![\"\"](\"https://i.postimg.cc/WzgyJfBr/image.png\")
Hello everyone.
\nIn this topic I will describe my approach, which placed me into gold zone in the first few days of the competition, but then New Year's holydays made me forgot about this competition for about a month. Also during that time I have faced something that I was never facing before in my kaggle competitions - lack of motivation to keep up. I trully admire people who kept working hard - no matter what you final score on the liderboard is - you are already winners.
\nI was using a combination of 3 datasets:
\nModels:
\n| Architecture | \nCV | \nPublic | \nPrivate | \nCode | \n
|---|---|---|---|---|
| resnet34 | \n0.8804 | \n0.892 | \n0.891 | \nlink | \n
| effnet-b3 | \n0.8871 | \n0.897 | \n0.897 | \nlink | \n
| effnet-b3 | \n0.8849 | \n0.898 | \n0.890 | \n--- | \n
As you can see on the picture below there were 3 submissions with 0.899 Private score, which might place me into a silver zone, but I did not choose any of them, which make a lot of sense - with such a low Cross-Validation scores I never even considered any of them as one of my final submissions.
\n
The ones I have chosen had better CV with higher Public LB.
\nBut as it was already said plenty of times - testing dataset contains a considerable amount of noise, just like training set. A challenge was to make a model robust enough to have good performance with such noise. And I have failed this challenge.
Anyway, besided combining 2020 and 2019 datasets with a dataset of my own I had no tricks. This is what worked for me from the very beginning:
\nWith everything that I have described I managed to score 0.902 on Public LB during the first week (or two?) of the competition and I remember that I was pretty surprised that this placed me into a gold zone at a time.
", "messages": [ { "id": "1210030", "postDate": "02/19/2021 06:58:03", "content": "Hello everyone.
\nIn this topic I will describe my approach, which placed me into gold zone in the first few days of the competition, but then New Year's holydays made me forgot about this competition for about a month. Also during that time I have faced something that I was never facing before in my kaggle competitions - lack of motivation to keep up. I trully admire people who kept working hard - no matter what you final score on the liderboard is - you are already winners.
\nI was using a combination of 3 datasets:
\nModels:
\n| Architecture | \nCV | \nPublic | \nPrivate | \nCode | \n
|---|---|---|---|---|
| resnet34 | \n0.8804 | \n0.892 | \n0.891 | \nlink | \n
| effnet-b3 | \n0.8871 | \n0.897 | \n0.897 | \nlink | \n
| effnet-b3 | \n0.8849 | \n0.898 | \n0.890 | \n--- | \n
As you can see on the picture below there were 3 submissions with 0.899 Private score, which might place me into a silver zone, but I did not choose any of them, which make a lot of sense - with such a low Cross-Validation scores I never even considered any of them as one of my final submissions.
\n
The ones I have chosen had better CV with higher Public LB.
\nBut as it was already said plenty of times - testing dataset contains a considerable amount of noise, just like training set. A challenge was to make a model robust enough to have good performance with such noise. And I have failed this challenge.
Anyway, besided combining 2020 and 2019 datasets with a dataset of my own I had no tricks. This is what worked for me from the very beginning:
\nWith everything that I have described I managed to score 0.902 on Public LB during the first week (or two?) of the competition and I remember that I was pretty surprised that this placed me into a gold zone at a time.
", "rawMarkdown": "## Greeting word\n\nHello everyone.\n\nIn this topic I will describe my approach, which placed me into gold zone in the first few days of the competition, but then New Year's holydays made me forgot about this competition for about a month. Also during that time I have faced something that I was never facing before in my kaggle competitions - lack of motivation to keep up. I trully admire people who kept working hard - no matter what you final score on the liderboard is - you are already winners.\n\n## Solution description\n\nI was using a combination of 3 datasets:\n1. Original dataset, provided for this competition.\n2. Dataset from [cassava 2019](https://www.kaggle.com/c/cassava-disease)\n3. A [dataset](https://www.kaggle.com/nroman/mendeley-leaves) that I have created. You can read more about it [here](https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/199902).\n\nModels:\n| Architecture | CV | Public | Private | Code |\n| --- | --- | --- | --- | --- |\n| resnet34 | 0.8804 | 0.892 | 0.891 | [link](https://github.com/CaypoH/kaggle_cassava/blob/main/resnet34.ipynb) |\n| effnet-b3 | 0.8871 | 0.897 | 0.897 | [link](https://github.com/CaypoH/kaggle_cassava/blob/main/efficientnet_b3_fold_3.ipynb) | \n| effnet-b3 | 0.8849 | 0.898 | 0.890 | --- |\n\n\nAs you can see on the picture below there were 3 submissions with 0.899 Private score, which might place me into a silver zone, but I did not choose any of them, which make a lot of sense - with such a low Cross-Validation scores I never even considered any of them as one of my final submissions. \n\n\n\nThe ones I have chosen had better CV with higher Public LB.\nBut as it was [already said](https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/202673) plenty of times - testing dataset contains a considerable amount of noise, just like training set. A challenge was to make a model robust enough to have good performance with such noise. And I have failed this challenge.\n\nAnyway, besided combining 2020 and 2019 datasets with a dataset of my own I had no tricks. This is what worked for me from the very beginning:\n* Smaller models -> better result: resnet34 > resnet101, effnet-b3 > effnet-b7\n* Lower TTA: No tta was better than 8-10 TTA rounds.\n* Less augmentations: heavy complex augmentations were losing to simple ones. CutMix and MixUp were even worse.\n\nWith everything that I have described I managed to score 0.902 on Public LB during the first week (or two?) of the competition and I remember that I was pretty surprised that this placed me into a gold zone at a time.", "votes": null }, { "id": "1210047", "postDate": "02/19/2021 07:09:52", "content": "Yeah, holidays and other stuff do get in the way.. However do note, family is more important than anything. So don't regret anything! You still did really good, and after reading your discussion, I would love to make more late submissions trying these tactics out. Thanks for solution, family first!
", "rawMarkdown": "Yeah, holidays and other stuff do get in the way.. However do note, family is more important than anything. So don't regret anything! You still did really good, and after reading your discussion, I would love to make more late submissions trying these tactics out. Thanks for solution, family first!", "votes": null }, { "id": "1210053", "postDate": "02/19/2021 07:13:26", "content": "Good job and well done! Next time you can get a medal. :)
", "rawMarkdown": "Good job and well done! Next time you can get a medal. :)", "votes": null } ], "comments": [ { "id": 1210047, "author_name": "andyjianzhou", "author_url": "", "post_date": "02/19/2021 07:09:52", "content": "Yeah, holidays and other stuff do get in the way.. However do note, family is more important than anything. So don't regret anything! You still did really good, and after reading your discussion, I would love to make more late submissions trying these tactics out. Thanks for solution, family first!
", "votes": null, "replies": [] }, { "id": 1210053, "author_name": "piantic", "author_url": "", "post_date": "02/19/2021 07:13:26", "content": "Good job and well done! Next time you can get a medal. :)
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1210030": "## Greeting word\n\nHello everyone.\n\nIn this topic I will describe my approach, which placed me into gold zone in the first few days of the competition, but then New Year's holydays made me forgot about this competition for about a month. Also during that time I have faced something that I was never facing before in my kaggle competitions - lack of motivation to keep up. I trully admire people who kept working hard - no matter what you final score on the liderboard is - you are already winners.\n\n## Solution description\n\nI was using a combination of 3 datasets:\n1. Original dataset, provided for this competition.\n2. Dataset from [cassava 2019](https://www.kaggle.com/c/cassava-disease)\n3. A [dataset](https://www.kaggle.com/nroman/mendeley-leaves) that I have created. You can read more about it [here](https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/199902).\n\nModels:\n| Architecture | CV | Public | Private | Code |\n| --- | --- | --- | --- | --- |\n| resnet34 | 0.8804 | 0.892 | 0.891 | [link](https://github.com/CaypoH/kaggle_cassava/blob/main/resnet34.ipynb) |\n| effnet-b3 | 0.8871 | 0.897 | 0.897 | [link](https://github.com/CaypoH/kaggle_cassava/blob/main/efficientnet_b3_fold_3.ipynb) | \n| effnet-b3 | 0.8849 | 0.898 | 0.890 | --- |\n\n\nAs you can see on the picture below there were 3 submissions with 0.899 Private score, which might place me into a silver zone, but I did not choose any of them, which make a lot of sense - with such a low Cross-Validation scores I never even considered any of them as one of my final submissions. \n\n\n\nThe ones I have chosen had better CV with higher Public LB.\nBut as it was [already said](https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/202673) plenty of times - testing dataset contains a considerable amount of noise, just like training set. A challenge was to make a model robust enough to have good performance with such noise. And I have failed this challenge.\n\nAnyway, besided combining 2020 and 2019 datasets with a dataset of my own I had no tricks. This is what worked for me from the very beginning:\n* Smaller models -> better result: resnet34 > resnet101, effnet-b3 > effnet-b7\n* Lower TTA: No tta was better than 8-10 TTA rounds.\n* Less augmentations: heavy complex augmentations were losing to simple ones. CutMix and MixUp were even worse.\n\nWith everything that I have described I managed to score 0.902 on Public LB during the first week (or two?) of the competition and I remember that I was pretty surprised that this placed me into a gold zone at a time.", "1210047": "Yeah, holidays and other stuff do get in the way.. However do note, family is more important than anything. So don't regret anything! You still did really good, and after reading your discussion, I would love to make more late submissions trying these tactics out. Thanks for solution, family first!", "1210053": "Good job and well done! Next time you can get a medal. :)" }, "source": "meta" }