Given that for the last few days I've done nothing beside watching a progress bar to finish, I guess it may reach ~0.6 with 512x512.
\n\nIf some of you millionaires decide to use 2048x2048 it can reach ~0.62 I think. What are your thoughts?
", "messages": [ { "id": "412964", "postDate": "10/31/2018 04:04:44", "content": "Given that for the last few days I've done nothing beside watching a progress bar to finish, I guess it may reach ~0.6 with 512x512.
\n\nIf some of you millionaires decide to use 2048x2048 it can reach ~0.62 I think. What are your thoughts?
", "rawMarkdown": "Given that for the last few days I've done nothing beside watching a progress bar to finish, I guess it may reach ~0.6 with 512x512.\n\nIf some of you millionaires decide to use 2048x2048 it can reach ~0.62 I think. What are your thoughts?", "votes": null }, { "id": "413077", "postDate": "10/31/2018 08:37:56", "content": "\"Combining the annotations of gamers and deep learning, we applied transfer learning to create a boosted learner that can characterize subcellular protein distribution with F1 score of 0.72\"
\n\nhttps://www.nature.com/articles/nbt.4225
\n\nhttps://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/70144#latest-413071
", "rawMarkdown": "\"Combining the annotations of gamers and deep learning, we applied transfer learning to create a boosted learner that can characterize subcellular protein distribution with F1 score of 0.72\"\n\n\nhttps://www.nature.com/articles/nbt.4225\n\nhttps://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/70144#latest-413071", "votes": null }, { "id": "413086", "postDate": "10/31/2018 08:56:41", "content": "", "rawMarkdown": "", "votes": null }, { "id": "413118", "postDate": "10/31/2018 09:53:46", "content": "Now the problem is: How do I gather my local nerds and convince them to improve my models?
", "rawMarkdown": "Now the problem is: How do I gather my local nerds and convince them to improve my models?", "votes": null }, { "id": "413474", "postDate": "11/01/2018 02:29:49", "content": "My bet is ~0.7-0.8...\nBTW, Khoi Nguyen, you wrote that you tried to use full resolution images, did you achieve any success with them?
", "rawMarkdown": "My bet is ~0.7-0.8...\nBTW, Khoi Nguyen, you wrote that you tried to use full resolution images, did you achieve any success with them?", "votes": null }, { "id": "413477", "postDate": "11/01/2018 02:40:31", "content": "It was a 3% increased in F1 so I won't be surprised if someone uses the 2048x2048.
", "rawMarkdown": "It was a 3% increased in F1 so I won't be surprised if someone uses the 2048x2048.", "votes": null }, { "id": "414083", "postDate": "11/02/2018 05:00:56", "content": "I just started with 1536x1536 based on the 2048x2048 images. A few days later maybe I'll have some results at the speed it is running, 2500 seconds per epoch.
", "rawMarkdown": "I just started with 1536x1536 based on the 2048x2048 images. A few days later maybe I'll have some results at the speed it is running, 2500 seconds per epoch.", "votes": null }, { "id": "414243", "postDate": "11/02/2018 11:44:38", "content": "256x256, local LB w. constant threshold 0.5: 0.72
\n\nI believe that 512x512 can bring it to 0.8.....
", "rawMarkdown": "256x256, local LB w. constant threshold 0.5: 0.72\n\nI believe that 512x512 can bring it to 0.8.....", "votes": null }, { "id": "414358", "postDate": "11/02/2018 15:48:21", "content": "Locally my best models perform over 0.8 at 512x512. Public LB is still much lower.
", "rawMarkdown": "Locally my best models perform over 0.8 at 512x512. Public LB is still much lower.", "votes": null }, { "id": "414387", "postDate": "11/02/2018 16:31:24", "content": "Yes it's easy to reach 0.8 in local validation. @Brian do you have any guesses why the drop happened?
", "rawMarkdown": "Yes it's easy to reach 0.8 in local validation. @Brian do you have any guesses why the drop happened?", "votes": null }, { "id": "414437", "postDate": "11/02/2018 19:07:08", "content": "I think it has to do with how the public LB is calculated. If some of the classes are not present in the public dataset, it will be treated as a class metric of 0. One of the implications of this is that private/public shakeup could be huge.
", "rawMarkdown": "I think it has to do with how the public LB is calculated. If some of the classes are not present in the public dataset, it will be treated as a class metric of 0. One of the implications of this is that private/public shakeup could be huge.", "votes": null }, { "id": "414480", "postDate": "11/02/2018 21:03:37", "content": "I agree with Alexander, I believe that the public LB may not be fully representative of the entire test set. My best submission was showing 0.87 locally.
", "rawMarkdown": "I agree with Alexander, I believe that the public LB may not be fully representative of the entire test set. My best submission was showing 0.87 locally.", "votes": null }, { "id": "418860", "postDate": "11/10/2018 18:41:14", "content": "@Brian, you have got 0.87 for local training or local validation set?
", "rawMarkdown": "Brian, you have got 0.87 for local training or local validation set?", "votes": null }, { "id": "418926", "postDate": "11/10/2018 21:59:00", "content": "Local validation set. Thinking back a week, it may have been due to a small validation set. With my latest models that still need way more training, locally .75 validation results a .373 public. Definitely needs more time...
", "rawMarkdown": "Local validation set. Thinking back a week, it may have been due to a small validation set. With my latest models that still need way more training, locally .75 validation results a .373 public. Definitely needs more time...", "votes": null }, { "id": "420356", "postDate": "11/13/2018 13:42:26", "content": "I guess it will be lower than 0.65
", "rawMarkdown": "I guess it will be lower than 0.65", "votes": null } ], "comments": [ { "id": 413077, "author_name": "hengck23", "author_url": "", "post_date": "10/31/2018 08:37:56", "content": "\"Combining the annotations of gamers and deep learning, we applied transfer learning to create a boosted learner that can characterize subcellular protein distribution with F1 score of 0.72\"
\n\nhttps://www.nature.com/articles/nbt.4225
\n\nhttps://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/70144#latest-413071
", "votes": null, "replies": [ { "id": 413118, "author_name": "suicaokhoailang", "author_url": "", "post_date": "10/31/2018 09:53:46", "content": "Now the problem is: How do I gather my local nerds and convince them to improve my models?
", "votes": null, "replies": [] } ] }, { "id": 413086, "author_name": "hengck23", "author_url": "", "post_date": "10/31/2018 08:56:41", "content": "", "votes": null, "replies": [] }, { "id": 413474, "author_name": "iafoss", "author_url": "", "post_date": "11/01/2018 02:29:49", "content": "My bet is ~0.7-0.8...\nBTW, Khoi Nguyen, you wrote that you tried to use full resolution images, did you achieve any success with them?
", "votes": null, "replies": [ { "id": 413477, "author_name": "suicaokhoailang", "author_url": "", "post_date": "11/01/2018 02:40:31", "content": "It was a 3% increased in F1 so I won't be surprised if someone uses the 2048x2048.
", "votes": null, "replies": [] }, { "id": 414083, "author_name": "ldm314", "author_url": "", "post_date": "11/02/2018 05:00:56", "content": "I just started with 1536x1536 based on the 2048x2048 images. A few days later maybe I'll have some results at the speed it is running, 2500 seconds per epoch.
", "votes": null, "replies": [] } ] }, { "id": 414243, "author_name": "alexanderliao", "author_url": "", "post_date": "11/02/2018 11:44:38", "content": "256x256, local LB w. constant threshold 0.5: 0.72
\n\nI believe that 512x512 can bring it to 0.8.....
", "votes": null, "replies": [ { "id": 414358, "author_name": "ldm314", "author_url": "", "post_date": "11/02/2018 15:48:21", "content": "Locally my best models perform over 0.8 at 512x512. Public LB is still much lower.
", "votes": null, "replies": [] }, { "id": 414387, "author_name": "suicaokhoailang", "author_url": "", "post_date": "11/02/2018 16:31:24", "content": "Yes it's easy to reach 0.8 in local validation. @Brian do you have any guesses why the drop happened?
", "votes": null, "replies": [] }, { "id": 414437, "author_name": "alexanderliao", "author_url": "", "post_date": "11/02/2018 19:07:08", "content": "I think it has to do with how the public LB is calculated. If some of the classes are not present in the public dataset, it will be treated as a class metric of 0. One of the implications of this is that private/public shakeup could be huge.
", "votes": null, "replies": [] }, { "id": 414480, "author_name": "ldm314", "author_url": "", "post_date": "11/02/2018 21:03:37", "content": "I agree with Alexander, I believe that the public LB may not be fully representative of the entire test set. My best submission was showing 0.87 locally.
", "votes": null, "replies": [] }, { "id": 418860, "author_name": "pinullmezon", "author_url": "", "post_date": "11/10/2018 18:41:14", "content": "@Brian, you have got 0.87 for local training or local validation set?
", "votes": null, "replies": [] }, { "id": 418926, "author_name": "ldm314", "author_url": "", "post_date": "11/10/2018 21:59:00", "content": "Local validation set. Thinking back a week, it may have been due to a small validation set. With my latest models that still need way more training, locally .75 validation results a .373 public. Definitely needs more time...
", "votes": null, "replies": [] } ] }, { "id": 420356, "author_name": "pinullmezon", "author_url": "", "post_date": "11/13/2018 13:42:26", "content": "I guess it will be lower than 0.65
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "412964": "Given that for the last few days I've done nothing beside watching a progress bar to finish, I guess it may reach ~0.6 with 512x512.\n\nIf some of you millionaires decide to use 2048x2048 it can reach ~0.62 I think. What are your thoughts?", "413077": "\"Combining the annotations of gamers and deep learning, we applied transfer learning to create a boosted learner that can characterize subcellular protein distribution with F1 score of 0.72\"\n\n\nhttps://www.nature.com/articles/nbt.4225\n\nhttps://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/70144#latest-413071", "413086": "", "413118": "Now the problem is: How do I gather my local nerds and convince them to improve my models?", "413474": "My bet is ~0.7-0.8...\nBTW, Khoi Nguyen, you wrote that you tried to use full resolution images, did you achieve any success with them?", "413477": "It was a 3% increased in F1 so I won't be surprised if someone uses the 2048x2048.", "414083": "I just started with 1536x1536 based on the 2048x2048 images. A few days later maybe I'll have some results at the speed it is running, 2500 seconds per epoch.", "414243": "256x256, local LB w. constant threshold 0.5: 0.72\n\nI believe that 512x512 can bring it to 0.8.....", "414358": "Locally my best models perform over 0.8 at 512x512. Public LB is still much lower.", "414387": "Yes it's easy to reach 0.8 in local validation. @Brian do you have any guesses why the drop happened?", "414437": "I think it has to do with how the public LB is calculated. If some of the classes are not present in the public dataset, it will be treated as a class metric of 0. One of the implications of this is that private/public shakeup could be huge.", "414480": "I agree with Alexander, I believe that the public LB may not be fully representative of the entire test set. My best submission was showing 0.87 locally.", "418860": "Brian, you have got 0.87 for local training or local validation set?", "418926": "Local validation set. Thinking back a week, it may have been due to a small validation set. With my latest models that still need way more training, locally .75 validation results a .373 public. Definitely needs more time...", "420356": "I guess it will be lower than 0.65" }, "source": "meta" }