GAN's - Inspired by this paper https://www.paperswithcode.com/paper/melanoma-detection-using-adversarial-training-1 and Rohit's post. ![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F413189%2F3dd7a16f80823987e41fcd2d9f252904%2FGAN%20Sample.png?generation=1597751688287238&alt=media\")
Chipnet - Even though the performance was impressive, we couldn't complete it due to paucity of time.
Congrats to winners..
\nI would like to thank my teammate Sebastien (@Anhmeow) for his amazing contribution in getting this position. Our best solution was rank ensemble of 8 different models which we had created in parallel and ensembled it using stacking.
\nWe had created 80 models in total and selected only those models which were less correlated.
\nThe mean ensemble of these models scored around 0.9432 in private LB while rank ensemble scored 0.9406.
\nSpecial thanks to @Cdeotte for providing wonderful notebooks and discussion threads
\nWhat we tried but couldn't complete:
\nGAN's - Inspired by this paper https://www.paperswithcode.com/paper/melanoma-detection-using-adversarial-training-1 and Rohit's post.
Chipnet - Even though the performance was impressive, we couldn't complete it due to paucity of time.
Congrats to winners..
\nI would like to thank my teammate Sebastien (@Anhmeow) for his amazing contribution in getting this position. Our best solution was rank ensemble of 8 different models which we had created in parallel and ensembled it using stacking.
\nWe had created 80 models in total and selected only those models which were less correlated.
\nThe mean ensemble of these models scored around 0.9432 in private LB while rank ensemble scored 0.9406.
\nSpecial thanks to @Cdeotte for providing wonderful notebooks and discussion threads
\nWhat we tried but couldn't complete:
\nGAN's - Inspired by this paper https://www.paperswithcode.com/paper/melanoma-detection-using-adversarial-training-1 and Rohit's post.
Chipnet - Even though the performance was impressive, we couldn't complete it due to paucity of time.
Could you also share how you ensembled these models and how you did CV?
\nThanks.
@fahad7 : it was just averaging the predictions for ensembling. And for CV, it was stratified K Fold cross validation.
", "rawMarkdown": "fahad7 : it was just averaging the predictions for ensembling. And for CV, it was stratified K Fold cross validation.", "votes": null } ], "comments": [ { "id": 975479, "author_name": "fahad7", "author_url": "", "post_date": "08/18/2020 10:06:52", "content": "Could you also share how you ensembled these models and how you did CV?
\nThanks.
@fahad7 : it was just averaging the predictions for ensembling. And for CV, it was stratified K Fold cross validation.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "975408": "Congrats to winners.. \n\n\nI would like to thank my teammate Sebastien (@Anhmeow) for his amazing contribution in getting this position. Our best solution was rank ensemble of 8 different models which we had created in parallel and ensembled it using stacking. \n\nWe had created 80 models in total and selected only those models which were less correlated.\n\n1. EfficientNet B4 - Image size - 512 - Sebastien\n2. EfficientNet B4 - Image Size - 768 - Sebastien\n3. EfficientNet B6 - Image Size - 768 - Sebastien\n4. EfficientNet B5 - Image Size - 384 - Manoj\n5. EfficientNet B6 - Image Size - 512 - Including external data - Manoj\n6. EfficientNet B5 - Image size - 384 - without external data - Manoj\n7. EfficientNet B5 - Image size - 384 - Including external data - Manoj\n8. EfficientNet B6 - Image size - 512 - without external data - Sebastien\n\nThe mean ensemble of these models scored around 0.9432 in private LB while rank ensemble scored 0.9406. \n\nSpecial thanks to @Cdeotte for providing wonderful notebooks and discussion threads\n\nWhat we tried but couldn't complete:\n\n1. GAN's - Inspired by this paper https://www.paperswithcode.com/paper/melanoma-detection-using-adversarial-training-1 and Rohit's post. \n\n2. Chipnet - Even though the performance was impressive, we couldn't complete it due to paucity of time.", "975479": "Could you also share how you ensembled these models and how you did CV?\nThanks.", "975608": "fahad7 : it was just averaging the predictions for ensembling. And for CV, it was stratified K Fold cross validation." }, "source": "meta" }