The training and test sets are different in this competition, and the only way to evaluate models that I see is public LB. Do you have other ideas for validation? What can we do to avoid overfitting?
", "messages": [ { "id": 1188304, "postDate": "2021-02-06T05:57:56.477Z", "content": "The training and test sets are different in this competition, and the only way to evaluate models that I see is public LB. Do you have other ideas for validation? What can we do to avoid overfitting?
", "rawMarkdown": "The training and test sets are different in this competition, and the only way to evaluate models that I see is public LB. Do you have other ideas for validation? What can we do to avoid overfitting? ", "votes": 7 }, { "id": 1188750, "postDate": "2021-02-06T13:43:31.780Z", "content": "What about stratified cross validation?
", "rawMarkdown": "What about **stratified cross validation**?", "votes": -1, "replies": [ { "id": 1189122, "postDate": "2021-02-06T18:20:21.533Z", "content": "Can you elaborate?
", "rawMarkdown": "Can you elaborate?" }, { "id": 1189133, "postDate": "2021-02-06T18:30:19.200Z", "content": "Check it out Workflow of the competition
", "rawMarkdown": "Check it out [Workflow of the competition](https://www.kaggle.com/c/hpa-single-cell-image-classification/discussion/217433)" }, { "id": 1189144, "postDate": "2021-02-06T18:37:54.847Z", "content": "There is no additional information given there. Can you elaborate how stratified Cross-fold validation would work for this competition?
", "rawMarkdown": "There is no additional information given there. Can you elaborate how stratified Cross-fold validation would work for this competition?" }, { "id": 1189158, "postDate": "2021-02-06T18:53:10.477Z", "content": "Cross-validation as you might know is a statistical method that serves to evaluate the generalization performance. Stratified version It's a way to balance distributions across the data.
\nP.D.: Before downvote me, if you still don't fully understand why it works here, i can elaborate more.
", "rawMarkdown": "**Cross-validation** as you might know is a statistical method that serves to evaluate the generalization performance. **Stratified** version It's a way to balance distributions across the data. \n\nP.D.: Before downvote me, if you still don't fully understand why it works here, i can elaborate more." }, { "id": 1189287, "postDate": "2021-02-06T21:51:44.467Z", "content": "This would work generally (for other competitions). But as there are no ground-truth masks available for either of the provided datasets (training or test), I don't believe that cross-validation is possible. Let alone a stratified version.
\nI asked for you to elaborate so that you might explain yourself and how it relates to this particular competition. I downvoted you because you didn't elaborate and still haven't. You also provided a link to a post that gives no new information or clarity. If you elaborate on cross-validation methodology for this competition, I'll make sure to upvote. I would be really interested to learn how this is possible!
", "rawMarkdown": "This would work generally (for other competitions). But as there are no ground-truth masks available for either of the provided datasets (training or test), I don't believe that cross-validation is possible. Let alone a stratified version.\n\n---\n\nI asked for you to elaborate so that you might explain yourself and how it relates to this particular competition. **I downvoted you because you didn't elaborate and still haven't. You also provided a link to a post that gives no new information or clarity. If you elaborate on cross-validation methodology for this competition, I'll make sure to upvote. I would be really interested to learn how this is possible!**", "votes": 2 }, { "id": 1189303, "postDate": "2021-02-06T22:30:01.723Z", "content": "Maybe i get confused with the first HPA competition, because it was a multi-label classification task. This is both, segmentation and multi-label classfication then, as you said it would be very hard to do cross-validation. I don't want to say that it's impossible because there are a lot skilled people that could realize a very complicated way to do that achivement.
\nP.S.: This is what i want it, feedback, thank you. Downvotes doesn't matter even upvotes, i never ask for it, the most important thing in Kaggle for my is knowledge.
", "rawMarkdown": "Maybe i get confused with the first HPA competition, because it was a multi-label classification task. This is both, segmentation and multi-label classfication then, as you said it would be very hard to do **cross-validation**. I don't want to say that it's impossible because there are a lot skilled people that could realize a very complicated way to do that achivement. \n\n----------------------------------------------------------------------------------------------------\nP.S.: This is what i want it, feedback, thank you. Downvotes doesn't matter even upvotes, i never ask for it, the most important thing in Kaggle for my is knowledge.", "votes": 1 } ] } ], "comments": [ { "id": 1188750, "author_name": "Hiram Coria 🧬", "author_url": "", "post_date": "2021-02-06T13:43:31.780000", "content": "What about stratified cross validation?
", "votes": -1, "replies": [ { "id": 1189122, "author_name": "Darien Schettler", "author_url": "", "post_date": "2021-02-06T18:20:21.533000", "content": "Can you elaborate?
", "votes": 0, "replies": [] }, { "id": 1189133, "author_name": "Hiram Coria 🧬", "author_url": "", "post_date": "2021-02-06T18:30:19.200000", "content": "Check it out Workflow of the competition
", "votes": 0, "replies": [] }, { "id": 1189144, "author_name": "Darien Schettler", "author_url": "", "post_date": "2021-02-06T18:37:54.847000", "content": "There is no additional information given there. Can you elaborate how stratified Cross-fold validation would work for this competition?
", "votes": 0, "replies": [] }, { "id": 1189158, "author_name": "Hiram Coria 🧬", "author_url": "", "post_date": "2021-02-06T18:53:10.477000", "content": "Cross-validation as you might know is a statistical method that serves to evaluate the generalization performance. Stratified version It's a way to balance distributions across the data.
\nP.D.: Before downvote me, if you still don't fully understand why it works here, i can elaborate more.
", "votes": 0, "replies": [] }, { "id": 1189287, "author_name": "Darien Schettler", "author_url": "", "post_date": "2021-02-06T21:51:44.467000", "content": "This would work generally (for other competitions). But as there are no ground-truth masks available for either of the provided datasets (training or test), I don't believe that cross-validation is possible. Let alone a stratified version.
\nI asked for you to elaborate so that you might explain yourself and how it relates to this particular competition. I downvoted you because you didn't elaborate and still haven't. You also provided a link to a post that gives no new information or clarity. If you elaborate on cross-validation methodology for this competition, I'll make sure to upvote. I would be really interested to learn how this is possible!
", "votes": 2, "replies": [] }, { "id": 1189303, "author_name": "Hiram Coria 🧬", "author_url": "", "post_date": "2021-02-06T22:30:01.723000", "content": "Maybe i get confused with the first HPA competition, because it was a multi-label classification task. This is both, segmentation and multi-label classfication then, as you said it would be very hard to do cross-validation. I don't want to say that it's impossible because there are a lot skilled people that could realize a very complicated way to do that achivement.
\nP.S.: This is what i want it, feedback, thank you. Downvotes doesn't matter even upvotes, i never ask for it, the most important thing in Kaggle for my is knowledge.
", "votes": 1, "replies": [] } ] } ], "raw_markdown_by_id": { "1188304": "The training and test sets are different in this competition, and the only way to evaluate models that I see is public LB. Do you have other ideas for validation? What can we do to avoid overfitting? ", "1188750": "What about **stratified cross validation**?" } }