![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1578933%2Ff6e84d2a4ea5be1eed73e6968a7a974f%2Fbelka.png?generation=1718407997709937&alt=media\")
I think it is an unnecessary check…
", "messages": [ { "id": "2872610", "postDate": "06/14/2024 23:40:55", "content": "
I think it is an unnecessary check…
", "rawMarkdown": "\n\nI think it is an unnecessary check...", "votes": null }, { "id": "2874135", "postDate": "06/16/2024 04:28:35", "content": "Oh, I have the same problem now.
\nIs your Notebook original code or copy any Notebooks?
I had this problem too, but it was caused by not setting index = false for the .to_csv argument.
", "rawMarkdown": "I had this problem too, but it was caused by not setting index = false for the .to_csv argument.", "votes": null }, { "id": "2874693", "postDate": "06/16/2024 14:07:06", "content": "I use this line of code to constrain my predictions to the range of 0 to 1 (for polars DataFrames). It's especially important for ensembling to prevent unintentionally overweighting the models that predict outside of this range.
\nsubmission = submission.with_columns(binds = submission['binds'].clip(0,1))
@sorkun - Did Kaggle count that as one of your five daily submissions? My experience is that most competitions don't debit the failures from one's daily allowance.
", "rawMarkdown": "sorkun - Did Kaggle count that as one of your five daily submissions? My experience is that most competitions don't debit the failures from one's daily allowance.", "votes": null }, { "id": "2877472", "postDate": "06/18/2024 11:42:43", "content": "I don't have much comp experience but what I see is that it doesn't debit if your kernel fails during the submission, but if there is an error during the evaluation it debits.
\nHowever, my frustration was related to the range check. The overall evaluation of the average precision depends on the order of these predictions, not their exact values.
", "rawMarkdown": "I don't have much comp experience but what I see is that it doesn't debit if your kernel fails during the submission, but if there is an error during the evaluation it debits.\n\nHowever, my frustration was related to the range check. The overall evaluation of the average precision depends on the order of these predictions, not their exact values.", "votes": null }, { "id": "2877692", "postDate": "06/18/2024 14:20:22", "content": "One could obtain raw predictions outside the range either by a quirk of the weights (it has happened to me when using 101% model_A - 1% model_B, for example), in which case I understand the frustration. But such an outcome can also indicate that something has gone more fundamentally wrong with the calculation, which has spat out numbers nothing like probabilities. In this case, I actually want the submission to fail to alert me if I was so careless as to press Submit without looking carefully at the output.
\nif you want to avoid the error, then it seems reasonable to fit the predictions to the range [0, 1] by first subtracting the smallest value from everything and then dividing all the results by the maximum. I personally prefer this to np.clipping, since the fitting avoids artificial ties (with clipping, everything predicted at or above 1.0 becomes equal, and similarly everything at or below 0.0).
", "rawMarkdown": "One could obtain raw predictions outside the range either by a quirk of the weights (it has happened to me when using 101% model_A - 1% model_B, for example), in which case I understand the frustration. But such an outcome can also indicate that something has gone more fundamentally wrong with the calculation, which has spat out numbers nothing like probabilities. In this case, I actually want the submission to fail to alert me if I was so careless as to press Submit without looking carefully at the output. \n\nif you want to avoid the error, then it seems reasonable to fit the predictions to the range [0, 1] by first subtracting the smallest value from everything and then dividing all the results by the maximum. I personally prefer this to np.clipping, since the fitting avoids artificial ties (with clipping, everything predicted at or above 1.0 becomes equal, and similarly everything at or below 0.0).", "votes": null } ], "comments": [ { "id": 2874135, "author_name": "megumih", "author_url": "", "post_date": "06/16/2024 04:28:35", "content": "Oh, I have the same problem now.
\nIs your Notebook original code or copy any Notebooks?
I had this problem too, but it was caused by not setting index = false for the .to_csv argument.
", "votes": null, "replies": [] }, { "id": 2874693, "author_name": "stevenhewitt", "author_url": "", "post_date": "06/16/2024 14:07:06", "content": "I use this line of code to constrain my predictions to the range of 0 to 1 (for polars DataFrames). It's especially important for ensembling to prevent unintentionally overweighting the models that predict outside of this range.
\nsubmission = submission.with_columns(binds = submission['binds'].clip(0,1))
@sorkun - Did Kaggle count that as one of your five daily submissions? My experience is that most competitions don't debit the failures from one's daily allowance.
", "votes": null, "replies": [ { "id": 2877472, "author_name": "sorkun", "author_url": "", "post_date": "06/18/2024 11:42:43", "content": "I don't have much comp experience but what I see is that it doesn't debit if your kernel fails during the submission, but if there is an error during the evaluation it debits.
\nHowever, my frustration was related to the range check. The overall evaluation of the average precision depends on the order of these predictions, not their exact values.
", "votes": null, "replies": [ { "id": 2877692, "author_name": "jbomitchell", "author_url": "", "post_date": "06/18/2024 14:20:22", "content": "One could obtain raw predictions outside the range either by a quirk of the weights (it has happened to me when using 101% model_A - 1% model_B, for example), in which case I understand the frustration. But such an outcome can also indicate that something has gone more fundamentally wrong with the calculation, which has spat out numbers nothing like probabilities. In this case, I actually want the submission to fail to alert me if I was so careless as to press Submit without looking carefully at the output.
\nif you want to avoid the error, then it seems reasonable to fit the predictions to the range [0, 1] by first subtracting the smallest value from everything and then dividing all the results by the maximum. I personally prefer this to np.clipping, since the fitting avoids artificial ties (with clipping, everything predicted at or above 1.0 becomes equal, and similarly everything at or below 0.0).
", "votes": null, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2872610": "\n\nI think it is an unnecessary check...", "2874135": "Oh, I have the same problem now.\nIs your Notebook original code or copy any Notebooks?", "2874242": "I had this problem too, but it was caused by not setting index = false for the .to_csv argument.", "2874693": "I use this line of code to constrain my predictions to the range of 0 to 1 (for polars DataFrames). It's especially important for ensembling to prevent unintentionally overweighting the models that predict outside of this range.\n`submission = submission.with_columns(binds = submission['binds'].clip(0,1))`", "2877415": "sorkun - Did Kaggle count that as one of your five daily submissions? My experience is that most competitions don't debit the failures from one's daily allowance.", "2877472": "I don't have much comp experience but what I see is that it doesn't debit if your kernel fails during the submission, but if there is an error during the evaluation it debits.\n\nHowever, my frustration was related to the range check. The overall evaluation of the average precision depends on the order of these predictions, not their exact values.", "2877692": "One could obtain raw predictions outside the range either by a quirk of the weights (it has happened to me when using 101% model_A - 1% model_B, for example), in which case I understand the frustration. But such an outcome can also indicate that something has gone more fundamentally wrong with the calculation, which has spat out numbers nothing like probabilities. In this case, I actually want the submission to fail to alert me if I was so careless as to press Submit without looking carefully at the output. \n\nif you want to avoid the error, then it seems reasonable to fit the predictions to the range [0, 1] by first subtracting the smallest value from everything and then dividing all the results by the maximum. I personally prefer this to np.clipping, since the fitting avoids artificial ties (with clipping, everything predicted at or above 1.0 becomes equal, and similarly everything at or below 0.0)." }, "source": "meta" }