I noticed that one of my private submissions topped the submission of the first price on the private leaderboard with score 0.9516. Just out of curiosity, how many of those submissions are there that have outperformed the winner of the competition.
\nObviously this submission with 0.9516 was not taken into account for the final rankings in my case and I understand the rules, so I am not disputing anything here. I fully accept the results and congratulate the winner.
\nI am just curious about other participants who outperformed the winner and are not visible.
I noticed that one of my private submissions topped the submission of the first price on the private leaderboard with score 0.9516. Just out of curiosity, how many of those submissions are there that have outperformed the winner of the competition.
\nObviously this submission with 0.9516 was not taken into account for the final rankings in my case and I understand the rules, so I am not disputing anything here. I fully accept the results and congratulate the winner.
\nI am just curious about other participants who outperformed the winner and are not visible.
I had a private score about 0.9499 which is the higher than the 5th place. Nobody can precisely submit the high Private Score and the only thing we can do is to train a model that fit very well not only in Public Score, but also in CV and Private Score so that we can submit our model with confident.
", "rawMarkdown": "I had a private score about 0.9499 which is the higher than the 5th place. Nobody can precisely submit the high Private Score and the only thing we can do is to train a model that fit very well not only in Public Score, but also in CV and Private Score so that we can submit our model with confident.", "votes": null }, { "id": "1314040", "postDate": "05/18/2021 22:45:33", "content": "It would be nice to describe the approach of that solution
", "rawMarkdown": "It would be nice to describe the approach of that solution", "votes": null }, { "id": "1314488", "postDate": "05/19/2021 07:32:53", "content": "You are right.
\nHere it is:
\nThis submission topped the first place solution on the private leaderboard (score: 0.9516), even though it fared quite modestly on the public leaderboard (score: 0.9166).
\nThe approach taken in the inference notebook is:
\namazing score for single model! No pseudo or hand-labelling I assume
\nFrom curiosity, what was your batch size for 1536 x 768, approx time/epoch, and how many epochs it is trained ?
You are right; we did not use any pseudo- or hand-labelling.
\nThe batch size was 12 using four RTX6000 GPUs (https://cloud.jarvislabs.ai/). Also trained for 18 epochs and the time per epoch was around six minutes.
", "rawMarkdown": "You are right; we did not use any pseudo- or hand-labelling.\n\nThe batch size was 12 using four RTX6000 GPUs (https://cloud.jarvislabs.ai/). Also trained for 18 epochs and the time per epoch was around six minutes.", "votes": null }, { "id": "1314718", "postDate": "05/19/2021 10:39:26", "content": "OK, with 4 GPUs all make sense :))) Thanks for your replies!
", "rawMarkdown": "OK, with 4 GPUs all make sense :))) Thanks for your replies!", "votes": null } ], "comments": [ { "id": 1307239, "author_name": "luciusk", "author_url": "", "post_date": "05/14/2021 10:32:09", "content": "I had a private score about 0.9499 which is the higher than the 5th place. Nobody can precisely submit the high Private Score and the only thing we can do is to train a model that fit very well not only in Public Score, but also in CV and Private Score so that we can submit our model with confident.
", "votes": null, "replies": [] }, { "id": 1314040, "author_name": "imeintanis", "author_url": "", "post_date": "05/18/2021 22:45:33", "content": "It would be nice to describe the approach of that solution
", "votes": null, "replies": [ { "id": 1314488, "author_name": "gilfernandes", "author_url": "", "post_date": "05/19/2021 07:32:53", "content": "You are right.
\nHere it is:
\nThis submission topped the first place solution on the private leaderboard (score: 0.9516), even though it fared quite modestly on the public leaderboard (score: 0.9166).
\nThe approach taken in the inference notebook is:
\namazing score for single model! No pseudo or hand-labelling I assume
\nFrom curiosity, what was your batch size for 1536 x 768, approx time/epoch, and how many epochs it is trained ?
You are right; we did not use any pseudo- or hand-labelling.
\nThe batch size was 12 using four RTX6000 GPUs (https://cloud.jarvislabs.ai/). Also trained for 18 epochs and the time per epoch was around six minutes.
", "votes": null, "replies": [] }, { "id": 1314718, "author_name": "imeintanis", "author_url": "", "post_date": "05/19/2021 10:39:26", "content": "OK, with 4 GPUs all make sense :))) Thanks for your replies!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1306952": "I noticed that one of my private submissions topped the submission of the first price on the private leaderboard with score 0.9516. Just out of curiosity, how many of those submissions are there that have outperformed the winner of the competition.\nObviously this submission with 0.9516 was not taken into account for the final rankings in my case and I understand the rules, so I am not disputing anything here. I fully accept the results and congratulate the winner.\nI am just curious about other participants who outperformed the winner and are not visible.", "1307239": "I had a private score about 0.9499 which is the higher than the 5th place. Nobody can precisely submit the high Private Score and the only thing we can do is to train a model that fit very well not only in Public Score, but also in CV and Private Score so that we can submit our model with confident.", "1314040": "It would be nice to describe the approach of that solution", "1314488": "You are right.\n\nHere it is:\n\nThis submission topped the first place solution on the private leaderboard (score: *0.9516*), even though it fared quite modestly on the public leaderboard (score: *0.9166*).\n\nThe approach taken in the inference notebook is:\n\n1. Use one single model trained with FPN and efficientnet-b7 back-end (Pytorch, using [segmentation models Pytorch](https://github.com/qubvel/segmentation_models.pytorch)), and with a window size of *1536* window size and *768* tile size. \n2. Perform inference on three grids, all with window size of *1536* window size and *768* tile size, but different overlaps with these sizes: *[32, 128, 256]*\n3. Take the predicted sets of masks and check if its average is above *0.49* and produce with that the final prediction set of masks.", "1314604": "amazing score for single model! No pseudo or hand-labelling I assume\nFrom curiosity, what was your batch size for 1536 x 768, approx time/epoch, and how many epochs it is trained ?", "1314663": "You are right; we did not use any pseudo- or hand-labelling.\n\nThe batch size was 12 using four RTX6000 GPUs (https://cloud.jarvislabs.ai/). Also trained for 18 epochs and the time per epoch was around six minutes.", "1314718": "OK, with 4 GPUs all make sense :))) Thanks for your replies!" }, "source": "meta" }