I'm using GroupKFold with 5 folds grouping by planet_id and for my best model I get 0.415 CV vs 0.346 LB. What about you?
", "messages": [ { "id": 3258870, "postDate": "2025-07-31T11:18:30.947Z", "content": "I'm using GroupKFold with 5 folds grouping by planet_id and for my best model I get 0.415 CV vs 0.346 LB. What about you?
", "rawMarkdown": "I'm using GroupKFold with 5 folds grouping by planet_id and for my best model I get 0.415 CV vs 0.346 LB. What about you?", "votes": 2 }, { "id": 3267007, "postDate": "2025-08-10T07:12:55.493Z", "content": "May I ask for your input and output shape of your model? BTW got 0.52 CV and 0 LB currently.
", "rawMarkdown": "May I ask for your input and output shape of your model? BTW got 0.52 CV and 0 LB currently.", "replies": [ { "id": 3267641, "postDate": "2025-08-11T15:43:45.013Z", "content": "Input shape is 32, output shape is 16. I approximate 283 target values with a 15-degree polynomial and try to learn its coeffs with a linear model. And what are you using to reach 0.52 CV?
", "rawMarkdown": "Input shape is 32, output shape is 16. I approximate 283 target values with a 15-degree polynomial and try to learn its coeffs with a linear model. And what are you using to reach 0.52 CV?", "votes": 1, "replies": [ { "id": 3267673, "postDate": "2025-08-11T16:28:58.660Z", "content": "I use the 2D CNN method; maybe I am overfitting, or I did something wrong in the data preprocessing. Currently I am using the official data preprocess pipeline. Is this the possible reason?
", "rawMarkdown": "I use the 2D CNN method; maybe I am overfitting, or I did something wrong in the data preprocessing. Currently I am using the official data preprocess pipeline. Is this the possible reason?" }, { "id": 3267676, "postDate": "2025-08-11T16:35:04.443Z", "content": "Do you clip the output? I use predictions.clip(0.003, 0.1)
", "rawMarkdown": "Do you clip the output? I use predictions.clip(0.003, 0.1)" }, { "id": 3268456, "postDate": "2025-08-13T01:26:14.343Z", "content": "I didn't do this; what's the reason for clip(0.003,0.1)? I mean, selecting a too strict threshold may somehow overfit to the lb.
", "rawMarkdown": "I didn't do this; what's the reason for clip(0.003,0.1)? I mean, selecting a too strict threshold may somehow overfit to the lb." }, { "id": 3269935, "postDate": "2025-08-15T11:05:06.587Z", "content": "I selected this thresholds by min/max of target in the training dataset. When I used gradient boosting models there was no problems. But when I switched to the linear model my score also dropped. And this prediction clipping helped. It is not optimal, to improve it we can select thresholds by-wavelength.
\nAs I see your LB score improved, has the clipping helped?
", "rawMarkdown": "I selected this thresholds by min/max of target in the training dataset. When I used gradient boosting models there was no problems. But when I switched to the linear model my score also dropped. And this prediction clipping helped. It is not optimal, to improve it we can select thresholds by-wavelength. \n\nAs I see your LB score improved, has the clipping helped?", "votes": 1 }, { "id": 3270237, "postDate": "2025-08-16T03:53:14.253Z", "content": "No, my model still scores 0; I just mimic some public kernel method to see if my data preprocessing works. It turns out the NN model may be overfitting, though I got stable and high cv locally. I have no experience in aerial data, and it seems my model design is irrational.
", "rawMarkdown": "No, my model still scores 0; I just mimic some public kernel method to see if my data preprocessing works. It turns out the NN model may be overfitting, though I got stable and high cv locally. I have no experience in aerial data, and it seems my model design is irrational." }, { "id": 3270252, "postDate": "2025-08-16T04:55:55.917Z", "content": "I too am getting 0 LB, i didn't use the official metric to train, that might alse be causing the problem, but still, performs very poorly, might have to change a few things.
", "rawMarkdown": "I too am getting 0 LB, i didn't use the official metric to train, that might alse be causing the problem, but still, performs very poorly, might have to change a few things.", "votes": 1 } ] } ] }, { "id": 3261982, "postDate": "2025-08-02T16:43:30.860Z", "content": "CV=0.397 LB=0.389
", "rawMarkdown": "CV=0.397 LB=0.389" }, { "id": 3258972, "postDate": "2025-07-31T16:02:06.727Z", "content": "CV/LB 0.39/0.35
", "rawMarkdown": "CV/LB 0.39/0.35" }, { "id": 3258946, "postDate": "2025-07-31T14:47:08.267Z", "content": "CV/LB: 0.332/0.334
", "rawMarkdown": "CV/LB: 0.332/0.334" }, { "id": 3258899, "postDate": "2025-07-31T12:41:59.253Z", "content": "CV 0.3805 LB 0.342
", "rawMarkdown": "CV 0.3805 LB 0.342" } ], "comments": [ { "id": 3267007, "author_name": "shanzhong8", "author_url": "", "post_date": "2025-08-10T07:12:55.493000", "content": "May I ask for your input and output shape of your model? BTW got 0.52 CV and 0 LB currently.
", "votes": 0, "replies": [ { "id": 3267641, "author_name": "Artem Veshkin", "author_url": "", "post_date": "2025-08-11T15:43:45.013000", "content": "Input shape is 32, output shape is 16. I approximate 283 target values with a 15-degree polynomial and try to learn its coeffs with a linear model. And what are you using to reach 0.52 CV?
", "votes": 1, "replies": [ { "id": 3267673, "author_name": "shanzhong8", "author_url": "", "post_date": "2025-08-11T16:28:58.660000", "content": "I use the 2D CNN method; maybe I am overfitting, or I did something wrong in the data preprocessing. Currently I am using the official data preprocess pipeline. Is this the possible reason?
", "votes": 0, "replies": [] }, { "id": 3267676, "author_name": "Artem Veshkin", "author_url": "", "post_date": "2025-08-11T16:35:04.443000", "content": "Do you clip the output? I use predictions.clip(0.003, 0.1)
", "votes": 0, "replies": [] }, { "id": 3268456, "author_name": "shanzhong8", "author_url": "", "post_date": "2025-08-13T01:26:14.343000", "content": "I didn't do this; what's the reason for clip(0.003,0.1)? I mean, selecting a too strict threshold may somehow overfit to the lb.
", "votes": 0, "replies": [] }, { "id": 3269935, "author_name": "Artem Veshkin", "author_url": "", "post_date": "2025-08-15T11:05:06.587000", "content": "I selected this thresholds by min/max of target in the training dataset. When I used gradient boosting models there was no problems. But when I switched to the linear model my score also dropped. And this prediction clipping helped. It is not optimal, to improve it we can select thresholds by-wavelength.
\nAs I see your LB score improved, has the clipping helped?
", "votes": 1, "replies": [] }, { "id": 3270237, "author_name": "shanzhong8", "author_url": "", "post_date": "2025-08-16T03:53:14.253000", "content": "No, my model still scores 0; I just mimic some public kernel method to see if my data preprocessing works. It turns out the NN model may be overfitting, though I got stable and high cv locally. I have no experience in aerial data, and it seems my model design is irrational.
", "votes": 0, "replies": [] }, { "id": 3270252, "author_name": "Manas Choudhary", "author_url": "", "post_date": "2025-08-16T04:55:55.917000", "content": "I too am getting 0 LB, i didn't use the official metric to train, that might alse be causing the problem, but still, performs very poorly, might have to change a few things.
", "votes": 1, "replies": [] } ] } ] }, { "id": 3261982, "author_name": "Carlos Pérez Ricardo", "author_url": "", "post_date": "2025-08-02T16:43:30.860000", "content": "CV=0.397 LB=0.389
", "votes": 0, "replies": [] }, { "id": 3258972, "author_name": "eugene", "author_url": "", "post_date": "2025-07-31T16:02:06.727000", "content": "CV/LB 0.39/0.35
", "votes": 0, "replies": [] }, { "id": 3258946, "author_name": "Oleh Kivernyk", "author_url": "", "post_date": "2025-07-31T14:47:08.267000", "content": "CV/LB: 0.332/0.334
", "votes": 0, "replies": [] }, { "id": 3258899, "author_name": "Sangram Patil", "author_url": "", "post_date": "2025-07-31T12:41:59.253000", "content": "CV 0.3805 LB 0.342
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "3258870": "I'm using GroupKFold with 5 folds grouping by planet_id and for my best model I get 0.415 CV vs 0.346 LB. What about you?", "3267007": "May I ask for your input and output shape of your model? BTW got 0.52 CV and 0 LB currently.", "3261982": "CV=0.397 LB=0.389", "3258972": "CV/LB 0.39/0.35", "3258946": "CV/LB: 0.332/0.334", "3258899": "CV 0.3805 LB 0.342" } }