It would be good to get the commentary from the winning solutions and also from the top Solutions. \nI did not concentrate much on this. However the competition I had concentrated was iCassava Challenge where also I received the 29th place among 87 teams. My solution was based on Densenet with fastai library.
\n\nThanks again!
", "messages": [ { "id": "541340", "postDate": "06/02/2019 09:03:34", "content": "It would be good to get the commentary from the winning solutions and also from the top Solutions. \nI did not concentrate much on this. However the competition I had concentrated was iCassava Challenge where also I received the 29th place among 87 teams. My solution was based on Densenet with fastai library.
\n\nThanks again!
", "rawMarkdown": "It would be good to get the commentary from the winning solutions and also from the top Solutions. \nI did not concentrate much on this. However the competition I had concentrated was iCassava Challenge where also I received the 29th place among 87 teams. My solution was based on Densenet with fastai library.\n\nThanks again!", "votes": null }, { "id": "544256", "postDate": "06/05/2019 10:13:49", "content": "I am very interested in the solution from @lzhbrian who clearly did something different as his/her error is ~4x smaller compared to the 2nd place. I have no idea how someone could achieve such a high score.
\n\nI finished 4th and very close to 2nd and 3rd place so I assume they did something similar to what I did which is a standard Kfold averaging (seresnext101). I did it on unprocessed images which scores 0.99+ on public. Errors are mainly images with only upper body present so I did Kfold with cropped upper body and scored 0.96+ and after averaging it reached 0.993+. Here the problematic images were those with a lot of pixels removed by organizers in their cropping pipeline. Perhaps @lzhbrian also used version1 of the dataset which had all the pixels present as this could certainly help and he/she had access to it unlike many who joined late. This is only my speculation motivated by 3 days of silence from the winning teams.
", "rawMarkdown": "I am very interested in the solution from [@lzhbrian](https://www.kaggle.com/lzhbrian) who clearly did something different as his/her error is ~4x smaller compared to the 2nd place. I have no idea how someone could achieve such a high score. \n\nI finished 4th and very close to 2nd and 3rd place so I assume they did something similar to what I did which is a standard Kfold averaging (seresnext101). I did it on unprocessed images which scores 0.99+ on public. Errors are mainly images with only upper body present so I did Kfold with cropped upper body and scored 0.96+ and after averaging it reached 0.993+. Here the problematic images were those with a lot of pixels removed by organizers in their cropping pipeline. Perhaps [@lzhbrian](https://www.kaggle.com/lzhbrian) also used version1 of the dataset which had all the pixels present as this could certainly help and he/she had access to it unlike many who joined late. This is only my speculation motivated by 3 days of silence from the winning teams.", "votes": null }, { "id": "544470", "postDate": "06/05/2019 14:41:01", "content": "Unfortunately I joined the competition late and didn't have access to the version 1 of the test set.\nI'm curious how my performance would be if I had access to test set version. \nMaybe someone can share it?
\n\nI wanted to use stratified K-fold averaging with 8 folds but only had time to train 2 of the folds for the following models:\n- pnasnet5large\n- senet154\n- se-resnext101-32x4d\n- se-resnet152
\n\nI used the Fast.ai framework and used the following 4 stage training:
\n\n```\nif STAGE == 1:\n EPOCHS = 40\n SIZE = 'S'\n TRANSFORM = False\n LR = 1e-2\n REDUCE_FACTOR = 0.9\n LAYERS = 'FREEZE'\n MIXUP = False
\n\nif STAGE == 2:\n EPOCHS = 24 # 12 or 24 (resnet)\n SIZE = 'L'\n TRANSFORM = True\n LR = 5e-3 # 1e-2 a 1e-3\n REDUCE_FACTOR = 0.8\n LAYERS = 'FREEZE'\n MIXUP = False
\n\nif STAGE == 3:\n EPOCHS = 16 #8 or 16 (resnet)\n SIZE = 'L'\n TRANSFORM = True\n LR = 1e-4 ### 1e-4 a 1e-5\n REDUCE_FACTOR = 0.8\n LAYERS = 'UNFREEZE'\n MIXUP = False
\n\nif STAGE == 4:\n EPOCHS = 16 #8 or 16 (resnet)\n SIZE = 'L'\n TRANSFORM = True\n LR = 1e-5 ### 1e-4 a 1e-5\n REDUCE_FACTOR = 0.8\n LAYERS = 'UNFREEZE'\n MIXUP = True
\n\nx,y = 226,80\nif SIZE == 'S':\n SIZE = (x,y)\nif SIZE == 'L':\n SIZE = (x*2-2,y*2)\n```
\n\nI came upon this stages by experimenting and reading upon best practices from fast.ai. \nFeel free to comment any suggestions!
", "rawMarkdown": "Unfortunately I joined the competition late and didn't have access to the version 1 of the test set.\nI'm curious how my performance would be if I had access to test set version. \nMaybe someone can share it?\n\nI wanted to use stratified K-fold averaging with 8 folds but only had time to train 2 of the folds for the following models:\n- pnasnet5large\n- senet154\n- se-resnext101-32x4d\n- se-resnet152\n\nI used the Fast.ai framework and used the following 4 stage training:\n\n```\nif STAGE == 1:\n EPOCHS = 40\n SIZE = 'S'\n TRANSFORM = False\n LR = 1e-2\n REDUCE_FACTOR = 0.9\n LAYERS = 'FREEZE'\n MIXUP = False\n \nif STAGE == 2:\n EPOCHS = 24 # 12 or 24 (resnet)\n SIZE = 'L'\n TRANSFORM = True\n LR = 5e-3 # 1e-2 a 1e-3\n REDUCE_FACTOR = 0.8\n LAYERS = 'FREEZE'\n MIXUP = False\n\nif STAGE == 3:\n EPOCHS = 16 #8 or 16 (resnet)\n SIZE = 'L'\n TRANSFORM = True\n LR = 1e-4 ### 1e-4 a 1e-5\n REDUCE_FACTOR = 0.8\n LAYERS = 'UNFREEZE'\n MIXUP = False\n \nif STAGE == 4:\n EPOCHS = 16 #8 or 16 (resnet)\n SIZE = 'L'\n TRANSFORM = True\n LR = 1e-5 ### 1e-4 a 1e-5\n REDUCE_FACTOR = 0.8\n LAYERS = 'UNFREEZE'\n MIXUP = True\n\nx,y = 226,80\nif SIZE == 'S':\n SIZE = (x,y)\nif SIZE == 'L':\n SIZE = (x*2-2,y*2)\n```\n\nI came upon this stages by experimenting and reading upon best practices from fast.ai. \nFeel free to comment any suggestions!", "votes": null }, { "id": "544472", "postDate": "06/05/2019 14:45:18", "content": "Thanks @kurtjanssens. Would it be possible to share the code ?
\n\nThe 4 stage training is very interesting! Would you point to the lecture in fast.ai they are using this ?
\n\nI also used fast.ai
", "rawMarkdown": "Thanks @kurtjanssens. Would it be possible to share the code ?\n\nThe 4 stage training is very interesting! Would you point to the lecture in fast.ai they are using this ?\n\nI also used fast.ai", "votes": null }, { "id": "544741", "postDate": "06/05/2019 20:54:35", "content": "Hi all, sorry for a silence from our side but we are participating in another challenges so we will share our description later on. Thank you for your patience.
\n\nBTW anyone attending CVPR?
", "rawMarkdown": "Hi all, sorry for a silence from our side but we are participating in another challenges so we will share our description later on. Thank you for your patience.\n\nBTW anyone attending CVPR?", "votes": null }, { "id": "546198", "postDate": "06/06/2019 10:50:38", "content": "I did about the same things and I'm also wondering what Izhbrian did to get ~4x smaller error compared to the 2nd place
", "rawMarkdown": "I did about the same things and I'm also wondering what Izhbrian did to get ~4x smaller error compared to the 2nd place", "votes": null }, { "id": "547576", "postDate": "06/07/2019 22:13:04", "content": "Greetings all,
\n\nThank you for the participation in this competition. If you are a top 3rd place winner, please email the Jupyter Notebook containing the approach for your winning solution to us to validate and then issue the prize money. We will confirm the accuracy score for all winning models were based on the final set of training images uploaded (i.e., not the earlier set as some had expressed concern). Our email address is hearstml@gmail.com
\n\nMany thanks,\nHearst ML
", "rawMarkdown": "Greetings all,\n\nThank you for the participation in this competition. If you are a top 3rd place winner, please email the Jupyter Notebook containing the approach for your winning solution to us to validate and then issue the prize money. We will confirm the accuracy score for all winning models were based on the final set of training images uploaded (i.e., not the earlier set as some had expressed concern). Our email address is hearstml@gmail.com\n\nMany thanks,\nHearst ML", "votes": null }, { "id": "551362", "postDate": "06/12/2019 16:16:20", "content": "Hey @picekl , @aleszita and @dsvolkov ,
\n\ncan you please share the code/Jupyter Notebook containing the approach for your winning solution to us to validate and then issue the prize money.?
\n\nFor the other participants:\nWe also want to say thank you all for the other participants and you are also welcome to post a description of your approach on Kaggle!
\n\nOur email address is hearstml@gmail.com
", "rawMarkdown": "**Hey @picekl , @aleszita and @dsvolkov ,**\n\ncan you please share the code/Jupyter Notebook containing the approach for your winning solution to us to validate and then issue the prize money.?\n\n- We will need to confirm the accuracy score for all winning models were based on the final set of training images uploaded (i.e., not the earlier set as some had expressed concern). \n- We will also need to make a presentation at FGVC2019 at CVPR based on your approaches. You can send us 1-2 google slides (or pdf) describing your methods. Or we can edit it on our side.\n- You also have access to a 4 foot x 4 foot poster board if you want to present your method at the workshop (we cannot do remote presentations or videos)\n- If you are planning to join the FGVC workshop next Monday, please let us know. Maybe we could meet and discuss some potential collaboration opportunities in the future :)\n\nThank you! \n---------------------------------------------------------------------------------------------------------------------------------------\n**For the other participants:**\nWe also want to say thank you all for the other participants and you are also welcome to post a description of your approach on Kaggle!\n\nOur email address is hearstml@gmail.com", "votes": null }, { "id": "551534", "postDate": "06/12/2019 20:32:55", "content": "Hello Hearst ML! Did not notice your first message on the forum. I'm going to send a solution in a couple of days. Thank you
", "rawMarkdown": "Hello Hearst ML! Did not notice your first message on the forum. I'm going to send a solution in a couple of days. Thank you", "votes": null }, { "id": "551567", "postDate": "06/12/2019 21:34:45", "content": "Hey @picekl , We will come to CVPR . Will you come to FGVC workshop next Monday? Maybe we can talk a little bit if you will come. Can you please also share your solution since we will need to present your solution at FGVC workshop at CVPR (please see my another message about prize below). Thanks!
", "rawMarkdown": "Hey @picekl , We will come to CVPR . Will you come to FGVC workshop next Monday? Maybe we can talk a little bit if you will come. Can you please also share your solution since we will need to present your solution at FGVC workshop at CVPR (please see my another message about prize below). Thanks!", "votes": null }, { "id": "551568", "postDate": "06/12/2019 21:36:47", "content": "Hey @dsvolkov , thank you for the message! If you could send us your solution during the weekend, it would be great. Just because we need to present your solution at FGVC workshop at CVPR. BTW, are you gonna be at FGVC workshop next Monday? Thank you!
", "rawMarkdown": "Hey @dsvolkov , thank you for the message! If you could send us your solution during the weekend, it would be great. Just because we need to present your solution at FGVC workshop at CVPR. BTW, are you gonna be at FGVC workshop next Monday? Thank you!", "votes": null }, { "id": "552012", "postDate": "06/13/2019 12:09:37", "content": "I am still very interested in the first place solution. Any news with that regard?
", "rawMarkdown": "I am still very interested in the first place solution. Any news with that regard?", "votes": null }, { "id": "552817", "postDate": "06/14/2019 15:31:35", "content": "Hey @picekl, @aleszita and @dsvolkov \nsorry to bother you again on this: If you could send us your solution during the weekend, it would be great. Just because we need to present your solution at FGVC workshop at CVPR next Monday (6/17). BTW, are you gonna be at FGVC workshop next Monday? Thank you!
", "rawMarkdown": "Hey @picekl, @aleszita and @dsvolkov \nsorry to bother you again on this: If you could send us your solution during the weekend, it would be great. Just because we need to present your solution at FGVC workshop at CVPR **next Monday (6/17).** BTW, are you gonna be at FGVC workshop next Monday? Thank you!", "votes": null }, { "id": "552818", "postDate": "06/14/2019 15:33:28", "content": "unfortunately, we didn't :(
", "rawMarkdown": "unfortunately, we didn't :(", "votes": null }, { "id": "552822", "postDate": "06/14/2019 15:44:41", "content": "Dear @valanm , @shentao , @garybios , @kurtjanssens , I am wondering if you could send us your solution of iDesigner during the weekend. I am trying to reach the Top 3 team. If I cannot receive their response before next Monday (for FGVC workshop), I hope we can present your solution at FGVC workshop at CVPR if you are interested in. BTW, are you gonna be at FGVC workshop next Monday? Thank you!
", "rawMarkdown": "Dear @valanm , @shentao , @garybios , @kurtjanssens , I am wondering if you could send us your solution of iDesigner during the weekend. I am trying to reach the Top 3 team. If I cannot receive their response before **next Monday (for FGVC workshop),** I hope we can present your solution at FGVC workshop at CVPR if you are interested in. BTW, are you gonna be at FGVC workshop next Monday? Thank you!", "votes": null }, { "id": "552847", "postDate": "06/14/2019 16:45:38", "content": "Hey Hearst ML! \nI already sent a code for the inference and all pretrained models to your mail address. I will definitely send you training pipeline during the weekend.
\n\nUnfortunately, I cannot participate the FGVC workshop, as it is 14 hours of flight from me (and I do not have a visa anyway)
", "rawMarkdown": "Hey Hearst ML! \nI already sent a code for the inference and all pretrained models to your mail address. I will definitely send you training pipeline during the weekend. \n\nUnfortunately, I cannot participate the FGVC workshop, as it is 14 hours of flight from me (and I do not have a visa anyway)", "votes": null }, { "id": "552896", "postDate": "06/14/2019 19:09:48", "content": "Hey @dsvolkov ,\nthanks for your reply!
\n1) we will wait your training pipeline during the weekend.\n2) we will issue the prize after validating your model.\n3) I see. Maybe hope to see you on the FGVC next year :)
Hi I am coming back from vacation and will get back to you tomorrow. Sorry for the inconvenience
", "rawMarkdown": "Hi I am coming back from vacation and will get back to you tomorrow. Sorry for the inconvenience", "votes": null }, { "id": "553296", "postDate": "06/15/2019 12:39:50", "content": "Hello @hearstml, \nI've given you access to my the notebook I used for training the models.\nI've used a 4 stage approach as outlined above.\nThe idea of training is stages is to increase the difficulty of the problem in each stage.\nPhase 2 = Increasing the size\nPhase 3 = Unfreezing\nPhase 4 = Adding mixup\nFeel free to contact me if I can provide some more details!
", "rawMarkdown": "Hello @hearstml, \nI've given you access to my the notebook I used for training the models.\nI've used a 4 stage approach as outlined above.\nThe idea of training is stages is to increase the difficulty of the problem in each stage.\nPhase 2 = Increasing the size\nPhase 3 = Unfreezing\nPhase 4 = Adding mixup\nFeel free to contact me if I can provide some more details!", "votes": null }, { "id": "553305", "postDate": "06/15/2019 12:57:28", "content": "Sure, I can share the notebook! But I can't find your username when trying to share the kernel. I've send you a direct message with the github link.
", "rawMarkdown": "Sure, I can share the notebook! But I can't find your username when trying to share the kernel. I've send you a direct message with the github link.", "votes": null }, { "id": "553323", "postDate": "06/15/2019 13:46:00", "content": "\n\nWould it be possible to share the code and the winning approaches to all of us ?
\n\nCheers
", "rawMarkdown": "hearstml \n\nWould it be possible to share the code and the winning approaches to all of us ?\n\nCheers", "votes": null }, { "id": "553417", "postDate": "06/15/2019 16:40:40", "content": "\n\nThanks. But I could not get from the github link. It is showing 404 error
\n\nMy user name is @ambarish. https://www.kaggle.com/ambarish is my profile page.
\n\nCheers\nBukun ( Ambarish )
", "rawMarkdown": "kurtjanssens \n\nThanks. But I could not get from the github link. It is showing 404 error\n\nMy user name is @ambarish. https://www.kaggle.com/ambarish is my profile page.\n\nCheers\nBukun ( Ambarish )", "votes": null }, { "id": "553635", "postDate": "06/16/2019 03:27:19", "content": "Hey @kurtjanssens , thank you very much! Appreciate that! :)
", "rawMarkdown": "Hey @kurtjanssens , thank you very much! Appreciate that! :)", "votes": null }, { "id": "554291", "postDate": "06/17/2019 08:32:37", "content": "HI, i have shared too
", "rawMarkdown": "HI, i have shared too", "votes": null }, { "id": "554614", "postDate": "06/17/2019 18:40:08", "content": "Hi @valanm -> here you have brief overview of used method. It's similar to my and my college article published last year and used for previous edition of FGVC. If you have any questions please feel free to ask. I will try to answer all of them.
\n\nProposed system is an ensemble of multiple Convolutional Neural Network architectures such as Inception-ResNet-v2, Inception-v4 [1] and PNASNet[2], improved by:\n - Multiple level training where:\n - 1st level - trains only last layer - 10 epochs - rmsprop - Initial LR 0.01 with exponential decay 0.9 per epoch.\n - 2nd level - trains whole net - 10 epochs - rmsprop - Initial LR 0.01 with exponential decay 0.9 per epoch.\n - 3rd same as first - 10 epochs - rmsprop - rmsprop - Initial LR 0.0075 with exponential decay 0.9 per epoch.\n - 4th same as second - 10 epochs - rmsprop - rmsprop - Initial LR 0.0075 with exponential decay 0.9 per epoch.\n - Using random crops to keep image aspect ration as well as get better detail on images with higher resolution.\n - Using running averages on the trained variables with an exponential weight decay.\n - Adjusting predictions according to the estimated change of the class prior probabilities [3,4].\n - Test-time data augmentation with multiple predictions that covers image with an overlap.\n - Summing the soft-max values from multiple crops/nets and combining that with MODUS “decision” on images with low top1 prediction accuracy.
\n\nFor training I used TensorFlow slim.\n\n[1] Szegedy, C., Ioffe, S., Vanhoucke, V.: Inception-v4, Inception-ResNet and the impact of residual connections on learning. arXiv:1602.07261 (2016)\n[2] C. Liu, B. Zoph, J. Shlens, W. Hua, L.-J. Li, L. Fei-Fei, A. Yuille, J. Huang, and K. Murphy. Progressive neural architecture search. ECCV, 2018.\n[3] Sulc, M., Matas, J.: Improving cnn classifiers by estimating test-time priors. arXiv:1805.08235 (2018)\n[4] Saerens, M., Latinne, P., Decaestecker, C.: Adjusting the outputs of a classifier to new a priori probabilities: a simple procedure. Neural computation 14(1), 21–41 (2002)\nI am very interested in the solution from @lzhbrian who clearly did something different as his/her error is ~4x smaller compared to the 2nd place. I have no idea how someone could achieve such a high score.
\n\nI finished 4th and very close to 2nd and 3rd place so I assume they did something similar to what I did which is a standard Kfold averaging (seresnext101). I did it on unprocessed images which scores 0.99+ on public. Errors are mainly images with only upper body present so I did Kfold with cropped upper body and scored 0.96+ and after averaging it reached 0.993+. Here the problematic images were those with a lot of pixels removed by organizers in their cropping pipeline. Perhaps @lzhbrian also used version1 of the dataset which had all the pixels present as this could certainly help and he/she had access to it unlike many who joined late. This is only my speculation motivated by 3 days of silence from the winning teams.
", "votes": null, "replies": [ { "id": 546198, "author_name": "dsvolkov", "author_url": "", "post_date": "06/06/2019 10:50:38", "content": "I did about the same things and I'm also wondering what Izhbrian did to get ~4x smaller error compared to the 2nd place
", "votes": null, "replies": [] }, { "id": 554614, "author_name": "picekl", "author_url": "", "post_date": "06/17/2019 18:40:08", "content": "Hi @valanm -> here you have brief overview of used method. It's similar to my and my college article published last year and used for previous edition of FGVC. If you have any questions please feel free to ask. I will try to answer all of them.
\n\nProposed system is an ensemble of multiple Convolutional Neural Network architectures such as Inception-ResNet-v2, Inception-v4 [1] and PNASNet[2], improved by:\n - Multiple level training where:\n - 1st level - trains only last layer - 10 epochs - rmsprop - Initial LR 0.01 with exponential decay 0.9 per epoch.\n - 2nd level - trains whole net - 10 epochs - rmsprop - Initial LR 0.01 with exponential decay 0.9 per epoch.\n - 3rd same as first - 10 epochs - rmsprop - rmsprop - Initial LR 0.0075 with exponential decay 0.9 per epoch.\n - 4th same as second - 10 epochs - rmsprop - rmsprop - Initial LR 0.0075 with exponential decay 0.9 per epoch.\n - Using random crops to keep image aspect ration as well as get better detail on images with higher resolution.\n - Using running averages on the trained variables with an exponential weight decay.\n - Adjusting predictions according to the estimated change of the class prior probabilities [3,4].\n - Test-time data augmentation with multiple predictions that covers image with an overlap.\n - Summing the soft-max values from multiple crops/nets and combining that with MODUS “decision” on images with low top1 prediction accuracy.
\n\nFor training I used TensorFlow slim.\n\n[1] Szegedy, C., Ioffe, S., Vanhoucke, V.: Inception-v4, Inception-ResNet and the impact of residual connections on learning. arXiv:1602.07261 (2016)\n[2] C. Liu, B. Zoph, J. Shlens, W. Hua, L.-J. Li, L. Fei-Fei, A. Yuille, J. Huang, and K. Murphy. Progressive neural architecture search. ECCV, 2018.\n[3] Sulc, M., Matas, J.: Improving cnn classifiers by estimating test-time priors. arXiv:1805.08235 (2018)\n[4] Saerens, M., Latinne, P., Decaestecker, C.: Adjusting the outputs of a classifier to new a priori probabilities: a simple procedure. Neural computation 14(1), 21–41 (2002)\nUnfortunately I joined the competition late and didn't have access to the version 1 of the test set.\nI'm curious how my performance would be if I had access to test set version. \nMaybe someone can share it?
\n\nI wanted to use stratified K-fold averaging with 8 folds but only had time to train 2 of the folds for the following models:\n- pnasnet5large\n- senet154\n- se-resnext101-32x4d\n- se-resnet152
\n\nI used the Fast.ai framework and used the following 4 stage training:
\n\n```\nif STAGE == 1:\n EPOCHS = 40\n SIZE = 'S'\n TRANSFORM = False\n LR = 1e-2\n REDUCE_FACTOR = 0.9\n LAYERS = 'FREEZE'\n MIXUP = False
\n\nif STAGE == 2:\n EPOCHS = 24 # 12 or 24 (resnet)\n SIZE = 'L'\n TRANSFORM = True\n LR = 5e-3 # 1e-2 a 1e-3\n REDUCE_FACTOR = 0.8\n LAYERS = 'FREEZE'\n MIXUP = False
\n\nif STAGE == 3:\n EPOCHS = 16 #8 or 16 (resnet)\n SIZE = 'L'\n TRANSFORM = True\n LR = 1e-4 ### 1e-4 a 1e-5\n REDUCE_FACTOR = 0.8\n LAYERS = 'UNFREEZE'\n MIXUP = False
\n\nif STAGE == 4:\n EPOCHS = 16 #8 or 16 (resnet)\n SIZE = 'L'\n TRANSFORM = True\n LR = 1e-5 ### 1e-4 a 1e-5\n REDUCE_FACTOR = 0.8\n LAYERS = 'UNFREEZE'\n MIXUP = True
\n\nx,y = 226,80\nif SIZE == 'S':\n SIZE = (x,y)\nif SIZE == 'L':\n SIZE = (x*2-2,y*2)\n```
\n\nI came upon this stages by experimenting and reading upon best practices from fast.ai. \nFeel free to comment any suggestions!
", "votes": null, "replies": [ { "id": 544472, "author_name": "ambarish", "author_url": "", "post_date": "06/05/2019 14:45:18", "content": "Thanks @kurtjanssens. Would it be possible to share the code ?
\n\nThe 4 stage training is very interesting! Would you point to the lecture in fast.ai they are using this ?
\n\nI also used fast.ai
", "votes": null, "replies": [] }, { "id": 553305, "author_name": "kurtjanssens", "author_url": "", "post_date": "06/15/2019 12:57:28", "content": "Sure, I can share the notebook! But I can't find your username when trying to share the kernel. I've send you a direct message with the github link.
", "votes": null, "replies": [] }, { "id": 553417, "author_name": "ambarish", "author_url": "", "post_date": "06/15/2019 16:40:40", "content": "\n\nThanks. But I could not get from the github link. It is showing 404 error
\n\nMy user name is @ambarish. https://www.kaggle.com/ambarish is my profile page.
\n\nCheers\nBukun ( Ambarish )
", "votes": null, "replies": [] } ] }, { "id": 544741, "author_name": "picekl", "author_url": "", "post_date": "06/05/2019 20:54:35", "content": "Hi all, sorry for a silence from our side but we are participating in another challenges so we will share our description later on. Thank you for your patience.
\n\nBTW anyone attending CVPR?
", "votes": null, "replies": [ { "id": 551567, "author_name": "hearstml", "author_url": "", "post_date": "06/12/2019 21:34:45", "content": "Hey @picekl , We will come to CVPR . Will you come to FGVC workshop next Monday? Maybe we can talk a little bit if you will come. Can you please also share your solution since we will need to present your solution at FGVC workshop at CVPR (please see my another message about prize below). Thanks!
", "votes": null, "replies": [] } ] }, { "id": 547576, "author_name": "hearstml", "author_url": "", "post_date": "06/07/2019 22:13:04", "content": "Greetings all,
\n\nThank you for the participation in this competition. If you are a top 3rd place winner, please email the Jupyter Notebook containing the approach for your winning solution to us to validate and then issue the prize money. We will confirm the accuracy score for all winning models were based on the final set of training images uploaded (i.e., not the earlier set as some had expressed concern). Our email address is hearstml@gmail.com
\n\nMany thanks,\nHearst ML
", "votes": null, "replies": [] }, { "id": 551362, "author_name": "hearstml", "author_url": "", "post_date": "06/12/2019 16:16:20", "content": "Hey @picekl , @aleszita and @dsvolkov ,
\n\ncan you please share the code/Jupyter Notebook containing the approach for your winning solution to us to validate and then issue the prize money.?
\n\nFor the other participants:\nWe also want to say thank you all for the other participants and you are also welcome to post a description of your approach on Kaggle!
\n\nOur email address is hearstml@gmail.com
", "votes": null, "replies": [ { "id": 551534, "author_name": "dsvolkov", "author_url": "", "post_date": "06/12/2019 20:32:55", "content": "Hello Hearst ML! Did not notice your first message on the forum. I'm going to send a solution in a couple of days. Thank you
", "votes": null, "replies": [] }, { "id": 551568, "author_name": "hearstml", "author_url": "", "post_date": "06/12/2019 21:36:47", "content": "Hey @dsvolkov , thank you for the message! If you could send us your solution during the weekend, it would be great. Just because we need to present your solution at FGVC workshop at CVPR. BTW, are you gonna be at FGVC workshop next Monday? Thank you!
", "votes": null, "replies": [] }, { "id": 552012, "author_name": "valanm", "author_url": "", "post_date": "06/13/2019 12:09:37", "content": "I am still very interested in the first place solution. Any news with that regard?
", "votes": null, "replies": [] }, { "id": 552818, "author_name": "hearstml", "author_url": "", "post_date": "06/14/2019 15:33:28", "content": "unfortunately, we didn't :(
", "votes": null, "replies": [] }, { "id": 553323, "author_name": "ambarish", "author_url": "", "post_date": "06/15/2019 13:46:00", "content": "\n\nWould it be possible to share the code and the winning approaches to all of us ?
\n\nCheers
", "votes": null, "replies": [] } ] }, { "id": 552817, "author_name": "hearstml", "author_url": "", "post_date": "06/14/2019 15:31:35", "content": "Hey @picekl, @aleszita and @dsvolkov \nsorry to bother you again on this: If you could send us your solution during the weekend, it would be great. Just because we need to present your solution at FGVC workshop at CVPR next Monday (6/17). BTW, are you gonna be at FGVC workshop next Monday? Thank you!
", "votes": null, "replies": [ { "id": 552847, "author_name": "dsvolkov", "author_url": "", "post_date": "06/14/2019 16:45:38", "content": "Hey Hearst ML! \nI already sent a code for the inference and all pretrained models to your mail address. I will definitely send you training pipeline during the weekend.
\n\nUnfortunately, I cannot participate the FGVC workshop, as it is 14 hours of flight from me (and I do not have a visa anyway)
", "votes": null, "replies": [] }, { "id": 552896, "author_name": "hearstml", "author_url": "", "post_date": "06/14/2019 19:09:48", "content": "Hey @dsvolkov ,\nthanks for your reply!
\n1) we will wait your training pipeline during the weekend.\n2) we will issue the prize after validating your model.\n3) I see. Maybe hope to see you on the FGVC next year :)
Dear @valanm , @shentao , @garybios , @kurtjanssens , I am wondering if you could send us your solution of iDesigner during the weekend. I am trying to reach the Top 3 team. If I cannot receive their response before next Monday (for FGVC workshop), I hope we can present your solution at FGVC workshop at CVPR if you are interested in. BTW, are you gonna be at FGVC workshop next Monday? Thank you!
", "votes": null, "replies": [ { "id": 553105, "author_name": "valanm", "author_url": "", "post_date": "06/15/2019 05:47:33", "content": "Hi I am coming back from vacation and will get back to you tomorrow. Sorry for the inconvenience
", "votes": null, "replies": [] }, { "id": 553296, "author_name": "kurtjanssens", "author_url": "", "post_date": "06/15/2019 12:39:50", "content": "Hello @hearstml, \nI've given you access to my the notebook I used for training the models.\nI've used a 4 stage approach as outlined above.\nThe idea of training is stages is to increase the difficulty of the problem in each stage.\nPhase 2 = Increasing the size\nPhase 3 = Unfreezing\nPhase 4 = Adding mixup\nFeel free to contact me if I can provide some more details!
", "votes": null, "replies": [] }, { "id": 553635, "author_name": "hearstml", "author_url": "", "post_date": "06/16/2019 03:27:19", "content": "Hey @kurtjanssens , thank you very much! Appreciate that! :)
", "votes": null, "replies": [] }, { "id": 554291, "author_name": "valanm", "author_url": "", "post_date": "06/17/2019 08:32:37", "content": "HI, i have shared too
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "541340": "It would be good to get the commentary from the winning solutions and also from the top Solutions. \nI did not concentrate much on this. However the competition I had concentrated was iCassava Challenge where also I received the 29th place among 87 teams. My solution was based on Densenet with fastai library.\n\nThanks again!", "544256": "I am very interested in the solution from [@lzhbrian](https://www.kaggle.com/lzhbrian) who clearly did something different as his/her error is ~4x smaller compared to the 2nd place. I have no idea how someone could achieve such a high score. \n\nI finished 4th and very close to 2nd and 3rd place so I assume they did something similar to what I did which is a standard Kfold averaging (seresnext101). I did it on unprocessed images which scores 0.99+ on public. Errors are mainly images with only upper body present so I did Kfold with cropped upper body and scored 0.96+ and after averaging it reached 0.993+. Here the problematic images were those with a lot of pixels removed by organizers in their cropping pipeline. Perhaps [@lzhbrian](https://www.kaggle.com/lzhbrian) also used version1 of the dataset which had all the pixels present as this could certainly help and he/she had access to it unlike many who joined late. This is only my speculation motivated by 3 days of silence from the winning teams.", "544470": "Unfortunately I joined the competition late and didn't have access to the version 1 of the test set.\nI'm curious how my performance would be if I had access to test set version. \nMaybe someone can share it?\n\nI wanted to use stratified K-fold averaging with 8 folds but only had time to train 2 of the folds for the following models:\n- pnasnet5large\n- senet154\n- se-resnext101-32x4d\n- se-resnet152\n\nI used the Fast.ai framework and used the following 4 stage training:\n\n```\nif STAGE == 1:\n EPOCHS = 40\n SIZE = 'S'\n TRANSFORM = False\n LR = 1e-2\n REDUCE_FACTOR = 0.9\n LAYERS = 'FREEZE'\n MIXUP = False\n \nif STAGE == 2:\n EPOCHS = 24 # 12 or 24 (resnet)\n SIZE = 'L'\n TRANSFORM = True\n LR = 5e-3 # 1e-2 a 1e-3\n REDUCE_FACTOR = 0.8\n LAYERS = 'FREEZE'\n MIXUP = False\n\nif STAGE == 3:\n EPOCHS = 16 #8 or 16 (resnet)\n SIZE = 'L'\n TRANSFORM = True\n LR = 1e-4 ### 1e-4 a 1e-5\n REDUCE_FACTOR = 0.8\n LAYERS = 'UNFREEZE'\n MIXUP = False\n \nif STAGE == 4:\n EPOCHS = 16 #8 or 16 (resnet)\n SIZE = 'L'\n TRANSFORM = True\n LR = 1e-5 ### 1e-4 a 1e-5\n REDUCE_FACTOR = 0.8\n LAYERS = 'UNFREEZE'\n MIXUP = True\n\nx,y = 226,80\nif SIZE == 'S':\n SIZE = (x,y)\nif SIZE == 'L':\n SIZE = (x*2-2,y*2)\n```\n\nI came upon this stages by experimenting and reading upon best practices from fast.ai. \nFeel free to comment any suggestions!", "544472": "Thanks @kurtjanssens. Would it be possible to share the code ?\n\nThe 4 stage training is very interesting! Would you point to the lecture in fast.ai they are using this ?\n\nI also used fast.ai", "544741": "Hi all, sorry for a silence from our side but we are participating in another challenges so we will share our description later on. Thank you for your patience.\n\nBTW anyone attending CVPR?", "546198": "I did about the same things and I'm also wondering what Izhbrian did to get ~4x smaller error compared to the 2nd place", "547576": "Greetings all,\n\nThank you for the participation in this competition. If you are a top 3rd place winner, please email the Jupyter Notebook containing the approach for your winning solution to us to validate and then issue the prize money. We will confirm the accuracy score for all winning models were based on the final set of training images uploaded (i.e., not the earlier set as some had expressed concern). Our email address is hearstml@gmail.com\n\nMany thanks,\nHearst ML", "551362": "**Hey @picekl , @aleszita and @dsvolkov ,**\n\ncan you please share the code/Jupyter Notebook containing the approach for your winning solution to us to validate and then issue the prize money.?\n\n- We will need to confirm the accuracy score for all winning models were based on the final set of training images uploaded (i.e., not the earlier set as some had expressed concern). \n- We will also need to make a presentation at FGVC2019 at CVPR based on your approaches. You can send us 1-2 google slides (or pdf) describing your methods. Or we can edit it on our side.\n- You also have access to a 4 foot x 4 foot poster board if you want to present your method at the workshop (we cannot do remote presentations or videos)\n- If you are planning to join the FGVC workshop next Monday, please let us know. Maybe we could meet and discuss some potential collaboration opportunities in the future :)\n\nThank you! \n---------------------------------------------------------------------------------------------------------------------------------------\n**For the other participants:**\nWe also want to say thank you all for the other participants and you are also welcome to post a description of your approach on Kaggle!\n\nOur email address is hearstml@gmail.com", "551534": "Hello Hearst ML! Did not notice your first message on the forum. I'm going to send a solution in a couple of days. Thank you", "551567": "Hey @picekl , We will come to CVPR . Will you come to FGVC workshop next Monday? Maybe we can talk a little bit if you will come. Can you please also share your solution since we will need to present your solution at FGVC workshop at CVPR (please see my another message about prize below). Thanks!", "551568": "Hey @dsvolkov , thank you for the message! If you could send us your solution during the weekend, it would be great. Just because we need to present your solution at FGVC workshop at CVPR. BTW, are you gonna be at FGVC workshop next Monday? Thank you!", "552012": "I am still very interested in the first place solution. Any news with that regard?", "552817": "Hey @picekl, @aleszita and @dsvolkov \nsorry to bother you again on this: If you could send us your solution during the weekend, it would be great. Just because we need to present your solution at FGVC workshop at CVPR **next Monday (6/17).** BTW, are you gonna be at FGVC workshop next Monday? Thank you!", "552818": "unfortunately, we didn't :(", "552822": "Dear @valanm , @shentao , @garybios , @kurtjanssens , I am wondering if you could send us your solution of iDesigner during the weekend. I am trying to reach the Top 3 team. If I cannot receive their response before **next Monday (for FGVC workshop),** I hope we can present your solution at FGVC workshop at CVPR if you are interested in. BTW, are you gonna be at FGVC workshop next Monday? Thank you!", "552847": "Hey Hearst ML! \nI already sent a code for the inference and all pretrained models to your mail address. I will definitely send you training pipeline during the weekend. \n\nUnfortunately, I cannot participate the FGVC workshop, as it is 14 hours of flight from me (and I do not have a visa anyway)", "552896": "Hey @dsvolkov ,\nthanks for your reply! \n1) we will wait your training pipeline during the weekend.\n2) we will issue the prize after validating your model.\n3) I see. Maybe hope to see you on the FGVC next year :)", "553105": "Hi I am coming back from vacation and will get back to you tomorrow. Sorry for the inconvenience", "553296": "Hello @hearstml, \nI've given you access to my the notebook I used for training the models.\nI've used a 4 stage approach as outlined above.\nThe idea of training is stages is to increase the difficulty of the problem in each stage.\nPhase 2 = Increasing the size\nPhase 3 = Unfreezing\nPhase 4 = Adding mixup\nFeel free to contact me if I can provide some more details!", "553305": "Sure, I can share the notebook! But I can't find your username when trying to share the kernel. I've send you a direct message with the github link.", "553323": "hearstml \n\nWould it be possible to share the code and the winning approaches to all of us ?\n\nCheers", "553417": "kurtjanssens \n\nThanks. But I could not get from the github link. It is showing 404 error\n\nMy user name is @ambarish. https://www.kaggle.com/ambarish is my profile page.\n\nCheers\nBukun ( Ambarish )", "553635": "Hey @kurtjanssens , thank you very much! Appreciate that! :)", "554291": "HI, i have shared too", "554614": "Hi @valanm -> here you have brief overview of used method. It's similar to my and my college article published last year and used for previous edition of FGVC. If you have any questions please feel free to ask. I will try to answer all of them. \n\nProposed system is an ensemble of multiple Convolutional Neural Network architectures such as Inception-ResNet-v2, Inception-v4 [1] and PNASNet[2], improved by:\n - Multiple level training where:\n - 1st level - trains only last layer - 10 epochs - rmsprop - Initial LR 0.01 with exponential decay 0.9 per epoch.\n - 2nd level - trains whole net - 10 epochs - rmsprop - Initial LR 0.01 with exponential decay 0.9 per epoch.\n - 3rd same as first - 10 epochs - rmsprop - rmsprop - Initial LR 0.0075 with exponential decay 0.9 per epoch.\n - 4th same as second - 10 epochs - rmsprop - rmsprop - Initial LR 0.0075 with exponential decay 0.9 per epoch.\n - Using random crops to keep image aspect ration as well as get better detail on images with higher resolution.\n - Using running averages on the trained variables with an exponential weight decay.\n - Adjusting predictions according to the estimated change of the class prior probabilities [3,4].\n - Test-time data augmentation with multiple predictions that covers image with an overlap.\n - Summing the soft-max values from multiple crops/nets and combining that with MODUS “decision” on images with low top1 prediction accuracy.\n \n For training I used TensorFlow slim.\n \n [1] Szegedy, C., Ioffe, S., Vanhoucke, V.: Inception-v4, Inception-ResNet and the impact of residual connections on learning. arXiv:1602.07261 (2016)\n [2] C. Liu, B. Zoph, J. Shlens, W. Hua, L.-J. Li, L. Fei-Fei, A. Yuille, J. Huang, and K. Murphy. Progressive neural architecture search. ECCV, 2018.\n [3] Sulc, M., Matas, J.: Improving cnn classifiers by estimating test-time priors. arXiv:1805.08235 (2018)\n [4] Saerens, M., Latinne, P., Decaestecker, C.: Adjusting the outputs of a classifier to new a priori probabilities: a simple procedure. Neural computation 14(1), 21–41 (2002)" }, "source": "meta" }