Instead of creating 1 single model.\nI thought of creating 3 different models for consonant ,grapheme,vowel.\nI made the model then i made the prediction.\nFirst the score was very low 0.016.\nThen I thought of comparing my result to a high score visible submission of some public kernel.\nI found that my and \nthe his score\ndiffer only in three rows one in vowel , one in grapheme and one in consonant.\nall rest 33 are identical.\nI can't understand the reason of my low score.\nI know grpaheme roots have double weights but mine only one is wrong.\nPlease can any one explain.\nI am new to image classification.
", "messages": [ { "id": "763659", "postDate": "03/04/2020 17:58:39", "content": "Instead of creating 1 single model.\nI thought of creating 3 different models for consonant ,grapheme,vowel.\nI made the model then i made the prediction.\nFirst the score was very low 0.016.\nThen I thought of comparing my result to a high score visible submission of some public kernel.\nI found that my and \nthe his score\ndiffer only in three rows one in vowel , one in grapheme and one in consonant.\nall rest 33 are identical.\nI can't understand the reason of my low score.\nI know grpaheme roots have double weights but mine only one is wrong.\nPlease can any one explain.\nI am new to image classification.
", "rawMarkdown": "Instead of creating 1 single model.\nI thought of creating 3 different models for consonant ,grapheme,vowel.\nI made the model then i made the prediction.\nFirst the score was very low **0.016**.\nThen I thought of comparing my result to a high score visible submission of some public kernel.\nI found that [**my**](https://www.kaggle.com/ianmoone0617/grapheme) and \nthe [**his score**](https://www.kaggle.com/marcelosanchezortega/version1-0-9696)\ndiffer only in **three rows** one in vowel , one in grapheme and one in consonant.\nall rest 33 are identical.\nI can't understand the reason of my low score.\nI know grpaheme roots have double weights but mine only one is wrong.\nPlease can any one explain.\nI am new to image classification.", "votes": null }, { "id": "763684", "postDate": "03/04/2020 18:32:05", "content": "Those 36 samples are the pubicly available test set. the leaderboard is based on 100000+ (don't know the exact number). Your score is probably low because of a bug in your inference kernel. Try predicting (committing) the train set. You will find the bug.
", "rawMarkdown": "Those 36 samples are the pubicly available test set. the leaderboard is based on 100000+ (don't know the exact number). Your score is probably low because of a bug in your inference kernel. Try predicting (committing) the **train** set. You will find the bug.", "votes": null }, { "id": "763717", "postDate": "03/04/2020 19:26:24", "content": "I understand there is private test set, but i can't understand the reason for low public set score
", "rawMarkdown": "I understand there is private test set, but i can't understand the reason for low public set score", "votes": null }, { "id": "763740", "postDate": "03/04/2020 20:01:10", "content": "In my case, I made a bug, but the kernel finished the commit/submit successfully. The problem was that I loaded the sample_submission.csv, but because of the bug, the script did not update the target values. I saved and submit the default target values.
@ianmoone0617 you are limiting your predictions to the 12 publicly available test images in cell 9 and 10 (range(12)). \nThe hidden test-parquets contain 1000000+ as stated below.
\n", "rawMarkdown": "> low public set score \nIt is computed from private dataset that you cannot see.", "votes": null }, { "id": "763860", "postDate": "03/04/2020 22:58:12", "content": "low public set score \n It is computed from private dataset that you cannot see.
\n
Did you do cross-validation? What're your train and validation set recall? It seems to me your models weren't trained well. The cv and lb won't be so different.
", "rawMarkdown": "Did you do cross-validation? What're your train and validation set recall? It seems to me your models weren't trained well. The cv and lb won't be so different.", "votes": null } ], "comments": [ { "id": 763684, "author_name": "pestipeti", "author_url": "", "post_date": "03/04/2020 18:32:05", "content": "Those 36 samples are the pubicly available test set. the leaderboard is based on 100000+ (don't know the exact number). Your score is probably low because of a bug in your inference kernel. Try predicting (committing) the train set. You will find the bug.
", "votes": null, "replies": [ { "id": 763717, "author_name": "", "author_url": "", "post_date": "03/04/2020 19:26:24", "content": "I understand there is private test set, but i can't understand the reason for low public set score
", "votes": null, "replies": [] }, { "id": 763740, "author_name": "pestipeti", "author_url": "", "post_date": "03/04/2020 20:01:10", "content": "In my case, I made a bug, but the kernel finished the commit/submit successfully. The problem was that I loaded the sample_submission.csv, but because of the bug, the script did not update the target values. I saved and submit the default target values.
\n", "votes": null, "replies": [] } ] }, { "id": 763833, "author_name": "joatom", "author_url": "", "post_date": "03/04/2020 22:24:33", "content": "low public set score \n It is computed from private dataset that you cannot see.
\n
@ianmoone0617 you are limiting your predictions to the 12 publicly available test images in cell 9 and 10 (range(12)). \nThe hidden test-parquets contain 1000000+ as stated below.
Did you do cross-validation? What're your train and validation set recall? It seems to me your models weren't trained well. The cv and lb won't be so different.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "763659": "Instead of creating 1 single model.\nI thought of creating 3 different models for consonant ,grapheme,vowel.\nI made the model then i made the prediction.\nFirst the score was very low **0.016**.\nThen I thought of comparing my result to a high score visible submission of some public kernel.\nI found that [**my**](https://www.kaggle.com/ianmoone0617/grapheme) and \nthe [**his score**](https://www.kaggle.com/marcelosanchezortega/version1-0-9696)\ndiffer only in **three rows** one in vowel , one in grapheme and one in consonant.\nall rest 33 are identical.\nI can't understand the reason of my low score.\nI know grpaheme roots have double weights but mine only one is wrong.\nPlease can any one explain.\nI am new to image classification.", "763684": "Those 36 samples are the pubicly available test set. the leaderboard is based on 100000+ (don't know the exact number). Your score is probably low because of a bug in your inference kernel. Try predicting (committing) the **train** set. You will find the bug.", "763717": "I understand there is private test set, but i can't understand the reason for low public set score", "763740": "In my case, I made a bug, but the kernel finished the commit/submit successfully. The problem was that I loaded the `sample_submission.csv`, but because of the bug, the script did not update the `target` values. I saved and submit the default `target` values.", "763833": "ianmoone0617 you are limiting your predictions to the 12 publicly available test images in cell 9 and 10 (`range(12)`). \nThe hidden test-parquets contain 1000000+ as stated below.", "763848": "> low public set score \nIt is computed from private dataset that you cannot see.", "763860": "Did you do cross-validation? What're your train and validation set recall? It seems to me your models weren't trained well. The cv and lb won't be so different." }, "source": "meta" }