I would be very interested to know the solutions of the top entries and also of any other entries after the competition ends. I tried some experiments with fast.ai but it seems that there were obviously far better entries.
\n\nIts great to see the entries of 1st,2nd and 3rd positions. I would also love to hear about the solutions in the top 20. Thank you again!
", "messages": [ { "id": "540099", "postDate": "05/31/2019 03:09:06", "content": "I would be very interested to know the solutions of the top entries and also of any other entries after the competition ends. I tried some experiments with fast.ai but it seems that there were obviously far better entries.
\n\nIts great to see the entries of 1st,2nd and 3rd positions. I would also love to hear about the solutions in the top 20. Thank you again!
", "rawMarkdown": "I would be very interested to know the solutions of the top entries and also of any other entries after the competition ends. I tried some experiments with fast.ai but it seems that there were obviously far better entries.\n\nIts great to see the entries of 1st,2nd and 3rd positions. I would also love to hear about the solutions in the top 20. Thank you again!", "votes": null }, { "id": "540694", "postDate": "05/31/2019 21:39:14", "content": "I would also be interested to see what approaches people took. I definitely hit a ceiling with the supplied training data. I'm guessing people were able to leverage the \"extra images\" but I didn't get a chance to dig into that. I would be curious to know if and how people were able to use them.
", "rawMarkdown": "I would also be interested to see what approaches people took. I definitely hit a ceiling with the supplied training data. I'm guessing people were able to leverage the \"extra images\" but I didn't get a chance to dig into that. I would be curious to know if and how people were able to use them.", "votes": null }, { "id": "540875", "postDate": "06/01/2019 09:23:49", "content": "I first hit the accuracy ceiling at 81% with MobileNetV2 trained for classification on the training set with heavy artificial data augmentation. \nThen I tried a few pretext learning/unsupervised learning techniques to take advantage of the extraimages dataset. Specifically, I tried Auto-colorization [1], Jigsaw puzzling [2] and Egomotion [3]. The first two did not really converge: my guesses are, respectively, that the dataset is basically composed on mostly green images so there wasn't much colorization to learn, and the images do not have strong structural properties (foliage) to help with jigsaw puzzling. The third approach did converge, and after transfer learning on to the training set it increased accuracy to about 83%.\nI recognize a few issues in my solution, for instance I crop images to a square the size of the shorter edge of the picture, and then resize to 224x224: a better cropping can probably increase accuracy. \nStill, the key to the competition probably lies in a better exploitation of the extraimages, so I'm curious as to how it was done by the winners.
\n\nCheers
\n\n[1] Learning Representations for Automatic Colorization. Larsson, Gustav and Maire, Michael and Shakhnarovich, Gregory. ECCV 2016\n[2] Unsupervised Learning of Visual Representations by Solving Jigsaw Puzzles. Noroozi, Mehdi and Favaro, Paolo. ECCV 2016\n[3] Learning to See by Moving. Agrawal, Pulkit and Carreira, Joao and Malik, Jitendra. ICCV 2015
", "rawMarkdown": "I first hit the accuracy ceiling at 81% with MobileNetV2 trained for classification on the training set with heavy artificial data augmentation. \nThen I tried a few pretext learning/unsupervised learning techniques to take advantage of the extraimages dataset. Specifically, I tried Auto-colorization [1], Jigsaw puzzling [2] and Egomotion [3]. The first two did not really converge: my guesses are, respectively, that the dataset is basically composed on mostly green images so there wasn't much colorization to learn, and the images do not have strong structural properties (foliage) to help with jigsaw puzzling. The third approach did converge, and after transfer learning on to the training set it increased accuracy to about 83%.\nI recognize a few issues in my solution, for instance I crop images to a square the size of the shorter edge of the picture, and then resize to 224x224: a better cropping can probably increase accuracy. \nStill, the key to the competition probably lies in a better exploitation of the extraimages, so I'm curious as to how it was done by the winners.\n\nCheers\n\n\n\n[1] *Learning Representations for Automatic Colorization*. Larsson, Gustav and Maire, Michael and Shakhnarovich, Gregory. ECCV 2016\n[2] *Unsupervised Learning of Visual Representations by Solving Jigsaw Puzzles.* Noroozi, Mehdi and Favaro, Paolo. ECCV 2016\n[3] *Learning to See by Moving.* Agrawal, Pulkit and Carreira, Joao and Malik, Jitendra. ICCV 2015", "votes": null }, { "id": "540903", "postDate": "06/01/2019 10:24:04", "content": "Thanks @pjkocks for your solution. \nI had used the fastai library with Densenet121 with data augmentations.
", "rawMarkdown": "Thanks @pjkocks for your solution. \nI had used the fastai library with Densenet121 with data augmentations.", "votes": null }, { "id": "541103", "postDate": "06/01/2019 18:14:40", "content": "Thank you @pjkocks! I will look into these learning techniques.
", "rawMarkdown": "Thank you @pjkocks! I will look into these learning techniques.", "votes": null }, { "id": "541261", "postDate": "06/02/2019 05:39:36", "content": "29th place solution is provided here
\n\nhttps://www.kaggle.com/ambarish/icassava-2019-data-analysis-dropout-params?scriptVersionId=13850110\nUses fastai\nUses densenet121\nUses dropout ps = 0.25\nUses fastai data augmentation\nepochs = 40
", "rawMarkdown": "29th place solution is provided here\n\nhttps://www.kaggle.com/ambarish/icassava-2019-data-analysis-dropout-params?scriptVersionId=13850110\nUses fastai\nUses densenet121\nUses dropout ps = 0.25\nUses fastai data augmentation\nepochs = 40", "votes": null } ], "comments": [ { "id": 540694, "author_name": "jedhassell", "author_url": "", "post_date": "05/31/2019 21:39:14", "content": "I would also be interested to see what approaches people took. I definitely hit a ceiling with the supplied training data. I'm guessing people were able to leverage the \"extra images\" but I didn't get a chance to dig into that. I would be curious to know if and how people were able to use them.
", "votes": null, "replies": [] }, { "id": 540875, "author_name": "pjkocks", "author_url": "", "post_date": "06/01/2019 09:23:49", "content": "I first hit the accuracy ceiling at 81% with MobileNetV2 trained for classification on the training set with heavy artificial data augmentation. \nThen I tried a few pretext learning/unsupervised learning techniques to take advantage of the extraimages dataset. Specifically, I tried Auto-colorization [1], Jigsaw puzzling [2] and Egomotion [3]. The first two did not really converge: my guesses are, respectively, that the dataset is basically composed on mostly green images so there wasn't much colorization to learn, and the images do not have strong structural properties (foliage) to help with jigsaw puzzling. The third approach did converge, and after transfer learning on to the training set it increased accuracy to about 83%.\nI recognize a few issues in my solution, for instance I crop images to a square the size of the shorter edge of the picture, and then resize to 224x224: a better cropping can probably increase accuracy. \nStill, the key to the competition probably lies in a better exploitation of the extraimages, so I'm curious as to how it was done by the winners.
\n\nCheers
\n\n[1] Learning Representations for Automatic Colorization. Larsson, Gustav and Maire, Michael and Shakhnarovich, Gregory. ECCV 2016\n[2] Unsupervised Learning of Visual Representations by Solving Jigsaw Puzzles. Noroozi, Mehdi and Favaro, Paolo. ECCV 2016\n[3] Learning to See by Moving. Agrawal, Pulkit and Carreira, Joao and Malik, Jitendra. ICCV 2015
", "votes": null, "replies": [ { "id": 540903, "author_name": "ambarish", "author_url": "", "post_date": "06/01/2019 10:24:04", "content": "Thanks @pjkocks for your solution. \nI had used the fastai library with Densenet121 with data augmentations.
", "votes": null, "replies": [] }, { "id": 541103, "author_name": "jedhassell", "author_url": "", "post_date": "06/01/2019 18:14:40", "content": "Thank you @pjkocks! I will look into these learning techniques.
", "votes": null, "replies": [] } ] }, { "id": 541261, "author_name": "ambarish", "author_url": "", "post_date": "06/02/2019 05:39:36", "content": "29th place solution is provided here
\n\nhttps://www.kaggle.com/ambarish/icassava-2019-data-analysis-dropout-params?scriptVersionId=13850110\nUses fastai\nUses densenet121\nUses dropout ps = 0.25\nUses fastai data augmentation\nepochs = 40
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "540099": "I would be very interested to know the solutions of the top entries and also of any other entries after the competition ends. I tried some experiments with fast.ai but it seems that there were obviously far better entries.\n\nIts great to see the entries of 1st,2nd and 3rd positions. I would also love to hear about the solutions in the top 20. Thank you again!", "540694": "I would also be interested to see what approaches people took. I definitely hit a ceiling with the supplied training data. I'm guessing people were able to leverage the \"extra images\" but I didn't get a chance to dig into that. I would be curious to know if and how people were able to use them.", "540875": "I first hit the accuracy ceiling at 81% with MobileNetV2 trained for classification on the training set with heavy artificial data augmentation. \nThen I tried a few pretext learning/unsupervised learning techniques to take advantage of the extraimages dataset. Specifically, I tried Auto-colorization [1], Jigsaw puzzling [2] and Egomotion [3]. The first two did not really converge: my guesses are, respectively, that the dataset is basically composed on mostly green images so there wasn't much colorization to learn, and the images do not have strong structural properties (foliage) to help with jigsaw puzzling. The third approach did converge, and after transfer learning on to the training set it increased accuracy to about 83%.\nI recognize a few issues in my solution, for instance I crop images to a square the size of the shorter edge of the picture, and then resize to 224x224: a better cropping can probably increase accuracy. \nStill, the key to the competition probably lies in a better exploitation of the extraimages, so I'm curious as to how it was done by the winners.\n\nCheers\n\n\n\n[1] *Learning Representations for Automatic Colorization*. Larsson, Gustav and Maire, Michael and Shakhnarovich, Gregory. ECCV 2016\n[2] *Unsupervised Learning of Visual Representations by Solving Jigsaw Puzzles.* Noroozi, Mehdi and Favaro, Paolo. ECCV 2016\n[3] *Learning to See by Moving.* Agrawal, Pulkit and Carreira, Joao and Malik, Jitendra. ICCV 2015", "540903": "Thanks @pjkocks for your solution. \nI had used the fastai library with Densenet121 with data augmentations.", "541103": "Thank you @pjkocks! I will look into these learning techniques.", "541261": "29th place solution is provided here\n\nhttps://www.kaggle.com/ambarish/icassava-2019-data-analysis-dropout-params?scriptVersionId=13850110\nUses fastai\nUses densenet121\nUses dropout ps = 0.25\nUses fastai data augmentation\nepochs = 40" }, "source": "meta" }