This was my first serious go at a kaggle competition, and I thought I'd share the work I did and my lessons learned.
\nMost of code I wrote was for image packing, and saliency/warp sampling. Saliency sampling seemed like a possible way to incorporate maximally high resolution information, and my intent was to bank on the possibility that that could improve scores.
\nI wasn't able to complete my project, but looking at winning/high ranked solutions, I'm increasingly convinced my focus was incorrect. Rather than focusing on overarching architecture, I should have focused on de-noising, model experimentation, and perhaps tweaking @iafoss's tile selection approach.
\nPerhaps in the future I'll take the time to investigate the fundamentals of what more successful participants ended up doing.
\nAnyways, I wanted to at least post something. This was my first time really working with modern deep learning tools and I ended up learning a lot despite not getting out an actual submission.
", "messages": [ { "id": 942311, "postDate": "2020-07-23T17:34:22.940Z", "content": "This was my first serious go at a kaggle competition, and I thought I'd share the work I did and my lessons learned.
\nMost of code I wrote was for image packing, and saliency/warp sampling. Saliency sampling seemed like a possible way to incorporate maximally high resolution information, and my intent was to bank on the possibility that that could improve scores.
\nI wasn't able to complete my project, but looking at winning/high ranked solutions, I'm increasingly convinced my focus was incorrect. Rather than focusing on overarching architecture, I should have focused on de-noising, model experimentation, and perhaps tweaking @iafoss's tile selection approach.
\nPerhaps in the future I'll take the time to investigate the fundamentals of what more successful participants ended up doing.
\nAnyways, I wanted to at least post something. This was my first time really working with modern deep learning tools and I ended up learning a lot despite not getting out an actual submission.
", "rawMarkdown": "This was my first serious go at a kaggle competition, and I thought I'd share the work I did and my lessons learned.\n\nMost of code I wrote was for [image packing], and [saliency/warp sampling]. Saliency sampling seemed like a possible way to incorporate maximally high resolution information, and my intent was to bank on the possibility that that could improve scores.\n\nI wasn't able to complete my project, but looking at winning/high ranked solutions, I'm increasingly convinced my focus was incorrect. Rather than focusing on overarching architecture, I should have focused on de-noising, model experimentation, and perhaps tweaking @iafoss's tile selection approach.\n\nPerhaps in the future I'll take the time to investigate the fundamentals of what more successful participants ended up doing. \n\nAnyways, I wanted to at least post something. This was my first time really working with modern deep learning tools and I ended up learning a lot despite not getting out an actual submission.\n\n[image packing]: https://www.kaggle.com/micimize/complex-packing\n[saliency/warp sampling]: https://colab.research.google.com/drive/16bw7Nut3GQwkKL4soX80g9YdYgcTqsF6?usp=sharing", "votes": 6 }, { "id": 942324, "postDate": "2020-07-23T17:42:20.150Z", "content": "read carefully the oraganiser notes . It was mentioned that noisy labels are there.\nI even created one regression model to predict labels and did not believe on model because of overlapping of labels .\nI though that model is wrong , but no , label were wrongs. Fall from 41 to 536.
", "rawMarkdown": "read carefully the oraganiser notes . It was mentioned that noisy labels are there.\nI even created one regression model to predict labels and did not believe on model because of overlapping of labels .\nI though that model is wrong , but no , label were wrongs. Fall from 41 to 536.\n", "votes": 1 } ], "comments": [ { "id": 942324, "author_name": "Rajnish Chauhan", "author_url": "", "post_date": "2020-07-23T17:42:20.150000", "content": "read carefully the oraganiser notes . It was mentioned that noisy labels are there.\nI even created one regression model to predict labels and did not believe on model because of overlapping of labels .\nI though that model is wrong , but no , label were wrongs. Fall from 41 to 536.
", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "942311": "This was my first serious go at a kaggle competition, and I thought I'd share the work I did and my lessons learned.\n\nMost of code I wrote was for [image packing], and [saliency/warp sampling]. Saliency sampling seemed like a possible way to incorporate maximally high resolution information, and my intent was to bank on the possibility that that could improve scores.\n\nI wasn't able to complete my project, but looking at winning/high ranked solutions, I'm increasingly convinced my focus was incorrect. Rather than focusing on overarching architecture, I should have focused on de-noising, model experimentation, and perhaps tweaking @iafoss's tile selection approach.\n\nPerhaps in the future I'll take the time to investigate the fundamentals of what more successful participants ended up doing. \n\nAnyways, I wanted to at least post something. This was my first time really working with modern deep learning tools and I ended up learning a lot despite not getting out an actual submission.\n\n[image packing]: https://www.kaggle.com/micimize/complex-packing\n[saliency/warp sampling]: https://colab.research.google.com/drive/16bw7Nut3GQwkKL4soX80g9YdYgcTqsF6?usp=sharing", "942324": "read carefully the oraganiser notes . It was mentioned that noisy labels are there.\nI even created one regression model to predict labels and did not believe on model because of overlapping of labels .\nI though that model is wrong , but no , label were wrongs. Fall from 41 to 536.\n" } }