebooks and I have also taken into account some of the concepts that has been discussed in the forum.
\nHello!
\nI'd suggest starting with these great notebooks by Kaggle community and by DimitreOliveira and then carefully replacing/adding some lines step by step, introducing your own ideas and style (?) on your way of understanding and learning. The most important part here is not to break the code, which you probably did somehow if you followed this way.
\nAs long as the accuracy of 65% is pretty close to this Naive Baseline of 60%, I suppose that your model didn't actually learn anything else but mostly predicting the most common class. This can be due to such small slip like:
\nI would suggest reading up https://www.kaggle.com/khyeh0719/pytorch-efficientnet-baseline-train-amp-aug which is a really good starter notebook (in Pytorch). And once you are comfortable with it, you can tune it further from there.
", "rawMarkdown": "I would suggest reading up https://www.kaggle.com/khyeh0719/pytorch-efficientnet-baseline-train-amp-aug which is a really good starter notebook (in Pytorch). And once you are comfortable with it, you can tune it further from there.", "votes": null }, { "id": "1179733", "postDate": "01/31/2021 19:10:34", "content": "Thank you so much @angqx95 and @nickuzmenkov for your valuable suggestions.
", "rawMarkdown": "Thank you so much @angqx95 and @nickuzmenkov for your valuable suggestions.", "votes": null }, { "id": "1179850", "postDate": "01/31/2021 22:17:59", "content": "More specific help can be obtained if you make your kernel public and than add a link here. There are just too many possible reasons for your accuracy values for folks to give you help that can quickly correct your issue.
", "rawMarkdown": "More specific help can be obtained if you make your kernel public and than add a link here. There are just too many possible reasons for your accuracy values for folks to give you help that can quickly correct your issue.", "votes": null }, { "id": "1180127", "postDate": "02/01/2021 05:32:16", "content": "Hi Arka, This is because of the class imbalance problem. Either do up sampling or down sampling technique to get uniform data. If you wish to use data as it is I would suggest go for Bagging or Boosting technique as they are very good at pattern recognition or you can also try Artificial Neural Networks to see unseen pattern within data. Hope it helps.
", "rawMarkdown": "Hi Arka, This is because of the class imbalance problem. Either do up sampling or down sampling technique to get uniform data. If you wish to use data as it is I would suggest go for Bagging or Boosting technique as they are very good at pattern recognition or you can also try Artificial Neural Networks to see unseen pattern within data. Hope it helps.", "votes": null }, { "id": "1181101", "postDate": "02/01/2021 16:44:03", "content": "Thank you @pcjimmmy for your reply. As suggested I have made the notebook public. Here is the URL: https://www.kaggle.com/arka1993/ab-cassava-leaf-disease-classification. This is not a clean notebook because I am working on it, but I hope you can get an overview on what I have done. If it is not at all understandable, then please let me know. I will try to clean as much as possible.
", "rawMarkdown": "Thank you @pcjimmmy for your reply. As suggested I have made the notebook public. Here is the URL: https://www.kaggle.com/arka1993/ab-cassava-leaf-disease-classification. This is not a clean notebook because I am working on it, but I hope you can get an overview on what I have done. If it is not at all understandable, then please let me know. I will try to clean as much as possible.", "votes": null }, { "id": "1181114", "postDate": "02/01/2021 16:51:26", "content": "Hi Harpreet @harpreet245 , thank you for your valuable suggestion. I have tried the up sampling and down sampling concept that you have mentioned and it did not work. I think I am messing up the data loading part. Let's see how it works out at the end.
", "rawMarkdown": "Hi Harpreet @harpreet245 , thank you for your valuable suggestion. I have tried the up sampling and down sampling concept that you have mentioned and it did not work. I think I am messing up the data loading part. Let's see how it works out at the end.", "votes": null }, { "id": "1184368", "postDate": "02/03/2021 14:02:39", "content": "I have created a post for the same purpose here
", "rawMarkdown": "I have created a post for the same purpose [here](https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/216461)", "votes": null } ], "comments": [ { "id": 1179626, "author_name": "nickuzmenkov", "author_url": "", "post_date": "01/31/2021 17:28:09", "content": "Hello!
\nI'd suggest starting with these great notebooks by Kaggle community and by DimitreOliveira and then carefully replacing/adding some lines step by step, introducing your own ideas and style (?) on your way of understanding and learning. The most important part here is not to break the code, which you probably did somehow if you followed this way.
\nAs long as the accuracy of 65% is pretty close to this Naive Baseline of 60%, I suppose that your model didn't actually learn anything else but mostly predicting the most common class. This can be due to such small slip like:
\nI would suggest reading up https://www.kaggle.com/khyeh0719/pytorch-efficientnet-baseline-train-amp-aug which is a really good starter notebook (in Pytorch). And once you are comfortable with it, you can tune it further from there.
", "votes": null, "replies": [] }, { "id": 1179733, "author_name": "arka1993", "author_url": "", "post_date": "01/31/2021 19:10:34", "content": "Thank you so much @angqx95 and @nickuzmenkov for your valuable suggestions.
", "votes": null, "replies": [] }, { "id": 1179850, "author_name": "pcjimmmy", "author_url": "", "post_date": "01/31/2021 22:17:59", "content": "More specific help can be obtained if you make your kernel public and than add a link here. There are just too many possible reasons for your accuracy values for folks to give you help that can quickly correct your issue.
", "votes": null, "replies": [] }, { "id": 1180127, "author_name": "harpreet245", "author_url": "", "post_date": "02/01/2021 05:32:16", "content": "Hi Arka, This is because of the class imbalance problem. Either do up sampling or down sampling technique to get uniform data. If you wish to use data as it is I would suggest go for Bagging or Boosting technique as they are very good at pattern recognition or you can also try Artificial Neural Networks to see unseen pattern within data. Hope it helps.
", "votes": null, "replies": [] }, { "id": 1181101, "author_name": "arka1993", "author_url": "", "post_date": "02/01/2021 16:44:03", "content": "Thank you @pcjimmmy for your reply. As suggested I have made the notebook public. Here is the URL: https://www.kaggle.com/arka1993/ab-cassava-leaf-disease-classification. This is not a clean notebook because I am working on it, but I hope you can get an overview on what I have done. If it is not at all understandable, then please let me know. I will try to clean as much as possible.
", "votes": null, "replies": [] }, { "id": 1181114, "author_name": "arka1993", "author_url": "", "post_date": "02/01/2021 16:51:26", "content": "Hi Harpreet @harpreet245 , thank you for your valuable suggestion. I have tried the up sampling and down sampling concept that you have mentioned and it did not work. I think I am messing up the data loading part. Let's see how it works out at the end.
", "votes": null, "replies": [] }, { "id": 1184368, "author_name": "vickygoyal", "author_url": "", "post_date": "02/03/2021 14:02:39", "content": "I have created a post for the same purpose here
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1179334": "Hi everyone , I know this is a silly question to ask when only 3 weeks are left for the deadline but unfortunately I cannot get an accuracy greater than 65%. I have tried to follow other notebooks and I have also taken into account some of the concepts that has been discussed in the forum.\n1. I have tried to use the merged dataset (2019+2020).\n2. I have tried different concepts like learning rate scheduling.\n3. I have tried EfficientNet, InceptionResnet, Only ResNet as some other people have implemented. But all of them are getting around 88%, whereas I am stuck with 65%. \nI am new in this domain. If someone have faced a similar issue, could you please share your experience. I mean even if I can get upto 75%, I will be happy. \nAny help is highly appreciated. Thank you.", "1179626": "Hello!\n\nI'd suggest starting with these great notebooks **[by Kaggle community](https://www.kaggle.com/jessemostipak/getting-started-tpus-cassava-leaf-disease)** and **[by DimitreOliveira](https://www.kaggle.com/dimitreoliveira/cassava-leaf-disease-tpu-tensorflow-training)** and then carefully replacing/adding some lines step by step, introducing your own ideas and style (?) on your way of understanding and learning. The most important part here is not to break the code, which you probably did somehow if you followed this way.\n\nAs long as the accuracy of 65% is pretty close to this **[Naive Baseline](https://www.kaggle.com/nickuzmenkov/cassava-leaf-disease-na-ve-baseline)** of 60%, I suppose that your model didn't actually learn anything else but mostly predicting the most common class. This can be due to such small slip like:\n* no image normalization\n* no dataset shuffle (e.g. the training dataset contains some labels, and the validation dataset - the other labels)", "1179665": "I would suggest reading up https://www.kaggle.com/khyeh0719/pytorch-efficientnet-baseline-train-amp-aug which is a really good starter notebook (in Pytorch). And once you are comfortable with it, you can tune it further from there.", "1179733": "Thank you so much @angqx95 and @nickuzmenkov for your valuable suggestions.", "1179850": "More specific help can be obtained if you make your kernel public and than add a link here. There are just too many possible reasons for your accuracy values for folks to give you help that can quickly correct your issue.", "1180127": "Hi Arka, This is because of the class imbalance problem. Either do up sampling or down sampling technique to get uniform data. If you wish to use data as it is I would suggest go for Bagging or Boosting technique as they are very good at pattern recognition or you can also try Artificial Neural Networks to see unseen pattern within data. Hope it helps.", "1181101": "Thank you @pcjimmmy for your reply. As suggested I have made the notebook public. Here is the URL: https://www.kaggle.com/arka1993/ab-cassava-leaf-disease-classification. This is not a clean notebook because I am working on it, but I hope you can get an overview on what I have done. If it is not at all understandable, then please let me know. I will try to clean as much as possible.", "1181114": "Hi Harpreet @harpreet245 , thank you for your valuable suggestion. I have tried the up sampling and down sampling concept that you have mentioned and it did not work. I think I am messing up the data loading part. Let's see how it works out at the end.", "1184368": "I have created a post for the same purpose [here](https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/216461)" }, "source": "meta" }