{ "id": 340392, "title": "Metric Weights", "url": "/competitions/rsna-2022-cervical-spine-fracture-detection/discussion/340392", "author_name": "FelipeKitamura, MD, PhD", "post_date": "2022-07-28T19:21:46.467000", "votes": 65, "comment_count": 12, "views": 0, "content": "

Thanks for joining this competition. We look forward to learning from your solution.

\n

To help you build your models, we provide the weights of the metric (weighted log loss) below:

\n

Category: Weight
\nVertebrae negative: 1
\nVertebrae positive: 2
\nPatient negative: 7
\nPatient positive: 14

", "messages": [ { "id": 1875198, "postDate": "2022-07-28T19:21:46.467Z", "content": "

Thanks for joining this competition. We look forward to learning from your solution.

\n

To help you build your models, we provide the weights of the metric (weighted log loss) below:

\n

Category: Weight
\nVertebrae negative: 1
\nVertebrae positive: 2
\nPatient negative: 7
\nPatient positive: 14

", "rawMarkdown": "Thanks for joining this competition. We look forward to learning from your solution.\n\nTo help you build your models, we provide the weights of the metric (weighted log loss) below:\n\nCategory: Weight\nVertebrae negative: 1\nVertebrae positive: 2\nPatient negative: 7\nPatient positive: 14", "votes": 65 }, { "id": 1885329, "postDate": "2022-08-05T05:23:33.060Z", "content": "

Thank you @barteksadlej123 for showing the code of the metric.

\n

@harshitsheoran
\nYou mean,

\n

def competiton_loss(y_hat, y):
\n loss = loss_fn(y_hat, y)
\n weights = y * competition_weights['+'] + (1 - y) * competition_weights['-']
\n loss = (loss * weights).sum(axis=1)#.mean()

\n
return loss / weights.sum()\n
\n

don't you?

", "rawMarkdown": "Thank you @barteksadlej123 for showing the code of the metric. \n\n@harshitsheoran\nYou mean,\n\ndef competiton_loss(y_hat, y):\n loss = loss_fn(y_hat, y)\n weights = y * competition_weights['+'] + (1 - y) * competition_weights['-']\n loss = (loss * weights).sum(axis=1)#.mean()\n\n return loss / weights.sum()\n\ndon't you?", "votes": 1, "replies": [ { "id": 1885335, "postDate": "2022-08-05T05:30:02.527Z", "content": "

yes, that is correct

", "rawMarkdown": "yes, that is correct", "votes": 2 }, { "id": 1885406, "postDate": "2022-08-05T07:01:57.377Z", "content": "

Yes, I have corrected it. Sorry for mislead.

", "rawMarkdown": "Yes, I have corrected it. Sorry for mislead.", "votes": 1 } ] }, { "id": 1977126, "postDate": "2022-10-07T19:49:39.423Z", "content": "

@felipekitamura , I am a bit confused

\n

If we look at Evaluation Page we can clearly see, that w_j (weight) depends on only certain class (C_1, …, C_7, patient_overall) and does NOT depend on class label (whether label is 0 or 1)

\n

BUT from your post it is obvious that w_j is same for C_1, …, C_7 classes (but still different for patient_overall class) but depends on class label

\n

So what is True ? Should we stick to your post logic loss computation ?

\n

Thanks for reply in advance!

", "rawMarkdown": "@felipekitamura , I am a bit confused \n\nIf we look at [Evaluation Page](https://www.kaggle.com/competitions/rsna-2022-cervical-spine-fracture-detection/overview/evaluation) we can clearly see, that w_j (weight) depends on only certain class (C_1, ..., C_7, patient_overall) and does **NOT** depend on class label (whether label is 0 or 1)\n\nBUT from your post it is obvious that w_j is same for C_1, ..., C_7 classes (but still different for patient_overall class) but depends on class label\n\nSo what is True ? Should we stick to your post logic loss computation ?\n\nThanks for reply in advance!", "replies": [ { "id": 2001770, "postDate": "2022-10-24T09:50:53.750Z", "content": "

Thanks for asking. This post has the most detailed information.

", "rawMarkdown": "Thanks for asking. This post has the most detailed information.", "votes": -1 } ] }, { "id": 1875748, "postDate": "2022-07-29T09:09:43.580Z", "content": "

@maggiemd Would you please pin this discussion? Thanks.

", "rawMarkdown": "@maggiemd Would you please pin this discussion? Thanks. ", "votes": 2, "replies": [ { "id": 1876143, "postDate": "2022-07-29T15:46:19.063Z", "content": "

Good idea, done.

", "rawMarkdown": "Good idea, done.", "votes": 4 } ] }, { "id": 2007958, "postDate": "2022-10-28T15:46:53.353Z", "content": "

ohhh tenuun

", "rawMarkdown": "ohhh tenuun\n" }, { "id": 1879772, "postDate": "2022-08-01T08:52:50.747Z", "content": "

This means that weights are different in rows, right?
\nThe difinition in \"Overview - Evaluation\" is written as 'w_j' and it looks to be same weights in all rows.

", "rawMarkdown": "This means that weights are different in rows, right?\nThe difinition in \"Overview - Evaluation\" is written as 'w_j' and it looks to be same weights in all rows.", "replies": [ { "id": 1879831, "postDate": "2022-08-01T09:39:58.340Z", "content": "

You are correct, @iwatatakuya. For example, if study one has no fracture, the weights would be like:

\n

Study1_C1: 1
\nStudy1_C2: 1
\nStudy1_C3: 1
\nStudy1_C4: 1
\nStudy1_C5: 1
\nStudy1_C6: 1
\nStudy1_C7: 1
\nStudy1_patient_overall: 7

", "rawMarkdown": "You are correct, @iwatatakuya. For example, if study one has no fracture, the weights would be like:\n\nStudy1_C1: 1 \nStudy1_C2: 1\nStudy1_C3: 1\nStudy1_C4: 1\nStudy1_C5: 1\nStudy1_C6: 1\nStudy1_C7: 1\nStudy1_patient_overall: 7", "votes": 4 }, { "id": 1936382, "postDate": "2022-09-12T17:45:02.590Z", "rawMarkdown": "", "isDeleted": true } ] }, { "id": 1878739, "postDate": "2022-07-31T15:39:41.063Z", "content": "

Thanks @felipekitamura

", "rawMarkdown": "Thanks @felipekitamura ", "votes": 1 } ], "comments": [ { "id": 1885329, "author_name": "Nidaime PostDoc", "author_url": "", "post_date": "2022-08-05T05:23:33.060000", "content": "

Thank you @barteksadlej123 for showing the code of the metric.

\n

@harshitsheoran
\nYou mean,

\n

def competiton_loss(y_hat, y):
\n loss = loss_fn(y_hat, y)
\n weights = y * competition_weights['+'] + (1 - y) * competition_weights['-']
\n loss = (loss * weights).sum(axis=1)#.mean()

\n
return loss / weights.sum()\n
\n

don't you?

", "votes": 1, "replies": [ { "id": 1885335, "author_name": "Harshit Sheoran", "author_url": "", "post_date": "2022-08-05T05:30:02.527000", "content": "

yes, that is correct

", "votes": 2, "replies": [] }, { "id": 1885406, "author_name": "Bartek Sadlej", "author_url": "", "post_date": "2022-08-05T07:01:57.377000", "content": "

Yes, I have corrected it. Sorry for mislead.

", "votes": 1, "replies": [] } ] }, { "id": 1977126, "author_name": "Volodymyr", "author_url": "", "post_date": "2022-10-07T19:49:39.423000", "content": "

@felipekitamura , I am a bit confused

\n

If we look at Evaluation Page we can clearly see, that w_j (weight) depends on only certain class (C_1, …, C_7, patient_overall) and does NOT depend on class label (whether label is 0 or 1)

\n

BUT from your post it is obvious that w_j is same for C_1, …, C_7 classes (but still different for patient_overall class) but depends on class label

\n

So what is True ? Should we stick to your post logic loss computation ?

\n

Thanks for reply in advance!

", "votes": 0, "replies": [ { "id": 2001770, "author_name": "FelipeKitamura, MD, PhD", "author_url": "", "post_date": "2022-10-24T09:50:53.750000", "content": "

Thanks for asking. This post has the most detailed information.

", "votes": -1, "replies": [] } ] }, { "id": 1875748, "author_name": "Innat", "author_url": "", "post_date": "2022-07-29T09:09:43.580000", "content": "

@maggiemd Would you please pin this discussion? Thanks.

", "votes": 2, "replies": [ { "id": 1876143, "author_name": "Sohier Dane", "author_url": "", "post_date": "2022-07-29T15:46:19.063000", "content": "

Good idea, done.

", "votes": 4, "replies": [] } ] }, { "id": 2007958, "author_name": "DEEPALI SINGH", "author_url": "", "post_date": "2022-10-28T15:46:53.353000", "content": "

ohhh tenuun

", "votes": 0, "replies": [] }, { "id": 1879772, "author_name": "ganchan", "author_url": "", "post_date": "2022-08-01T08:52:50.747000", "content": "

This means that weights are different in rows, right?
\nThe difinition in \"Overview - Evaluation\" is written as 'w_j' and it looks to be same weights in all rows.

", "votes": 0, "replies": [ { "id": 1879831, "author_name": "FelipeKitamura, MD, PhD", "author_url": "", "post_date": "2022-08-01T09:39:58.340000", "content": "

You are correct, @iwatatakuya. For example, if study one has no fracture, the weights would be like:

\n

Study1_C1: 1
\nStudy1_C2: 1
\nStudy1_C3: 1
\nStudy1_C4: 1
\nStudy1_C5: 1
\nStudy1_C6: 1
\nStudy1_C7: 1
\nStudy1_patient_overall: 7

", "votes": 4, "replies": [] }, { "id": 1936382, "author_name": "", "author_url": "", "post_date": "2022-09-12T17:45:02.590000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 1878739, "author_name": "R SIVA KUMAR", "author_url": "", "post_date": "2022-07-31T15:39:41.063000", "content": "

Thanks @felipekitamura

", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "1875198": "Thanks for joining this competition. We look forward to learning from your solution.\n\nTo help you build your models, we provide the weights of the metric (weighted log loss) below:\n\nCategory: Weight\nVertebrae negative: 1\nVertebrae positive: 2\nPatient negative: 7\nPatient positive: 14", "1885329": "Thank you @barteksadlej123 for showing the code of the metric. \n\n@harshitsheoran\nYou mean,\n\ndef competiton_loss(y_hat, y):\n loss = loss_fn(y_hat, y)\n weights = y * competition_weights['+'] + (1 - y) * competition_weights['-']\n loss = (loss * weights).sum(axis=1)#.mean()\n\n return loss / weights.sum()\n\ndon't you?", "1977126": "@felipekitamura , I am a bit confused \n\nIf we look at [Evaluation Page](https://www.kaggle.com/competitions/rsna-2022-cervical-spine-fracture-detection/overview/evaluation) we can clearly see, that w_j (weight) depends on only certain class (C_1, ..., C_7, patient_overall) and does **NOT** depend on class label (whether label is 0 or 1)\n\nBUT from your post it is obvious that w_j is same for C_1, ..., C_7 classes (but still different for patient_overall class) but depends on class label\n\nSo what is True ? Should we stick to your post logic loss computation ?\n\nThanks for reply in advance!", "1875748": "@maggiemd Would you please pin this discussion? Thanks. ", "2007958": "ohhh tenuun\n", "1879772": "This means that weights are different in rows, right?\nThe difinition in \"Overview - Evaluation\" is written as 'w_j' and it looks to be same weights in all rows.", "1878739": "Thanks @felipekitamura " } }