![\"Kaggle](\"https://raw.githubusercontent.com/stared/random_data_explorations/master/201704_dstl_kaggle_classes/kaggle_dstl_iou_per_class.png\" "\\"\\"")
https://docs.google.com/spreadsheets/d/1wi0ydVfaZ9hRFT8-Wgw0FSpLv55QPJclgyZZUXuz8Es/edit?usp=sharing
", "messages": [ { "id": 178476, "postDate": "2017-04-27T22:44:20.233Z", "content": "https://docs.google.com/spreadsheets/d/1wi0ydVfaZ9hRFT8-Wgw0FSpLv55QPJclgyZZUXuz8Es/edit?usp=sharing
", "rawMarkdown": "https://docs.google.com/spreadsheets/d/1wi0ydVfaZ9hRFT8-Wgw0FSpLv55QPJclgyZZUXuz8Es/edit?usp=sharing", "votes": 7 }, { "id": 178639, "postDate": "2017-04-28T13:37:17.860Z", "content": "Great!
\n\nI created plots of of them, for cleaner comparison: Jupyter Notebook here:
\n\n
![\"Kaggle](\"https://raw.githubusercontent.com/stared/random_data_explorations/master/201704_dstl_kaggle_classes/kaggle_dstl_classes_zscore_per_team.png\" "\\"\\"")
Hi, sir, I am using your code for satellite imagery object detect, is possible to feed only RGB images instead of 16 bands to your pre-trained model of dstl competition 3rd position holder architecture. Thanks
", "rawMarkdown": "Hi, sir, I am using your code for satellite imagery object detect, is possible to feed only RGB images instead of 16 bands to your pre-trained model of dstl competition 3rd position holder architecture. Thanks " }, { "id": 178618, "postDate": "2017-04-28T12:41:09.787Z", "content": "Hi Vladimir,
\n\nThank you for sharing top players' LB scores.
\nWhat is Mega merge ? Do you try ensemble all result ?
It is just an estimate of what will happen if top 5 teams merged into one and picked the best result per class.
", "rawMarkdown": "It is just an estimate of what will happen if top 5 teams merged into one and picked the best result per class.", "votes": 1 } ] }, { "id": 178497, "postDate": "2017-04-28T01:38:17.187Z", "content": "Hi Vladimir, where do these numbers come from? Do you have the top5 submissions and used them to probe the private LB?
\n\nWhy does your 3rd place solution (Congratulations by the way, outstanding work!!!!!) score 0.47874 on the private LB and this sheet shows 0.47542? Is there just a typo in one of the classes?
", "rawMarkdown": "Hi Vladimir, where do these numbers come from? Do you have the top5 submissions and used them to probe the private LB?\n\nWhy does your 3rd place solution (Congratulations by the way, outstanding work!!!!!) score 0.47874 on the private LB and this sheet shows 0.47542? Is there just a typo in one of the classes?\n\n", "replies": [ { "id": 178504, "postDate": "2017-04-28T02:08:21.257Z", "content": "In our final submission there was an extra 0.00331 that corresponds to the Large Vehicle class from one of the tries, that I did not remember how to reproduce => I excluded it from the joint top places class table. Motivation is that for all other classes we can easily reproduce our results (we have code to do it), but not fo rit.
\n\n=> We did not include any code for a Vehicle predictions in the code that we provided to organizers.
\n\nBut! When we were cleaning the code and verified that we can reproduce our results, we released that we able to improve our predictions:
\n\nWhich allowed us to have code that would score 0.4876 at the Private LB using only first 8 classes. Which is more than was in our final submission => good enough to claim a prize.
", "rawMarkdown": " 1. Shared in the [blog post][1] \n 2. Shared results with us \n 3. We know our score :)\n 4. Shared at the forum\n 5. Shared results with us\n\nIn our final submission there was an extra 0.00331 that corresponds to the Large Vehicle class from one of the tries, that I did not remember how to reproduce => I excluded it from the joint top places class table. Motivation is that for all other classes we can easily reproduce our results (we have code to do it), but not fo rit. \n\n=> We did not include any code for a Vehicle predictions in the code that we provided to organizers. \n\nBut! When we were cleaning the code and verified that we can reproduce our results, we released that we able to improve our predictions:\n\n - 0.06290 => 0.06855 for buildings\n - 0.02015 => 0.02414 for structures\n - 0.08280 => 0.08534 for crops\n\nWhich allowed us to have code that would score 0.4876 at the Private LB using only first 8 classes. Which is more than was in our final submission => good enough to claim a prize.\n\n\n [1]: http://blog.kaggle.com/2017/04/26/dstl-satellite-imagery-competition-1st-place-winners-interview-kyle-lee/", "votes": 1 }, { "id": 196485, "postDate": "2017-06-27T13:28:45.563Z", "rawMarkdown": "", "isDeleted": true } ] }, { "id": 196889, "postDate": "2017-06-28T10:24:04.730Z", "content": "Hi @Vladimir, thanks for this and the great 3rd place write-up.
\n\nHow did you work out your \"per-class\" scores?
\n\nI'm currently looking into this problem as a learning exercise. I'm hoping to compare my individual class scores against the leaders, but Kaggle only returns me a global score (i.e. my score based on the entire datasets and all classes).
\n\nMany thanks!
", "rawMarkdown": "Hi @Vladimir, thanks for this and the great 3rd place write-up.\n\nHow did you work out your \"per-class\" scores?\n\nI'm currently looking into this problem as a learning exercise. I'm hoping to compare my individual class scores against the leaders, but Kaggle only returns me a global score (i.e. my score based on the entire datasets and all classes).\n\nMany thanks!", "isDeleted": true, "replies": [ { "id": 197106, "postDate": "2017-06-28T20:00:27.400Z", "content": "Metrics in this competition allows LB probing to get per class score. If you take your submission and zero out everything except rows in which you predict, say buildings - you will get the score for buildings. If you zero out everything except crops - you will get the score for crops, etc.
", "rawMarkdown": "Metrics in this competition allows LB probing to get per class score. If you take your submission and zero out everything except rows in which you predict, say buildings - you will get the score for buildings. If you zero out everything except crops - you will get the score for crops, etc." } ] } ], "comments": [ { "id": 178639, "author_name": "Piotr Migdał", "author_url": "", "post_date": "2017-04-28T13:37:17.860000", "content": "Great!
\n\nI created plots of of them, for cleaner comparison: Jupyter Notebook here:
\n\n
Hi, sir, I am using your code for satellite imagery object detect, is possible to feed only RGB images instead of 16 bands to your pre-trained model of dstl competition 3rd position holder architecture. Thanks
", "votes": 0, "replies": [] }, { "id": 178618, "author_name": "toshi_k", "author_url": "", "post_date": "2017-04-28T12:41:09.787000", "content": "Hi Vladimir,
\n\nThank you for sharing top players' LB scores.
\nWhat is Mega merge ? Do you try ensemble all result ?
It is just an estimate of what will happen if top 5 teams merged into one and picked the best result per class.
", "votes": 1, "replies": [] } ] }, { "id": 178497, "author_name": "JeffH", "author_url": "", "post_date": "2017-04-28T01:38:17.187000", "content": "Hi Vladimir, where do these numbers come from? Do you have the top5 submissions and used them to probe the private LB?
\n\nWhy does your 3rd place solution (Congratulations by the way, outstanding work!!!!!) score 0.47874 on the private LB and this sheet shows 0.47542? Is there just a typo in one of the classes?
", "votes": 0, "replies": [ { "id": 178504, "author_name": "Vladimir Iglovikov", "author_url": "", "post_date": "2017-04-28T02:08:21.257000", "content": "In our final submission there was an extra 0.00331 that corresponds to the Large Vehicle class from one of the tries, that I did not remember how to reproduce => I excluded it from the joint top places class table. Motivation is that for all other classes we can easily reproduce our results (we have code to do it), but not fo rit.
\n\n=> We did not include any code for a Vehicle predictions in the code that we provided to organizers.
\n\nBut! When we were cleaning the code and verified that we can reproduce our results, we released that we able to improve our predictions:
\n\nWhich allowed us to have code that would score 0.4876 at the Private LB using only first 8 classes. Which is more than was in our final submission => good enough to claim a prize.
", "votes": 1, "replies": [] }, { "id": 196485, "author_name": "", "author_url": "", "post_date": "2017-06-27T13:28:45.563000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 196889, "author_name": "", "author_url": "", "post_date": "2017-06-28T10:24:04.730000", "content": "Hi @Vladimir, thanks for this and the great 3rd place write-up.
\n\nHow did you work out your \"per-class\" scores?
\n\nI'm currently looking into this problem as a learning exercise. I'm hoping to compare my individual class scores against the leaders, but Kaggle only returns me a global score (i.e. my score based on the entire datasets and all classes).
\n\nMany thanks!
", "votes": 0, "replies": [ { "id": 197106, "author_name": "Vladimir Iglovikov", "author_url": "", "post_date": "2017-06-28T20:00:27.400000", "content": "Metrics in this competition allows LB probing to get per class score. If you take your submission and zero out everything except rows in which you predict, say buildings - you will get the score for buildings. If you zero out everything except crops - you will get the score for crops, etc.
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "178476": "https://docs.google.com/spreadsheets/d/1wi0ydVfaZ9hRFT8-Wgw0FSpLv55QPJclgyZZUXuz8Es/edit?usp=sharing", "178639": "Great!\n\nI created plots of of them, for cleaner comparison: [Jupyter Notebook here][1]:\n\n![Kaggle Dstl IoU per class][2]\n\n![Kaggle Dstl IoU per class z-score per team][3]\n\n [1]: https://github.com/stared/random_data_explorations/blob/master/201704_dstl_kaggle_classes/per_class_plots_of_winners.ipynb\n [2]: https://raw.githubusercontent.com/stared/random_data_explorations/master/201704_dstl_kaggle_classes/kaggle_dstl_iou_per_class.png\n [3]: https://raw.githubusercontent.com/stared/random_data_explorations/master/201704_dstl_kaggle_classes/kaggle_dstl_classes_zscore_per_team.png", "318740": "Hi, sir, I am using your code for satellite imagery object detect, is possible to feed only RGB images instead of 16 bands to your pre-trained model of dstl competition 3rd position holder architecture. Thanks ", "178618": "Hi Vladimir,\n\nThank you for sharing top players' LB scores.