There was a kernel that scores 0.66+ in public LB yesterday but now it's gone. I just forked the kernel yesterday. So, I'd like to share what I know about it to make this competition fair. I'm not sure the version I forked is the version scores 0.66+ because in my understanding the highest score among all version is shown in the Kernels tab and I hadn't check which one scores 0.66+. But in the version that I forked, the submission file is made using pre-trained weights that are loaded from private datasets. Therefore, you don't have to worry about it so much. If you know something about the kernel please share it here. I hope the author does some explains (sorry I forgot who she/he is).
", "messages": [ { "id": "536788", "postDate": "05/25/2019 09:27:53", "content": "There was a kernel that scores 0.66+ in public LB yesterday but now it's gone. I just forked the kernel yesterday. So, I'd like to share what I know about it to make this competition fair. I'm not sure the version I forked is the version scores 0.66+ because in my understanding the highest score among all version is shown in the Kernels tab and I hadn't check which one scores 0.66+. But in the version that I forked, the submission file is made using pre-trained weights that are loaded from private datasets. Therefore, you don't have to worry about it so much. If you know something about the kernel please share it here. I hope the author does some explains (sorry I forgot who she/he is).
", "rawMarkdown": "There was a kernel that scores 0.66+ in public LB yesterday but now it's gone. I just forked the kernel yesterday. So, I'd like to share what I know about it to make this competition fair. I'm not sure the version I forked is the version scores 0.66+ because in my understanding the highest score among all version is shown in the Kernels tab and I hadn't check which one scores 0.66+. But in the version that I forked, the submission file is made using pre-trained weights that are loaded from private datasets. Therefore, you don't have to worry about it so much. If you know something about the kernel please share it here. I hope the author does some explains (sorry I forgot who she/he is).", "votes": null }, { "id": "536794", "postDate": "05/25/2019 09:46:31", "content": "The latest version was published this morning (GMT +9). It achieved 0.673 LB. \nI didnt save or fork the kernel. However, I already got those information from this kernel:
InceptionV3 logits and aux of InceptionV3 If I remember correctly it seems that the local training was actually just from a previous version of the kernel. Do you remember any details of the training in that kernel?
", "rawMarkdown": "If I remember correctly it seems that the local training was actually just from a previous version of the kernel. Do you remember any details of the training in that kernel?", "votes": null }, { "id": "536833", "postDate": "05/25/2019 11:38:15", "content": "The part that the aux loss is used looks interesting! Could anyone please provide some more details ? @osciiart Could you please share this information if possible?
@tanlikesmath
\nThe training part is not special. It is quite same as the previous kernel
@ratthachat
\nloss = loss_logit + 0.4 * loss_aux. That is all I can remember
well, there is also \"RandomResizedCrop\"\nand I guess croping both dimensions allows to better capture specific frequencies
", "rawMarkdown": "well, there is also \"RandomResizedCrop\"\nand I guess croping both dimensions allows to better capture specific frequencies", "votes": null }, { "id": "537132", "postDate": "05/26/2019 09:05:55", "content": "Sorry for the confusion.\nHi, @osciiart it was my kernel. @backaggle all important tips are listed.\nI made that kernel with unpretrained resnet18 and LB 0.634, that's the initial commit version. Because my local lwlrap was always low, I wished guys can check and discuss my code.\nYesterday, when I changed some train setup online with one commit to get a new model, used the output model generate a submission file. The public LB is 0.66+.\nThat LB was in the bronze zone, and only 2 weeks left for this task. I thought that may be unfair. I donot know how I can delete that high LB commit, so I made the kernel private again after several hours.
\n\nI'm not quite familiar with the rule. If it is not unfair to put a high LB commit with only 2 weeks left to deadline, I will make it public then.
", "rawMarkdown": "Sorry for the confusion.\nHi, @osciiart it was my kernel. @backaggle all important tips are listed.\nI made that kernel with unpretrained resnet18 and LB 0.634, that's the initial commit version. Because my local lwlrap was always low, I wished guys can check and discuss my code.\nYesterday, when I changed some train setup online with one commit to get a new model, used the output model generate a submission file. The public LB is 0.66+.\nThat LB was in the bronze zone, and only 2 weeks left for this task. I thought that may be unfair. I donot know how I can delete that high LB commit, so I made the kernel private again after several hours.\n\nI'm not quite familiar with the rule. If it is not unfair to put a high LB commit with only 2 weeks left to deadline, I will make it public then.", "votes": null }, { "id": "537135", "postDate": "05/26/2019 09:08:08", "content": "Yes, @vandalko I am confused about that. RandomResizedCrop can not explained for sound physically. But it gave a higher LB than no crop.
", "rawMarkdown": "Yes, @vandalko I am confused about that. RandomResizedCrop can not explained for sound physically. But it gave a higher LB than no crop.", "votes": null }, { "id": "537186", "postDate": "05/26/2019 11:41:01", "content": "@sailorwei Thank you for your explanation. I don't mean to blame you. I just made this topic because I found some competitors at twitter who were confused about the kernel had gone before they read in detail. By your explanation, what's happening became clear for me. I think it's onto you to publish the kernel or not. Coming a strong kernel in the last few weeks is a usual event at kaggle so that it's not against the rule, I think. But some may think it's unfair. It's no problem to publish the kernel for me, but of course, you don't have to consider my opinion. You know, we are the team that will not be affected by your kernel most ;)
", "rawMarkdown": "sailorwei Thank you for your explanation. I don't mean to blame you. I just made this topic because I found some competitors at twitter who were confused about the kernel had gone before they read in detail. By your explanation, what's happening became clear for me. I think it's onto you to publish the kernel or not. Coming a strong kernel in the last few weeks is a usual event at kaggle so that it's not against the rule, I think. But some may think it's unfair. It's no problem to publish the kernel for me, but of course, you don't have to consider my opinion. You know, we are the team that will not be affected by your kernel most ;)", "votes": null }, { "id": "537227", "postDate": "05/26/2019 14:02:19", "content": "If single fold is in the bronze zone, k-folds can be in the silver zone then.
\n\nAnyway, I am not affected too, since I am not even in the bronze zone :)
\n\n*EDIT : but there is an official warning about releasing high-score kernel during last week here : \nhttps://www.kaggle.com/spoiler-alert *
", "rawMarkdown": "If single fold is in the bronze zone, k-folds can be in the silver zone then. \n\nAnyway, I am not affected too, since I am not even in the bronze zone :)\n\n**EDIT : but there is an official warning about releasing high-score kernel during last week here : \nhttps://www.kaggle.com/spoiler-alert **", "votes": null }, { "id": "538008", "postDate": "05/28/2019 01:44:09", "content": "As the above link shows, it may not proper to make the kernel public now. I can do it once the competition finished. Yes, your team cannot be affected as all @osciiart . In addition, if you want to share some tips, what are the key point for your high score? model ensemble, ssl, well-designed nn?
", "rawMarkdown": "As the above link shows, it may not proper to make the kernel public now. I can do it once the competition finished. Yes, your team cannot be affected as all @osciiart . In addition, if you want to share some tips, what are the key point for your high score? model ensemble, ssl, well-designed nn?", "votes": null }, { "id": "538024", "postDate": "05/28/2019 02:39:02", "content": "For me, the key points of this competition are preprocessing, augmentation and how to use noisy data. I use a popular ensemble technique. I tried some ssl but failed at all. I'm not stacking on model architecture search.
", "rawMarkdown": "For me, the key points of this competition are preprocessing, augmentation and how to use noisy data. I use a popular ensemble technique. I tried some ssl but failed at all. I'm not stacking on model architecture search.", "votes": null }, { "id": "550319", "postDate": "06/11/2019 13:55:26", "content": "Make it public again while waiting for the 14 day shakeup.\nhttps://www.kaggle.com/sailorwei/fat2019-2d-cnn-with-mixup-lb-0-673
", "rawMarkdown": "Make it public again while waiting for the 14 day shakeup.\nhttps://www.kaggle.com/sailorwei/fat2019-2d-cnn-with-mixup-lb-0-673", "votes": null }, { "id": "551119", "postDate": "06/12/2019 11:12:42", "content": "Thank you for sharing!
", "rawMarkdown": "Thank you for sharing!", "votes": null }, { "id": "551677", "postDate": "06/13/2019 02:19:55", "content": "Yeah, it should be public, I wished guys can help me check the problem. \nLooking forward to your solution.
", "rawMarkdown": "Yeah, it should be public, I wished guys can help me check the problem. \nLooking forward to your solution.", "votes": null } ], "comments": [ { "id": 536794, "author_name": "backaggle", "author_url": "", "post_date": "05/25/2019 09:46:31", "content": "The latest version was published this morning (GMT +9). It achieved 0.673 LB. \nI didnt save or fork the kernel. However, I already got those information from this kernel:
InceptionV3 logits and aux of InceptionV3 If I remember correctly it seems that the local training was actually just from a previous version of the kernel. Do you remember any details of the training in that kernel?
", "votes": null, "replies": [] }, { "id": 536833, "author_name": "ratthachat", "author_url": "", "post_date": "05/25/2019 11:38:15", "content": "The part that the aux loss is used looks interesting! Could anyone please provide some more details ? @osciiart Could you please share this information if possible?
@tanlikesmath
\nThe training part is not special. It is quite same as the previous kernel
@ratthachat
\nloss = loss_logit + 0.4 * loss_aux. That is all I can remember
well, there is also \"RandomResizedCrop\"\nand I guess croping both dimensions allows to better capture specific frequencies
", "votes": null, "replies": [] }, { "id": 537135, "author_name": "sailorwei", "author_url": "", "post_date": "05/26/2019 09:08:08", "content": "Yes, @vandalko I am confused about that. RandomResizedCrop can not explained for sound physically. But it gave a higher LB than no crop.
", "votes": null, "replies": [] } ] }, { "id": 537132, "author_name": "sailorwei", "author_url": "", "post_date": "05/26/2019 09:05:55", "content": "Sorry for the confusion.\nHi, @osciiart it was my kernel. @backaggle all important tips are listed.\nI made that kernel with unpretrained resnet18 and LB 0.634, that's the initial commit version. Because my local lwlrap was always low, I wished guys can check and discuss my code.\nYesterday, when I changed some train setup online with one commit to get a new model, used the output model generate a submission file. The public LB is 0.66+.\nThat LB was in the bronze zone, and only 2 weeks left for this task. I thought that may be unfair. I donot know how I can delete that high LB commit, so I made the kernel private again after several hours.
\n\nI'm not quite familiar with the rule. If it is not unfair to put a high LB commit with only 2 weeks left to deadline, I will make it public then.
", "votes": null, "replies": [ { "id": 537186, "author_name": "osciiart", "author_url": "", "post_date": "05/26/2019 11:41:01", "content": "@sailorwei Thank you for your explanation. I don't mean to blame you. I just made this topic because I found some competitors at twitter who were confused about the kernel had gone before they read in detail. By your explanation, what's happening became clear for me. I think it's onto you to publish the kernel or not. Coming a strong kernel in the last few weeks is a usual event at kaggle so that it's not against the rule, I think. But some may think it's unfair. It's no problem to publish the kernel for me, but of course, you don't have to consider my opinion. You know, we are the team that will not be affected by your kernel most ;)
", "votes": null, "replies": [] }, { "id": 537227, "author_name": "ratthachat", "author_url": "", "post_date": "05/26/2019 14:02:19", "content": "If single fold is in the bronze zone, k-folds can be in the silver zone then.
\n\nAnyway, I am not affected too, since I am not even in the bronze zone :)
\n\n*EDIT : but there is an official warning about releasing high-score kernel during last week here : \nhttps://www.kaggle.com/spoiler-alert *
", "votes": null, "replies": [] }, { "id": 538008, "author_name": "sailorwei", "author_url": "", "post_date": "05/28/2019 01:44:09", "content": "As the above link shows, it may not proper to make the kernel public now. I can do it once the competition finished. Yes, your team cannot be affected as all @osciiart . In addition, if you want to share some tips, what are the key point for your high score? model ensemble, ssl, well-designed nn?
", "votes": null, "replies": [] }, { "id": 538024, "author_name": "osciiart", "author_url": "", "post_date": "05/28/2019 02:39:02", "content": "For me, the key points of this competition are preprocessing, augmentation and how to use noisy data. I use a popular ensemble technique. I tried some ssl but failed at all. I'm not stacking on model architecture search.
", "votes": null, "replies": [] } ] }, { "id": 550319, "author_name": "sailorwei", "author_url": "", "post_date": "06/11/2019 13:55:26", "content": "Make it public again while waiting for the 14 day shakeup.\nhttps://www.kaggle.com/sailorwei/fat2019-2d-cnn-with-mixup-lb-0-673
", "votes": null, "replies": [ { "id": 551119, "author_name": "osciiart", "author_url": "", "post_date": "06/12/2019 11:12:42", "content": "Thank you for sharing!
", "votes": null, "replies": [] }, { "id": 551677, "author_name": "sailorwei", "author_url": "", "post_date": "06/13/2019 02:19:55", "content": "Yeah, it should be public, I wished guys can help me check the problem. \nLooking forward to your solution.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "536788": "There was a kernel that scores 0.66+ in public LB yesterday but now it's gone. I just forked the kernel yesterday. So, I'd like to share what I know about it to make this competition fair. I'm not sure the version I forked is the version scores 0.66+ because in my understanding the highest score among all version is shown in the Kernels tab and I hadn't check which one scores 0.66+. But in the version that I forked, the submission file is made using pre-trained weights that are loaded from private datasets. Therefore, you don't have to worry about it so much. If you know something about the kernel please share it here. I hope the author does some explains (sorry I forgot who she/he is).", "536794": "The latest version was published this morning (GMT +9). It achieved `0.673 LB`. \nI didnt save or fork the kernel. However, I already got those information from this kernel: \n\n- The author loaded the best weight that he/she already trained in the local \n- The architecture is `InceptionV3` \n- The code shows that he/she computes two loss from `logits` and `aux` of `InceptionV3` \n- There are some special preprocessing methods. But, The author still uses data published in [here](https://www.kaggle.com/daisukelab/fat2019_prep_mels1). It confuses me. If the author uses different preprocessing methods, whey does model still work with this kind of data? \n- 25TTA is used.", "536798": "If I remember correctly it seems that the local training was actually just from a previous version of the kernel. Do you remember any details of the training in that kernel?", "536833": "The part that the `aux` loss is used looks interesting! Could anyone please provide some more details ? @osciiart Could you please share this information if possible?", "536839": "tanlikesmath \nThe training part is not special. It is quite same as the previous kernel \n\n@ratthachat \n`loss = loss_logit + 0.4 * loss_aux`. That is all I can remember", "536981": "well, there is also \"RandomResizedCrop\"\nand I guess croping both dimensions allows to better capture specific frequencies", "537132": "Sorry for the confusion.\nHi, @osciiart it was my kernel. @backaggle all important tips are listed.\nI made that kernel with unpretrained resnet18 and LB 0.634, that's the initial commit version. Because my local lwlrap was always low, I wished guys can check and discuss my code.\nYesterday, when I changed some train setup online with one commit to get a new model, used the output model generate a submission file. The public LB is 0.66+.\nThat LB was in the bronze zone, and only 2 weeks left for this task. I thought that may be unfair. I donot know how I can delete that high LB commit, so I made the kernel private again after several hours.\n\nI'm not quite familiar with the rule. If it is not unfair to put a high LB commit with only 2 weeks left to deadline, I will make it public then.", "537135": "Yes, @vandalko I am confused about that. RandomResizedCrop can not explained for sound physically. But it gave a higher LB than no crop.", "537186": "sailorwei Thank you for your explanation. I don't mean to blame you. I just made this topic because I found some competitors at twitter who were confused about the kernel had gone before they read in detail. By your explanation, what's happening became clear for me. I think it's onto you to publish the kernel or not. Coming a strong kernel in the last few weeks is a usual event at kaggle so that it's not against the rule, I think. But some may think it's unfair. It's no problem to publish the kernel for me, but of course, you don't have to consider my opinion. You know, we are the team that will not be affected by your kernel most ;)", "537227": "If single fold is in the bronze zone, k-folds can be in the silver zone then. \n\nAnyway, I am not affected too, since I am not even in the bronze zone :)\n\n**EDIT : but there is an official warning about releasing high-score kernel during last week here : \nhttps://www.kaggle.com/spoiler-alert **", "538008": "As the above link shows, it may not proper to make the kernel public now. I can do it once the competition finished. Yes, your team cannot be affected as all @osciiart . In addition, if you want to share some tips, what are the key point for your high score? model ensemble, ssl, well-designed nn?", "538024": "For me, the key points of this competition are preprocessing, augmentation and how to use noisy data. I use a popular ensemble technique. I tried some ssl but failed at all. I'm not stacking on model architecture search.", "550319": "Make it public again while waiting for the 14 day shakeup.\nhttps://www.kaggle.com/sailorwei/fat2019-2d-cnn-with-mixup-lb-0-673", "551119": "Thank you for sharing!", "551677": "Yeah, it should be public, I wished guys can help me check the problem. \nLooking forward to your solution." }, "source": "meta" }