{ "id": 215212, "title": "Best model score", "url": "/competitions/ranzcr-clip-catheter-line-classification/discussion/215212", "author_name": "Aman Deep Gupta", "post_date": "2021-01-29T05:21:19.815000", "votes": 7, "comment_count": 10, "views": 0, "content": "

Which model is providing better result?
\nI get CV 0.96 using efficient-net. Which model you guys are using and what is it's CV and LB on that?

", "messages": [ { "id": 1175353, "postDate": "2021-01-29T05:21:19.817Z", "content": "

Which model is providing better result?
\nI get CV 0.96 using efficient-net. Which model you guys are using and what is it's CV and LB on that?

", "rawMarkdown": "Which model is providing better result?\nI get CV 0.96 using efficient-net. Which model you guys are using and what is it's CV and LB on that?", "votes": 7 }, { "id": 1185848, "postDate": "2021-02-04T12:18:40.667Z", "content": "

I have a ResNet-50d (with a few custom tweeks in architecture) with CV 0.965 and LB of 0.966 (single-model + tta flip)
\nNot sure why but for me bigger models does not provide gains over r50 or r101, could be a bug in my code though

", "rawMarkdown": "I have a ResNet-50d (with a few custom tweeks in architecture) with CV 0.965 and LB of 0.966 (single-model + tta flip)\nNot sure why but for me bigger models does not provide gains over r50 or r101, could be a bug in my code though", "votes": 3, "replies": [ { "id": 1189193, "postDate": "2021-02-06T19:29:00.973Z", "content": "

It is a blessing to find a small model with so high auc. Low costs with high accuracy too.

", "rawMarkdown": "It is a blessing to find a small model with so high auc. Low costs with high accuracy too." }, { "id": 1189450, "postDate": "2021-02-07T03:12:35.833Z", "content": "

Hi,can you share more train details about you use the resnet-50d.Thanks!

", "rawMarkdown": "Hi,can you share more train details about you use the resnet-50d.Thanks!", "votes": -1 }, { "id": 1213211, "postDate": "2021-02-22T00:59:45.080Z", "rawMarkdown": "", "isDeleted": true }, { "id": 1213398, "postDate": "2021-02-22T05:08:49.627Z", "content": "

@arc144 Congrats for such a huge score with a single model. Can I ask you if the above-mentioned model trained by a Teacher-Student method with annotations or Directly trained on the dataset?

", "rawMarkdown": "@arc144 Congrats for such a huge score with a single model. Can I ask you if the above-mentioned model trained by a Teacher-Student method with annotations or Directly trained on the dataset?" } ] }, { "id": 1176615, "postDate": "2021-01-29T17:41:13.577Z", "content": "

The highest score from efficient published here was 0.959 on LB. Your eff model is very good. For ResNet200D is about 0.963 and SeResnet is around 0.962 on LB. (all single models)

", "rawMarkdown": "The highest score from efficient published here was 0.959 on LB. Your eff model is very good. For ResNet200D is about 0.963 and SeResnet is around 0.962 on LB. (all single models)", "votes": 2, "replies": [ { "id": 1183557, "postDate": "2021-02-03T04:58:25.433Z", "content": "

Sir can you share the intuitions about your model ( LB ~ 0.969 )
\nThank you 🙌🙌

", "rawMarkdown": "Sir can you share the intuitions about your model ( LB ~ 0.969 ) \nThank you 🙌🙌", "votes": 1 }, { "id": 1184722, "postDate": "2021-02-03T17:44:30.383Z", "content": "

Its all about ensembles and a lot of tries. In ensembles the models must have a high disagreement ratio, while each base model must be as much accurate as possible. In order to build base models with high disagreement ratio, you can do some of the things bellow:

\n

[1] Mix diferent models like ResNet and Eff
\n[2] Mix models trained on diferent number of epochs (e.g one model close to overfitting, and another close to bias)
\n[3] Mix diferent trained folds models, eg 5 models trained on a 5 fold cyrcle.
\n[4] Mix diferent trained seeds models (models trained on diferent initial random states)
\n[5] Mix models trained on diferent images resolutions.
\n[6] Mix models trained on diferent images filters.
\n[7] TTA (Test Time Augmentation)

", "rawMarkdown": "Its all about ensembles and a lot of tries. In ensembles the models must have a high disagreement ratio, while each base model must be as much accurate as possible. In order to build base models with high disagreement ratio, you can do some of the things bellow:\n\n[1] Mix diferent models like ResNet and Eff\n[2] Mix models trained on diferent number of epochs (e.g one model close to overfitting, and another close to bias)\n[3] Mix diferent trained folds models, eg 5 models trained on a 5 fold cyrcle.\n[4] Mix diferent trained seeds models (models trained on diferent initial random states)\n[5] Mix models trained on diferent images resolutions.\n[6] Mix models trained on diferent images filters.\n[7] TTA (Test Time Augmentation)", "votes": 17 }, { "id": 1193205, "postDate": "2021-02-09T14:08:07.693Z", "content": "

Thanks Manolis for these points.
\nIn this sense, any good resource (e.g. notebook) about ensembling models?

", "rawMarkdown": "Thanks Manolis for these points. \nIn this sense, any good resource (e.g. notebook) about ensembling models?" }, { "id": 1212548, "postDate": "2021-02-21T10:20:14.047Z", "content": "

Thanks @manolispintelas for such a detailed information. If you find something new please share here.

", "rawMarkdown": "Thanks @manolispintelas for such a detailed information. If you find something new please share here." } ] } ], "comments": [ { "id": 1185848, "author_name": "Eduardo Rocha de Andrade", "author_url": "", "post_date": "2021-02-04T12:18:40.667000", "content": "

I have a ResNet-50d (with a few custom tweeks in architecture) with CV 0.965 and LB of 0.966 (single-model + tta flip)
\nNot sure why but for me bigger models does not provide gains over r50 or r101, could be a bug in my code though

", "votes": 3, "replies": [ { "id": 1189193, "author_name": "", "author_url": "", "post_date": "2021-02-06T19:29:00.973000", "content": "

It is a blessing to find a small model with so high auc. Low costs with high accuracy too.

", "votes": 0, "replies": [] }, { "id": 1189450, "author_name": "Bcw93", "author_url": "", "post_date": "2021-02-07T03:12:35.833000", "content": "

Hi,can you share more train details about you use the resnet-50d.Thanks!

", "votes": -1, "replies": [] }, { "id": 1213211, "author_name": "", "author_url": "", "post_date": "2021-02-22T00:59:45.080000", "content": "", "votes": 0, "replies": [] }, { "id": 1213398, "author_name": "Yerram Varun", "author_url": "", "post_date": "2021-02-22T05:08:49.627000", "content": "

@arc144 Congrats for such a huge score with a single model. Can I ask you if the above-mentioned model trained by a Teacher-Student method with annotations or Directly trained on the dataset?

", "votes": 0, "replies": [] } ] }, { "id": 1176615, "author_name": "", "author_url": "", "post_date": "2021-01-29T17:41:13.577000", "content": "

The highest score from efficient published here was 0.959 on LB. Your eff model is very good. For ResNet200D is about 0.963 and SeResnet is around 0.962 on LB. (all single models)

", "votes": 2, "replies": [ { "id": 1183557, "author_name": "Akhilesh D. Kapse", "author_url": "", "post_date": "2021-02-03T04:58:25.433000", "content": "

Sir can you share the intuitions about your model ( LB ~ 0.969 )
\nThank you 🙌🙌

", "votes": 1, "replies": [] }, { "id": 1184722, "author_name": "", "author_url": "", "post_date": "2021-02-03T17:44:30.383000", "content": "

Its all about ensembles and a lot of tries. In ensembles the models must have a high disagreement ratio, while each base model must be as much accurate as possible. In order to build base models with high disagreement ratio, you can do some of the things bellow:

\n

[1] Mix diferent models like ResNet and Eff
\n[2] Mix models trained on diferent number of epochs (e.g one model close to overfitting, and another close to bias)
\n[3] Mix diferent trained folds models, eg 5 models trained on a 5 fold cyrcle.
\n[4] Mix diferent trained seeds models (models trained on diferent initial random states)
\n[5] Mix models trained on diferent images resolutions.
\n[6] Mix models trained on diferent images filters.
\n[7] TTA (Test Time Augmentation)

", "votes": 17, "replies": [] }, { "id": 1193205, "author_name": "albertofv", "author_url": "", "post_date": "2021-02-09T14:08:07.693000", "content": "

Thanks Manolis for these points.
\nIn this sense, any good resource (e.g. notebook) about ensembling models?

", "votes": 0, "replies": [] }, { "id": 1212548, "author_name": "Aman Deep Gupta", "author_url": "", "post_date": "2021-02-21T10:20:14.047000", "content": "

Thanks @manolispintelas for such a detailed information. If you find something new please share here.

", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "1175353": "Which model is providing better result?\nI get CV 0.96 using efficient-net. Which model you guys are using and what is it's CV and LB on that?", "1185848": "I have a ResNet-50d (with a few custom tweeks in architecture) with CV 0.965 and LB of 0.966 (single-model + tta flip)\nNot sure why but for me bigger models does not provide gains over r50 or r101, could be a bug in my code though", "1176615": "The highest score from efficient published here was 0.959 on LB. Your eff model is very good. For ResNet200D is about 0.963 and SeResnet is around 0.962 on LB. (all single models)" } }