{ "id": 115081, "title": "Model overfit too soon ? ", "url": "/competitions/understanding_cloud_organization/discussion/115081", "author_name": "Nirjhar Roy", "post_date": "2019-10-31T05:49:43.722000", "votes": 6, "comment_count": 11, "views": 0, "content": "

I am experience a weird situation . Deep models are overfitting and converging too soon . Almost in 5-10 epochs . Anyone is facing similar situation ?

", "messages": [ { "id": 662126, "postDate": "2019-10-31T05:49:43.723Z", "content": "

I am experience a weird situation . Deep models are overfitting and converging too soon . Almost in 5-10 epochs . Anyone is facing similar situation ?

", "rawMarkdown": "I am experience a weird situation . Deep models are overfitting and converging too soon . Almost in 5-10 epochs . Anyone is facing similar situation ? ", "votes": 6 }, { "id": 668982, "postDate": "2019-11-09T08:31:31.750Z", "content": "

EarlyStopping stops the training at ~10-11 epochs for me too. I have tried a lot of things but nothing works for me in this competition... seems like no matter what I do my score is stuck at the public best.

", "rawMarkdown": "EarlyStopping stops the training at ~10-11 epochs for me too. I have tried a lot of things but nothing works for me in this competition... seems like no matter what I do my score is stuck at the public best.", "votes": 1, "replies": [ { "id": 669011, "postDate": "2019-11-09T09:45:11.107Z", "content": "

Kind of same story for me as well :) . There is something small we are missing .

", "rawMarkdown": "Kind of same story for me as well :) . There is something small we are missing ." }, { "id": 669033, "postDate": "2019-11-09T11:25:25.957Z", "content": "

stops too early is not problem, 12 epochs gives me .660 lb

", "rawMarkdown": "stops too early is not problem, 12 epochs gives me .660 lb", "votes": 2 }, { "id": 669049, "postDate": "2019-11-09T11:45:46.453Z", "content": "

@niuddd Completely Agreed .

", "rawMarkdown": "@niuddd Completely Agreed ." }, { "id": 669051, "postDate": "2019-11-09T11:51:28.943Z", "content": "

@niuddd I see, thanks for sharing :) ... I was a little worried about my early stoppings.

", "rawMarkdown": "@niuddd I see, thanks for sharing :) ... I was a little worried about my early stoppings. " }, { "id": 669053, "postDate": "2019-11-09T11:58:52.210Z", "content": "

Though, with lower lr, heavier augmentation, different layers different lr, I can train 25 epochs, but not give significant lb improve... it depends on other things at least for me threshold optimize is still an unstable factor to overall

", "rawMarkdown": "Though, with lower lr, heavier augmentation, different layers different lr, I can train 25 epochs, but not give significant lb improve... it depends on other things at least for me threshold optimize is still an unstable factor to overall", "votes": 2 }, { "id": 669058, "postDate": "2019-11-09T12:08:02.463Z", "content": "

I also didn't gain improvement on longer epochs... So I think maybe it won't be an issue if the training stop too early.

", "rawMarkdown": "I also didn't gain improvement on longer epochs... So I think maybe it won't be an issue if the training stop too early.", "votes": 1 }, { "id": 669529, "postDate": "2019-11-10T06:11:01.923Z", "content": "

@niuddd I am seeing similar findings. Unstable thresholds seem to be more of an issue for me as I try larger models as well and longer epochs...

", "rawMarkdown": "@niuddd I am seeing similar findings. Unstable thresholds seem to be more of an issue for me as I try larger models as well and longer epochs..." } ] }, { "id": 668920, "postDate": "2019-11-09T05:01:59.807Z", "content": "

Similiar

", "rawMarkdown": "Similiar " }, { "id": 662152, "postDate": "2019-10-31T06:24:36.777Z", "content": "

I have the same problem, I kinda stuck at current score.. But I think 0.66 is reachable with more fold training and reasonable augmentation. I can get 0.658 with ensembling two single fold model which trained on BCE+Dice loss. But I don't have time to do more fold training, still tuning the loss function and augmentation..

", "rawMarkdown": "I have the same problem, I kinda stuck at current score.. But I think 0.66 is reachable with more fold training and reasonable augmentation. I can get 0.658 with ensembling two single fold model which trained on BCE+Dice loss. But I don't have time to do more fold training, still tuning the loss function and augmentation.." }, { "id": 662137, "postDate": "2019-10-31T06:02:58.193Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 668982, "author_name": "timetraveller", "author_url": "", "post_date": "2019-11-09T08:31:31.750000", "content": "

EarlyStopping stops the training at ~10-11 epochs for me too. I have tried a lot of things but nothing works for me in this competition... seems like no matter what I do my score is stuck at the public best.

", "votes": 1, "replies": [ { "id": 669011, "author_name": "Nirjhar Roy", "author_url": "", "post_date": "2019-11-09T09:45:11.107000", "content": "

Kind of same story for me as well :) . There is something small we are missing .

", "votes": 0, "replies": [] }, { "id": 669033, "author_name": "Endi Niu", "author_url": "", "post_date": "2019-11-09T11:25:25.957000", "content": "

stops too early is not problem, 12 epochs gives me .660 lb

", "votes": 2, "replies": [] }, { "id": 669049, "author_name": "Nirjhar Roy", "author_url": "", "post_date": "2019-11-09T11:45:46.453000", "content": "

@niuddd Completely Agreed .

", "votes": 0, "replies": [] }, { "id": 669051, "author_name": "timetraveller", "author_url": "", "post_date": "2019-11-09T11:51:28.943000", "content": "

@niuddd I see, thanks for sharing :) ... I was a little worried about my early stoppings.

", "votes": 0, "replies": [] }, { "id": 669053, "author_name": "Endi Niu", "author_url": "", "post_date": "2019-11-09T11:58:52.210000", "content": "

Though, with lower lr, heavier augmentation, different layers different lr, I can train 25 epochs, but not give significant lb improve... it depends on other things at least for me threshold optimize is still an unstable factor to overall

", "votes": 2, "replies": [] }, { "id": 669058, "author_name": "Tsai29", "author_url": "", "post_date": "2019-11-09T12:08:02.463000", "content": "

I also didn't gain improvement on longer epochs... So I think maybe it won't be an issue if the training stop too early.

", "votes": 1, "replies": [] }, { "id": 669529, "author_name": "Thomas Yokota", "author_url": "", "post_date": "2019-11-10T06:11:01.923000", "content": "

@niuddd I am seeing similar findings. Unstable thresholds seem to be more of an issue for me as I try larger models as well and longer epochs...

", "votes": 0, "replies": [] } ] }, { "id": 668920, "author_name": "哈尔的移动城堡", "author_url": "", "post_date": "2019-11-09T05:01:59.807000", "content": "

Similiar

", "votes": 0, "replies": [] }, { "id": 662152, "author_name": "Tsai29", "author_url": "", "post_date": "2019-10-31T06:24:36.777000", "content": "

I have the same problem, I kinda stuck at current score.. But I think 0.66 is reachable with more fold training and reasonable augmentation. I can get 0.658 with ensembling two single fold model which trained on BCE+Dice loss. But I don't have time to do more fold training, still tuning the loss function and augmentation..

", "votes": 0, "replies": [] }, { "id": 662137, "author_name": "", "author_url": "", "post_date": "2019-10-31T06:02:58.193000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "662126": "I am experience a weird situation . Deep models are overfitting and converging too soon . Almost in 5-10 epochs . Anyone is facing similar situation ? ", "668982": "EarlyStopping stops the training at ~10-11 epochs for me too. I have tried a lot of things but nothing works for me in this competition... seems like no matter what I do my score is stuck at the public best.", "668920": "Similiar ", "662152": "I have the same problem, I kinda stuck at current score.. But I think 0.66 is reachable with more fold training and reasonable augmentation. I can get 0.658 with ensembling two single fold model which trained on BCE+Dice loss. But I don't have time to do more fold training, still tuning the loss function and augmentation..", "662137": "" } }