Sharing some ideas here that worked for me here:
\n['site_id', 'laterality', 'view', 'implant', 'biopsy', 'invasive', 'BIRADS', 'density', 'difficult_negative_case', 'machine_id', 'age']. Auxilliary targets help to learn combined distribution of *.CSV + *.PNG data which helps to improve performance of the main classifier.Sharing some ideas here that worked for me here:
\n['site_id', 'laterality', 'view', 'implant', 'biopsy', 'invasive', 'BIRADS', 'density', 'difficult_negative_case', 'machine_id', 'age']. Auxilliary targets help to learn combined distribution of *.CSV + *.PNG data which helps to improve performance of the main classifier.hey @vslaykovsky! Really cool work, thank you for sharing! You using DEBUG as you are working on stuff to check that it works is very awesome! 🙂
Just wanted to stop by and say thank you. And also to let you know that you might want to fix the links in the OP -- some people might have a hard time getting to your NB 🙂
", "rawMarkdown": "hey @vslaykovsky! Really cool work, thank you for sharing! You using `DEBUG` as you are working on stuff to check that it works is very awesome! 🙂 \n\nJust wanted to stop by and say thank you. And also to let you know that you might want to fix the links in the OP -- some people might have a hard time getting to your NB 🙂", "votes": 2, "replies": [ { "id": 2055632, "postDate": "2022-12-05T08:56:20.323Z", "content": "Thanks for reporting!
", "rawMarkdown": "Thanks for reporting!", "votes": 1 } ] }, { "id": 2056520, "postDate": "2022-12-06T07:32:55.927Z", "content": "Thanks for sharing your nice work! Did you do experiment to test how much auxilliary can boost the CV score?
", "rawMarkdown": "Thanks for sharing your nice work! Did you do experiment to test how much auxilliary can boost the CV score? " }, { "id": 2055257, "postDate": "2022-12-04T23:54:38.980Z", "content": "Thanks for sharing, @vslaykovsky Did you try other pre-trained models than EffNetV2? I'm planning to go with EffNet as a starter, but I'm thinking to try ResNet.
", "rawMarkdown": "Thanks for sharing, @vslaykovsky Did you try other pre-trained models than EffNetV2? I'm planning to go with EffNet as a starter, but I'm thinking to try ResNet.", "replies": [ { "id": 2055633, "postDate": "2022-12-05T08:56:46.710Z", "content": "I haven't tried other models yet. Seems like a promissing idea
", "rawMarkdown": "I haven't tried other models yet. Seems like a promissing idea", "votes": 1 } ] } ], "comments": [ { "id": 2055423, "author_name": "Radek Osmulski", "author_url": "", "post_date": "2022-12-05T04:30:41.453000", "content": "hey @vslaykovsky! Really cool work, thank you for sharing! You using DEBUG as you are working on stuff to check that it works is very awesome! 🙂
Just wanted to stop by and say thank you. And also to let you know that you might want to fix the links in the OP -- some people might have a hard time getting to your NB 🙂
", "votes": 2, "replies": [ { "id": 2055632, "author_name": "Vladimir Slaykovskiy", "author_url": "", "post_date": "2022-12-05T08:56:20.323000", "content": "Thanks for reporting!
", "votes": 1, "replies": [] } ] }, { "id": 2056520, "author_name": "Mr.Fire", "author_url": "", "post_date": "2022-12-06T07:32:55.927000", "content": "Thanks for sharing your nice work! Did you do experiment to test how much auxilliary can boost the CV score?
", "votes": 0, "replies": [] }, { "id": 2055257, "author_name": "Gabriel Lins", "author_url": "", "post_date": "2022-12-04T23:54:38.980000", "content": "Thanks for sharing, @vslaykovsky Did you try other pre-trained models than EffNetV2? I'm planning to go with EffNet as a starter, but I'm thinking to try ResNet.
", "votes": 0, "replies": [ { "id": 2055633, "author_name": "Vladimir Slaykovskiy", "author_url": "", "post_date": "2022-12-05T08:56:46.710000", "content": "I haven't tried other models yet. Seems like a promissing idea
", "votes": 1, "replies": [] } ] } ], "raw_markdown_by_id": { "2055207": "Sharing some ideas here that worked for me here: \n* [Train - EffNetV2:aux targets+weighted loss+thres](https://www.kaggle.com/vslaykovsky/train-effnetv2-aux-targets-weighted-loss-thres)\n* [infer - EffNetV2:aux targets+weighted loss+thres](https://www.kaggle.com/vslaykovsky/infer-effnetv2-aux-targets-weighted-loss-thres)\n\n\n## Ideas:\n1. Both thresholding and weight-balancing improved my CV. \"cancer\" targets are highly imbalanced, so I use weighted loss to counteract (Positive weight ~50). Thresholding works well for F1-ish metrics, so optimized threshold with evaluation set. \n2. Used this [preprocessed dataset of 512x512 png images](https://www.kaggle.com/datasets/theoviel/rsna-breast-cancer-512-pngs) to speed up dataloader by ~10x with no performance degradation.\n3. Used additional classification auxilliary targets `['site_id', 'laterality', 'view', 'implant', 'biopsy', 'invasive', 'BIRADS', 'density', 'difficult_negative_case', 'machine_id', 'age']`. Auxilliary targets help to learn combined distribution of *.CSV + *.PNG data which helps to improve performance of the main classifier.\n", "2055423": "hey @vslaykovsky! Really cool work, thank you for sharing! You using `DEBUG` as you are working on stuff to check that it works is very awesome! 🙂 \n\nJust wanted to stop by and say thank you. And also to let you know that you might want to fix the links in the OP -- some people might have a hard time getting to your NB 🙂", "2056520": "Thanks for sharing your nice work! Did you do experiment to test how much auxilliary can boost the CV score? ", "2055257": "Thanks for sharing, @vslaykovsky Did you try other pre-trained models than EffNetV2? I'm planning to go with EffNet as a starter, but I'm thinking to try ResNet." } }