I am a beginner in programming. I started learning python in December last year. When I saw the results of the competition, I was very disappointed. I shake down from 102 to 244 and couldn't get any medals. After thinking for a long time, I think for beginners, what has been learned is the key point. The ranking of the competition is not the most important. When I found out, I was surprised that I had lost a lot of things.
\nIn the course of the competition, I over-pursued the LB score that may be overfitting. I overlooked many things, such as annotated training and doing local cross-validation. And when I learned of a new method. If I may not get better grades, or feel that I cannot learn well, I will give up learning new methods. I think the annotated training is more interesting than other CNN competitions, but I think it’s too much trouble and I don’t think it will increase much. Also, I felt that I couldn't implement this method well, so I gave up. So I lost learning how to use annotated images for training.
\nBecause I don't know how to use pytorch, I haven't trained the resnet200d model myself from start to finish. This made me unable to obtain my CV scores and was unable to cross-validate. And can not use the start point @ammarali32 provided. Although I know that if I use the start point @ammarali32 provided, I can get good results. But I felt that I only learned python for three months and couldn't learn to use pytorch in a short time, I gave up learning to use pytorch. I believe that if I work hard to learn, I will definitely learn to use pytorch. This should be my biggest regret in this game.
\nAlthough I am the loser of this competition . Before @ammarali32 provided a start point, I was 17th in this competition for a while. The methods I use are all models trained with tensorflow in ensemble using resnet200d of the public notebook. My method can increase the scores of public notebooks by about 0.001~0.002. But the increase in the private data set is not as much as the public, I think it is because I did not do cross-validation.
\nWhat I mainly do is to use the softmax function in this competition, so that my model is not highly correlated with the model output by the sigmoid function. This can slightly improve performance in ensemble.
\nThe ETT is relatively simple, just add a column ETT000.
\nIn the NGT, only a small part of the images will have 2 different labels at the same time, so the image with 2 labels is removed and the NGT000 column is added. Only ensemble the output with the highest ranking for each NGT image.
\nFinally, in CVC, I put all combinations with a label, so I will get 8 columns.
\nCVC001 CVC010 CVC100 CVC101 CVC011 CVC110 CVC000 CVC111
submission['CVC - Abnormal'] = sub_df['CVC100'] +sub_df['CVC101'] +sub_df['CVC111'] +sub_df['CVC110']\nsubmission['CVC - Borderline'] = sub_df['CVC110'] + sub_df['CVC010'] + sub_df['CVC111'] + sub_df['CVC011']\nsubmission['CVC - Normal'] = sub_df['CVC101'] +sub_df['CVC001'] +sub_df['CVC111'] +sub_df['CVC011'] \nIn this way, ETT NGT CVC can be output with softmax function.
\nFinally thanks @xhlulu. The notebook @xhlulu shared made me quickly familiar with the architecture of tensorflow and how to use TPU. Thanks also to everyone who answered my questions during this competition.
", "messages": [ { "id": "1242409", "postDate": "03/17/2021 15:41:21", "content": "I am a beginner in programming. I started learning python in December last year. When I saw the results of the competition, I was very disappointed. I shake down from 102 to 244 and couldn't get any medals. After thinking for a long time, I think for beginners, what has been learned is the key point. The ranking of the competition is not the most important. When I found out, I was surprised that I had lost a lot of things.
\nIn the course of the competition, I over-pursued the LB score that may be overfitting. I overlooked many things, such as annotated training and doing local cross-validation. And when I learned of a new method. If I may not get better grades, or feel that I cannot learn well, I will give up learning new methods. I think the annotated training is more interesting than other CNN competitions, but I think it’s too much trouble and I don’t think it will increase much. Also, I felt that I couldn't implement this method well, so I gave up. So I lost learning how to use annotated images for training.
\nBecause I don't know how to use pytorch, I haven't trained the resnet200d model myself from start to finish. This made me unable to obtain my CV scores and was unable to cross-validate. And can not use the start point @ammarali32 provided. Although I know that if I use the start point @ammarali32 provided, I can get good results. But I felt that I only learned python for three months and couldn't learn to use pytorch in a short time, I gave up learning to use pytorch. I believe that if I work hard to learn, I will definitely learn to use pytorch. This should be my biggest regret in this game.
\nAlthough I am the loser of this competition . Before @ammarali32 provided a start point, I was 17th in this competition for a while. The methods I use are all models trained with tensorflow in ensemble using resnet200d of the public notebook. My method can increase the scores of public notebooks by about 0.001~0.002. But the increase in the private data set is not as much as the public, I think it is because I did not do cross-validation.
\nWhat I mainly do is to use the softmax function in this competition, so that my model is not highly correlated with the model output by the sigmoid function. This can slightly improve performance in ensemble.
\nThe ETT is relatively simple, just add a column ETT000.
\nIn the NGT, only a small part of the images will have 2 different labels at the same time, so the image with 2 labels is removed and the NGT000 column is added. Only ensemble the output with the highest ranking for each NGT image.
\nFinally, in CVC, I put all combinations with a label, so I will get 8 columns.
\nCVC001 CVC010 CVC100 CVC101 CVC011 CVC110 CVC000 CVC111
submission['CVC - Abnormal'] = sub_df['CVC100'] +sub_df['CVC101'] +sub_df['CVC111'] +sub_df['CVC110']\nsubmission['CVC - Borderline'] = sub_df['CVC110'] + sub_df['CVC010'] + sub_df['CVC111'] + sub_df['CVC011']\nsubmission['CVC - Normal'] = sub_df['CVC101'] +sub_df['CVC001'] +sub_df['CVC111'] +sub_df['CVC011'] \nIn this way, ETT NGT CVC can be output with softmax function.
\nFinally thanks @xhlulu. The notebook @xhlulu shared made me quickly familiar with the architecture of tensorflow and how to use TPU. Thanks also to everyone who answered my questions during this competition.
", "rawMarkdown": "I am a beginner in programming. I started learning python in December last year. When I saw the results of the competition, I was very disappointed. I shake down from 102 to 244 and couldn't get any medals. After thinking for a long time, I think for beginners, what has been learned is the key point. The ranking of the competition is not the most important. When I found out, I was surprised that I had lost a lot of things.\n\nIn the course of the competition, I over-pursued the LB score that may be overfitting. I overlooked many things, such as annotated training and doing local cross-validation. And when I learned of a new method. If I may not get better grades, or feel that I cannot learn well, I will give up learning new methods. I think the annotated training is more interesting than other CNN competitions, but I think it’s too much trouble and I don’t think it will increase much. Also, I felt that I couldn't implement this method well, so I gave up. So I lost learning how to use annotated images for training.\n\nBecause I don't know how to use pytorch, I haven't trained the resnet200d model myself from start to finish. This made me unable to obtain my CV scores and was unable to cross-validate. And can not use the start point @ammarali32 provided. Although I know that if I use the start point @ammarali32 provided, I can get good results. But I felt that I only learned python for three months and couldn't learn to use pytorch in a short time, I gave up learning to use pytorch. I believe that if I work hard to learn, I will definitely learn to use pytorch. This should be my biggest regret in this game.\n\nAlthough I am the loser of this competition . Before @ammarali32 provided a start point, I was 17th in this competition for a while. The methods I use are all models trained with tensorflow in ensemble using resnet200d of the public notebook. My method can increase the scores of public notebooks by about 0.001~0.002. But the increase in the private data set is not as much as the public, I think it is because I did not do cross-validation.\n\nWhat I mainly do is to use the softmax function in this competition, so that my model is not highly correlated with the model output by the sigmoid function. This can slightly improve performance in ensemble.\n\nThe ETT is relatively simple, just add a column ETT000.\nIn the NGT, only a small part of the images will have 2 different labels at the same time, so the image with 2 labels is removed and the NGT000 column is added. Only ensemble the output with the highest ranking for each NGT image.\nFinally, in CVC, I put all combinations with a label, so I will get 8 columns.\nCVC001 CVC010 CVC100 CVC101 CVC011 CVC110 CVC000 CVC111\n```\nsubmission['CVC - Abnormal'] = sub_df['CVC100'] +sub_df['CVC101'] +sub_df['CVC111'] +sub_df['CVC110']\nsubmission['CVC - Borderline'] = sub_df['CVC110'] + sub_df['CVC010'] + sub_df['CVC111'] + sub_df['CVC011']\nsubmission['CVC - Normal'] = sub_df['CVC101'] +sub_df['CVC001'] +sub_df['CVC111'] +sub_df['CVC011'] \n```\nIn this way, ETT NGT CVC can be output with softmax function.\n\nFinally thanks @xhlulu. The notebook @xhlulu shared made me quickly familiar with the architecture of tensorflow and how to use TPU. Thanks also to everyone who answered my questions during this competition.", "votes": null }, { "id": "1242485", "postDate": "03/17/2021 16:34:13", "content": "I have been learning for an year!
", "rawMarkdown": "I have been learning for an year!", "votes": null }, { "id": "1243279", "postDate": "03/18/2021 05:55:18", "content": "@h053473666 Thanks for the writeup Alien about your experience . Dont worry about Shakeup . You will win Medal in VinBig Data
", "rawMarkdown": "h053473666 Thanks for the writeup Alien about your experience . Dont worry about Shakeup . You will win Medal in VinBig Data", "votes": null }, { "id": "1243314", "postDate": "03/18/2021 06:19:23", "content": "Thanks for your encouragement.
", "rawMarkdown": "Thanks for your encouragement.", "votes": null } ], "comments": [ { "id": 1242485, "author_name": "morizin", "author_url": "", "post_date": "03/17/2021 16:34:13", "content": "I have been learning for an year!
", "votes": null, "replies": [] }, { "id": 1243279, "author_name": "usharengaraju", "author_url": "", "post_date": "03/18/2021 05:55:18", "content": "@h053473666 Thanks for the writeup Alien about your experience . Dont worry about Shakeup . You will win Medal in VinBig Data
", "votes": null, "replies": [ { "id": 1243314, "author_name": "h053473666", "author_url": "", "post_date": "03/18/2021 06:19:23", "content": "Thanks for your encouragement.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1242409": "I am a beginner in programming. I started learning python in December last year. When I saw the results of the competition, I was very disappointed. I shake down from 102 to 244 and couldn't get any medals. After thinking for a long time, I think for beginners, what has been learned is the key point. The ranking of the competition is not the most important. When I found out, I was surprised that I had lost a lot of things.\n\nIn the course of the competition, I over-pursued the LB score that may be overfitting. I overlooked many things, such as annotated training and doing local cross-validation. And when I learned of a new method. If I may not get better grades, or feel that I cannot learn well, I will give up learning new methods. I think the annotated training is more interesting than other CNN competitions, but I think it’s too much trouble and I don’t think it will increase much. Also, I felt that I couldn't implement this method well, so I gave up. So I lost learning how to use annotated images for training.\n\nBecause I don't know how to use pytorch, I haven't trained the resnet200d model myself from start to finish. This made me unable to obtain my CV scores and was unable to cross-validate. And can not use the start point @ammarali32 provided. Although I know that if I use the start point @ammarali32 provided, I can get good results. But I felt that I only learned python for three months and couldn't learn to use pytorch in a short time, I gave up learning to use pytorch. I believe that if I work hard to learn, I will definitely learn to use pytorch. This should be my biggest regret in this game.\n\nAlthough I am the loser of this competition . Before @ammarali32 provided a start point, I was 17th in this competition for a while. The methods I use are all models trained with tensorflow in ensemble using resnet200d of the public notebook. My method can increase the scores of public notebooks by about 0.001~0.002. But the increase in the private data set is not as much as the public, I think it is because I did not do cross-validation.\n\nWhat I mainly do is to use the softmax function in this competition, so that my model is not highly correlated with the model output by the sigmoid function. This can slightly improve performance in ensemble.\n\nThe ETT is relatively simple, just add a column ETT000.\nIn the NGT, only a small part of the images will have 2 different labels at the same time, so the image with 2 labels is removed and the NGT000 column is added. Only ensemble the output with the highest ranking for each NGT image.\nFinally, in CVC, I put all combinations with a label, so I will get 8 columns.\nCVC001 CVC010 CVC100 CVC101 CVC011 CVC110 CVC000 CVC111\n```\nsubmission['CVC - Abnormal'] = sub_df['CVC100'] +sub_df['CVC101'] +sub_df['CVC111'] +sub_df['CVC110']\nsubmission['CVC - Borderline'] = sub_df['CVC110'] + sub_df['CVC010'] + sub_df['CVC111'] + sub_df['CVC011']\nsubmission['CVC - Normal'] = sub_df['CVC101'] +sub_df['CVC001'] +sub_df['CVC111'] +sub_df['CVC011'] \n```\nIn this way, ETT NGT CVC can be output with softmax function.\n\nFinally thanks @xhlulu. The notebook @xhlulu shared made me quickly familiar with the architecture of tensorflow and how to use TPU. Thanks also to everyone who answered my questions during this competition.", "1242485": "I have been learning for an year!", "1243279": "h053473666 Thanks for the writeup Alien about your experience . Dont worry about Shakeup . You will win Medal in VinBig Data", "1243314": "Thanks for your encouragement." }, "source": "meta" }