I have seen that combining different loss functions gives an increase to my CV.
\nHere I have combined Taylor Cross Entropy with Label Smoothing loss, NOTEBOOK.
\nThis gives me a really good result for this fold.
Is there any discussion or study about using more than one loss function for learning? some sources will be really helpful!
", "messages": [ { "id": "1154017", "postDate": "01/15/2021 10:37:16", "content": "I have seen that combining different loss functions gives an increase to my CV.
\nHere I have combined Taylor Cross Entropy with Label Smoothing loss, NOTEBOOK.
\nThis gives me a really good result for this fold.
Is there any discussion or study about using more than one loss function for learning? some sources will be really helpful!
", "rawMarkdown": "I have seen that combining different loss functions gives an increase to my CV. \nHere I have combined Taylor Cross Entropy with Label Smoothing loss, [NOTEBOOK](https://www.kaggle.com/yerramvarun/cassava-taylorce-loss-label-smoothing-combo).\nThis gives me a really good result for this fold.\n\nIs there any discussion or study about using more than one loss function for learning? some sources will be really helpful!", "votes": null }, { "id": "1154022", "postDate": "01/15/2021 10:42:43", "content": "top work @yerramvarun
\ni used combo loss : https://arxiv.org/abs/1805.02798 in past semantic segmentation competitions, it works well but your idea for this competition is truly magnificent, thank you for sharing with us
Thank you for the praise! Glad that it helps.
\nThe paper looks interesting will try it out 👍.
Here's an Update using this Custom Loss function
\nThe PerFold results.
\nFOLD 0 - 0.89182
\nFOLD 1 - 0.91191
\nFOLD 2 - 0.90114
\nFOLD 3 - 0.89343
\nFOLD 4 - 0.89319
\nFor implementation - Notebook
\nhttps://www.kaggle.com/yerramvarun/cassava-taylorce-loss-label-smoothing-combo
For using this loss with other architectures check out @piantic 's comprehensive Notebook
\nhttps://www.kaggle.com/piantic/how-to-finetuning-models-pytorch-xla-tpu
Hi,did you test this result?How many score it get on LB?Thanks!
", "rawMarkdown": "Hi,did you test this result?How many score it get on LB?Thanks!", "votes": null }, { "id": "1158375", "postDate": "01/18/2021 14:35:33", "content": "Yes I tested this result. I haven't tuned the smoothing parameter for now. With the present parameters and the model in my notebook, it scores an LB of 0.897 alone. but when added to my ensemble my lb increased much more than that. So its true potential is yet to be explored by me.
", "rawMarkdown": "Yes I tested this result. I haven't tuned the smoothing parameter for now. With the present parameters and the model in my notebook, it scores an LB of 0.897 alone. but when added to my ensemble my lb increased much more than that. So its true potential is yet to be explored by me.", "votes": null }, { "id": "1162391", "postDate": "01/21/2021 05:57:50", "content": "I see super progress @mobassir congrats 👏
\nIs it attrubuted to completely this or other things you must have tried as well
@jaideepvalani thank you,i am working on other ideas(not spending a lot of time on this)
", "rawMarkdown": "jaideepvalani thank you,i am working on other ideas(not spending a lot of time on this)", "votes": null } ], "comments": [ { "id": 1154022, "author_name": "mobassir", "author_url": "", "post_date": "01/15/2021 10:42:43", "content": "top work @yerramvarun
\ni used combo loss : https://arxiv.org/abs/1805.02798 in past semantic segmentation competitions, it works well but your idea for this competition is truly magnificent, thank you for sharing with us
Thank you for the praise! Glad that it helps.
\nThe paper looks interesting will try it out 👍.
I see super progress @mobassir congrats 👏
\nIs it attrubuted to completely this or other things you must have tried as well
@jaideepvalani thank you,i am working on other ideas(not spending a lot of time on this)
", "votes": null, "replies": [] } ] }, { "id": 1156952, "author_name": "yerramvarun", "author_url": "", "post_date": "01/17/2021 14:44:21", "content": "Here's an Update using this Custom Loss function
\nThe PerFold results.
\nFOLD 0 - 0.89182
\nFOLD 1 - 0.91191
\nFOLD 2 - 0.90114
\nFOLD 3 - 0.89343
\nFOLD 4 - 0.89319
\nFor implementation - Notebook
\nhttps://www.kaggle.com/yerramvarun/cassava-taylorce-loss-label-smoothing-combo
For using this loss with other architectures check out @piantic 's comprehensive Notebook
\nhttps://www.kaggle.com/piantic/how-to-finetuning-models-pytorch-xla-tpu
Hi,did you test this result?How many score it get on LB?Thanks!
", "votes": null, "replies": [] }, { "id": 1158375, "author_name": "yerramvarun", "author_url": "", "post_date": "01/18/2021 14:35:33", "content": "Yes I tested this result. I haven't tuned the smoothing parameter for now. With the present parameters and the model in my notebook, it scores an LB of 0.897 alone. but when added to my ensemble my lb increased much more than that. So its true potential is yet to be explored by me.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1154017": "I have seen that combining different loss functions gives an increase to my CV. \nHere I have combined Taylor Cross Entropy with Label Smoothing loss, [NOTEBOOK](https://www.kaggle.com/yerramvarun/cassava-taylorce-loss-label-smoothing-combo).\nThis gives me a really good result for this fold.\n\nIs there any discussion or study about using more than one loss function for learning? some sources will be really helpful!", "1154022": "top work @yerramvarun \ni used combo loss : https://arxiv.org/abs/1805.02798 in past semantic segmentation competitions, it works well but your idea for this competition is truly magnificent, thank you for sharing with us", "1154096": "Thank you for the praise! Glad that it helps.\nThe paper looks interesting will try it out 👍.", "1156952": "Here's an Update using this Custom Loss function\nThe PerFold results.\nFOLD 0 - 0.89182\nFOLD 1 - 0.91191\nFOLD 2 - 0.90114\nFOLD 3 - 0.89343\nFOLD 4 - 0.89319\nFor implementation - [Notebook](https://www.kaggle.com/yerramvarun/cassava-taylorce-loss-label-smoothing-combo)\nhttps://www.kaggle.com/yerramvarun/cassava-taylorce-loss-label-smoothing-combo\n\nFor using this loss with other architectures check out @piantic 's comprehensive [Notebook](https://www.kaggle.com/piantic/how-to-finetuning-models-pytorch-xla-tpu)\nhttps://www.kaggle.com/piantic/how-to-finetuning-models-pytorch-xla-tpu", "1158169": "Hi,did you test this result?How many score it get on LB?Thanks!", "1158375": "Yes I tested this result. I haven't tuned the smoothing parameter for now. With the present parameters and the model in my notebook, it scores an LB of 0.897 alone. but when added to my ensemble my lb increased much more than that. So its true potential is yet to be explored by me.", "1162391": "I see super progress @mobassir congrats 👏 \nIs it attrubuted to completely this or other things you must have tried as well", "1162445": "jaideepvalani thank you,i am working on other ideas(not spending a lot of time on this)" }, "source": "meta" }