![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3924325%2F0123b313bdc492661d572fc4c3dab6ed%2Fsdgp.JPG?generation=1596093943582809&alt=media\")
It's not every day a new version of SGD and Adam is released, and after reading the research paper it seems promising, worth testing and evaluating. Anyone tried them yet?
\n\n
https://github.com/clovaai/AdamP\nhttps://arxiv.org/abs/2006.08217
", "messages": [ { "id": "951512", "postDate": "07/30/2020 07:26:35", "content": "It's not every day a new version of SGD and Adam is released, and after reading the research paper it seems promising, worth testing and evaluating. Anyone tried them yet?
\n\n
https://github.com/clovaai/AdamP\nhttps://arxiv.org/abs/2006.08217
", "rawMarkdown": "It's not every day a new version of SGD and Adam is released, and after reading the research paper it seems promising, worth testing and evaluating. Anyone tried them yet?\n\n\n\nhttps://github.com/clovaai/AdamP\nhttps://arxiv.org/abs/2006.08217", "votes": null }, { "id": "955639", "postDate": "08/02/2020 19:08:24", "content": "Are we allowed to use them if it wasn't posted in external data thread before the deadline?
", "rawMarkdown": "Are we allowed to use them if it wasn't posted in external data thread before the deadline?", "votes": null }, { "id": "955651", "postDate": "08/02/2020 19:23:11", "content": "Good question, don't know if optimizers and schedulers are part of External Data Thread, have posted a question to the Kaggle Team. But have never seen such report/post before, like for Radam, AdamW, special schedulers, dataloaders etc that not are built-in.
", "rawMarkdown": "Good question, don't know if optimizers and schedulers are part of External Data Thread, have posted a question to the Kaggle Team. But have never seen such report/post before, like for Radam, AdamW, special schedulers, dataloaders etc that not are built-in.", "votes": null }, { "id": "958098", "postDate": "08/04/2020 19:03:48", "content": "Anyway, I tried AdamP, and every time it resulted in about +0.002 LB compared to AdamW. So thank you 😊
", "rawMarkdown": "Anyway, I tried AdamP, and every time it resulted in about +0.002 LB compared to AdamW. So thank you 😊", "votes": null }, { "id": "958766", "postDate": "08/05/2020 06:15:10", "content": "Did anyone try these optimizers? If so, how did it go?
", "rawMarkdown": "Did anyone try these optimizers? If so, how did it go?", "votes": null }, { "id": "959010", "postDate": "08/05/2020 09:33:56", "content": "Yes it worked well and slightly better results but had problem with those and apex. I think apex works best with built-in features and not so complex structure, just a personal reflection, or it is just a question of knowledge.
", "rawMarkdown": "Yes it worked well and slightly better results but had problem with those and apex. I think apex works best with built-in features and not so complex structure, just a personal reflection, or it is just a question of knowledge.", "votes": null } ], "comments": [ { "id": 955639, "author_name": "alexxx", "author_url": "", "post_date": "08/02/2020 19:08:24", "content": "Are we allowed to use them if it wasn't posted in external data thread before the deadline?
", "votes": null, "replies": [ { "id": 955651, "author_name": "kirderf", "author_url": "", "post_date": "08/02/2020 19:23:11", "content": "Good question, don't know if optimizers and schedulers are part of External Data Thread, have posted a question to the Kaggle Team. But have never seen such report/post before, like for Radam, AdamW, special schedulers, dataloaders etc that not are built-in.
", "votes": null, "replies": [ { "id": 958098, "author_name": "alexxx", "author_url": "", "post_date": "08/04/2020 19:03:48", "content": "Anyway, I tried AdamP, and every time it resulted in about +0.002 LB compared to AdamW. So thank you 😊
", "votes": null, "replies": [] } ] } ] }, { "id": 958766, "author_name": "tanmaypandey", "author_url": "", "post_date": "08/05/2020 06:15:10", "content": "Did anyone try these optimizers? If so, how did it go?
", "votes": null, "replies": [ { "id": 959010, "author_name": "kirderf", "author_url": "", "post_date": "08/05/2020 09:33:56", "content": "Yes it worked well and slightly better results but had problem with those and apex. I think apex works best with built-in features and not so complex structure, just a personal reflection, or it is just a question of knowledge.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "951512": "It's not every day a new version of SGD and Adam is released, and after reading the research paper it seems promising, worth testing and evaluating. Anyone tried them yet?\n\n\n\nhttps://github.com/clovaai/AdamP\nhttps://arxiv.org/abs/2006.08217", "955639": "Are we allowed to use them if it wasn't posted in external data thread before the deadline?", "955651": "Good question, don't know if optimizers and schedulers are part of External Data Thread, have posted a question to the Kaggle Team. But have never seen such report/post before, like for Radam, AdamW, special schedulers, dataloaders etc that not are built-in.", "958098": "Anyway, I tried AdamP, and every time it resulted in about +0.002 LB compared to AdamW. So thank you 😊", "958766": "Did anyone try these optimizers? If so, how did it go?", "959010": "Yes it worked well and slightly better results but had problem with those and apex. I think apex works best with built-in features and not so complex structure, just a personal reflection, or it is just a question of knowledge." }, "source": "meta" }