{ "id": 216247, "title": "Which are the best loss functions to use other than cross entropy loss ?", "url": "/competitions/cassava-leaf-disease-classification/discussion/216247", "author_name": "", "post_date": "2021-02-02T06:05:23.186234800Z", "votes": 2, "comment_count": 5, "views": 0, "content": "

Hello all,

\n

What kind of loss function (other than cross-entropy loss) should I try in this type of competition which are known for better results?
\nIt will be helpful to explore different loss function which can solve the class imbalance.
\nAny suggestions?

", "messages": [ { "id": "1181802", "postDate": "02/02/2021 06:05:23", "content": "

Hello all,

\n

What kind of loss function (other than cross-entropy loss) should I try in this type of competition which are known for better results?
\nIt will be helpful to explore different loss function which can solve the class imbalance.
\nAny suggestions?

", "rawMarkdown": "Hello all,\n\nWhat kind of loss function (other than cross-entropy loss) should I try in this type of competition which are known for better results?\nIt will be helpful to explore different loss function which can solve the class imbalance. \nAny suggestions?", "votes": null }, { "id": "1181972", "postDate": "02/02/2021 08:26:46", "content": "

Categorical cross entropy is the best option I think.

", "rawMarkdown": "Categorical cross entropy is the best option I think.", "votes": null }, { "id": "1182108", "postDate": "02/02/2021 09:57:11", "content": "

Well, you can try some, just to give some directions:

\n

(I got these with typing \"imbalanced\" into the discussion forum in 5 seconds from top 10 sorted by \"Relevance\")
\nHandling imbalanced datasets: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/198842
\nLoss Functions for Multiclass imbalanced data: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/200696

\n

(another 5 seconds search result…)
\nWhich loss function is better?: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/210918
\nTaylor Cross Entropy Loss: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/209782
\nCombining Different Loss Functions: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/211475
\n[Tips] How to use Label Smoothing in Pytorch: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/203103
\nBi-Tempered Loss [Tensorflow 2.0]: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/209773
\nif you are not getting lb 0.901 and above, here is why: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/202017

\n

(after this, if you want to eat, you can go fishing -> just type short words into the search field and hit enter, it does magix…)

", "rawMarkdown": "Well, you can try some, just to give some directions:\n\n(I got these with typing \"imbalanced\" into the discussion forum in 5 seconds from top 10 sorted by \"Relevance\")\nHandling imbalanced datasets: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/198842\nLoss Functions for Multiclass imbalanced data: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/200696\n\n(another 5 seconds search result...)\nWhich loss function is better?: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/210918\nTaylor Cross Entropy Loss: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/209782\nCombining Different Loss Functions: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/211475\n[Tips] How to use Label Smoothing in Pytorch: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/203103\nBi-Tempered Loss [Tensorflow 2.0]: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/209773\nif you are not getting lb 0.901 and above, here is why: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/202017\n\n(after this, if you want to eat, you can go fishing -> just type short words into the search field and hit enter, it does magix...)", "votes": null }, { "id": "1182425", "postDate": "02/02/2021 12:58:04", "content": "

I saw many loss in this competitions. Categorical crossentropy, of course, is basic. And I saw bi-tempered log loss, taylorcrossentropy, and symmetric cross entropy for noisy labels.

", "rawMarkdown": "I saw many loss in this competitions. Categorical crossentropy, of course, is basic. And I saw bi-tempered log loss, taylorcrossentropy, and symmetric cross entropy for noisy labels.", "votes": null }, { "id": "1189178", "postDate": "02/06/2021 19:14:37", "content": "

I used Taylor cross entropy, showing me good results 👍

", "rawMarkdown": "I used Taylor cross entropy, showing me good results 👍", "votes": null }, { "id": "1189180", "postDate": "02/06/2021 19:15:34", "content": "

Well Thank You so much I have used taylor cross entropy loss with label smoothing and showing me good results

", "rawMarkdown": "Well Thank You so much I have used taylor cross entropy loss with label smoothing and showing me good results", "votes": null } ], "comments": [ { "id": 1181972, "author_name": "shanmukh05", "author_url": "", "post_date": "02/02/2021 08:26:46", "content": "

Categorical cross entropy is the best option I think.

", "votes": null, "replies": [] }, { "id": 1182108, "author_name": "killimi", "author_url": "", "post_date": "02/02/2021 09:57:11", "content": "

Well, you can try some, just to give some directions:

\n

(I got these with typing \"imbalanced\" into the discussion forum in 5 seconds from top 10 sorted by \"Relevance\")
\nHandling imbalanced datasets: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/198842
\nLoss Functions for Multiclass imbalanced data: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/200696

\n

(another 5 seconds search result…)
\nWhich loss function is better?: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/210918
\nTaylor Cross Entropy Loss: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/209782
\nCombining Different Loss Functions: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/211475
\n[Tips] How to use Label Smoothing in Pytorch: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/203103
\nBi-Tempered Loss [Tensorflow 2.0]: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/209773
\nif you are not getting lb 0.901 and above, here is why: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/202017

\n

(after this, if you want to eat, you can go fishing -> just type short words into the search field and hit enter, it does magix…)

", "votes": null, "replies": [ { "id": 1189180, "author_name": "parthdhameliya77", "author_url": "", "post_date": "02/06/2021 19:15:34", "content": "

Well Thank You so much I have used taylor cross entropy loss with label smoothing and showing me good results

", "votes": null, "replies": [] } ] }, { "id": 1182425, "author_name": "vkehfdl1", "author_url": "", "post_date": "02/02/2021 12:58:04", "content": "

I saw many loss in this competitions. Categorical crossentropy, of course, is basic. And I saw bi-tempered log loss, taylorcrossentropy, and symmetric cross entropy for noisy labels.

", "votes": null, "replies": [ { "id": 1189178, "author_name": "parthdhameliya77", "author_url": "", "post_date": "02/06/2021 19:14:37", "content": "

I used Taylor cross entropy, showing me good results 👍

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1181802": "Hello all,\n\nWhat kind of loss function (other than cross-entropy loss) should I try in this type of competition which are known for better results?\nIt will be helpful to explore different loss function which can solve the class imbalance. \nAny suggestions?", "1181972": "Categorical cross entropy is the best option I think.", "1182108": "Well, you can try some, just to give some directions:\n\n(I got these with typing \"imbalanced\" into the discussion forum in 5 seconds from top 10 sorted by \"Relevance\")\nHandling imbalanced datasets: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/198842\nLoss Functions for Multiclass imbalanced data: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/200696\n\n(another 5 seconds search result...)\nWhich loss function is better?: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/210918\nTaylor Cross Entropy Loss: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/209782\nCombining Different Loss Functions: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/211475\n[Tips] How to use Label Smoothing in Pytorch: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/203103\nBi-Tempered Loss [Tensorflow 2.0]: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/209773\nif you are not getting lb 0.901 and above, here is why: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/202017\n\n(after this, if you want to eat, you can go fishing -> just type short words into the search field and hit enter, it does magix...)", "1182425": "I saw many loss in this competitions. Categorical crossentropy, of course, is basic. And I saw bi-tempered log loss, taylorcrossentropy, and symmetric cross entropy for noisy labels.", "1189178": "I used Taylor cross entropy, showing me good results 👍", "1189180": "Well Thank You so much I have used taylor cross entropy loss with label smoothing and showing me good results" }, "source": "meta" }