I apply cutmix & fmix to effnet04 and it's reduced the valid accuracy. Despite I found many discussions talk about this as the ways to improve the accuracy.
\nI try it many times and I didn't get the fault, Anyone has an idea about what's going on?
", "messages": [ { "id": "1187406", "postDate": "02/05/2021 12:22:26", "content": "I apply cutmix & fmix to effnet04 and it's reduced the valid accuracy. Despite I found many discussions talk about this as the ways to improve the accuracy.
\nI try it many times and I didn't get the fault, Anyone has an idea about what's going on?
", "rawMarkdown": "I apply cutmix & fmix to effnet04 and it's reduced the valid accuracy. Despite I found many discussions talk about this as the ways to improve the accuracy.\n\nI try it many times and I didn't get the fault, **Anyone has an idea about what's going on?**", "votes": null }, { "id": "1187443", "postDate": "02/05/2021 12:59:35", "content": "cutmix was not helpful for me either. I didn't try fmix.
", "rawMarkdown": "cutmix was not helpful for me either. I didn't try fmix.", "votes": null }, { "id": "1187451", "postDate": "02/05/2021 13:02:34", "content": "thanks, @serigne,
\ndid you try label smoothing?
I don't use Label Smoothing (I use a custom loss)
\nBut LabelSmoothing is clearly more helpful than regular Cross Entropy here.
", "rawMarkdown": "I don't use Label Smoothing (I use a custom loss)\n\nBut LabelSmoothing is clearly more helpful than regular Cross Entropy here.", "votes": null }, { "id": "1187632", "postDate": "02/05/2021 15:36:09", "content": "I suspect what's going on is what I describe in more detail here. In short, on some image datasets (let's say CIFAR-10) it is very clear that a sensible label after applying CutMix indeed corresponds to the proportion of area from each image, but for this competition I don't think that's the case. I think that, because a leaf with one diseased corner of the leaf is still diseased, even if you replace 80% of the leaf with something from an image of healthy leaf.
", "rawMarkdown": "I suspect what's going on is what I describe in more detail [here](https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/217150#1187623). In short, on some image datasets (let's say CIFAR-10) it is very clear that a sensible label after applying CutMix indeed corresponds to the proportion of area from each image, but for this competition I don't think that's the case. I think that, because a leaf with one diseased corner of the leaf is still diseased, even if you replace 80% of the leaf with something from an image of healthy leaf.", "votes": null }, { "id": "1187670", "postDate": "02/05/2021 16:09:23", "content": "For me cutmix works but you cant just add it to your original pipeline with hard regularization because you will underfit. I had to change many things to make it work
", "rawMarkdown": "For me cutmix works but you cant just add it to your original pipeline with hard regularization because you will underfit. I had to change many things to make it work", "votes": null }, { "id": "1187687", "postDate": "02/05/2021 16:25:13", "content": "Cutmix alone did improve the performance of my model. I apply cutmix only with 50% probability on each batch
", "rawMarkdown": "Cutmix alone did improve the performance of my model. I apply cutmix only with 50% probability on each batch", "votes": null }, { "id": "1187927", "postDate": "02/05/2021 19:00:35", "content": "Heavy augmentation such as cutout, cutmix or snapmix requires resetting of training parameters. Adjust trainning epochs, lr and scheduler may help.
", "rawMarkdown": "Heavy augmentation such as cutout, cutmix or snapmix requires resetting of training parameters. Adjust trainning epochs, lr and scheduler may help.", "votes": null }, { "id": "1187935", "postDate": "02/05/2021 19:08:48", "content": "same for me, you can check CutMix and MixUp help? for reference.
", "rawMarkdown": "same for me, you can check [CutMix and MixUp help?](https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/214225) for reference.", "votes": null }, { "id": "1189159", "postDate": "02/06/2021 18:53:54", "content": "Thanks, @angqx95!
\nwhat model did you use?
i used effnetb4
", "rawMarkdown": "i used effnetb4", "votes": null } ], "comments": [ { "id": 1187443, "author_name": "serigne", "author_url": "", "post_date": "02/05/2021 12:59:35", "content": "cutmix was not helpful for me either. I didn't try fmix.
", "votes": null, "replies": [ { "id": 1187451, "author_name": "elcaiseri", "author_url": "", "post_date": "02/05/2021 13:02:34", "content": "thanks, @serigne,
\ndid you try label smoothing?
I don't use Label Smoothing (I use a custom loss)
\nBut LabelSmoothing is clearly more helpful than regular Cross Entropy here.
", "votes": null, "replies": [] } ] }, { "id": 1187632, "author_name": "bjoernholzhauer", "author_url": "", "post_date": "02/05/2021 15:36:09", "content": "I suspect what's going on is what I describe in more detail here. In short, on some image datasets (let's say CIFAR-10) it is very clear that a sensible label after applying CutMix indeed corresponds to the proportion of area from each image, but for this competition I don't think that's the case. I think that, because a leaf with one diseased corner of the leaf is still diseased, even if you replace 80% of the leaf with something from an image of healthy leaf.
", "votes": null, "replies": [] }, { "id": 1187670, "author_name": "yannmajewski", "author_url": "", "post_date": "02/05/2021 16:09:23", "content": "For me cutmix works but you cant just add it to your original pipeline with hard regularization because you will underfit. I had to change many things to make it work
", "votes": null, "replies": [] }, { "id": 1187687, "author_name": "angqx95", "author_url": "", "post_date": "02/05/2021 16:25:13", "content": "Cutmix alone did improve the performance of my model. I apply cutmix only with 50% probability on each batch
", "votes": null, "replies": [ { "id": 1189159, "author_name": "elcaiseri", "author_url": "", "post_date": "02/06/2021 18:53:54", "content": "Thanks, @angqx95!
\nwhat model did you use?
i used effnetb4
", "votes": null, "replies": [] } ] }, { "id": 1187927, "author_name": "steamedsheep", "author_url": "", "post_date": "02/05/2021 19:00:35", "content": "Heavy augmentation such as cutout, cutmix or snapmix requires resetting of training parameters. Adjust trainning epochs, lr and scheduler may help.
", "votes": null, "replies": [] }, { "id": 1187935, "author_name": "leonshangguan", "author_url": "", "post_date": "02/05/2021 19:08:48", "content": "same for me, you can check CutMix and MixUp help? for reference.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1187406": "I apply cutmix & fmix to effnet04 and it's reduced the valid accuracy. Despite I found many discussions talk about this as the ways to improve the accuracy.\n\nI try it many times and I didn't get the fault, **Anyone has an idea about what's going on?**", "1187443": "cutmix was not helpful for me either. I didn't try fmix.", "1187451": "thanks, @serigne,\ndid you try label smoothing?", "1187459": "I don't use Label Smoothing (I use a custom loss)\n\nBut LabelSmoothing is clearly more helpful than regular Cross Entropy here.", "1187632": "I suspect what's going on is what I describe in more detail [here](https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/217150#1187623). In short, on some image datasets (let's say CIFAR-10) it is very clear that a sensible label after applying CutMix indeed corresponds to the proportion of area from each image, but for this competition I don't think that's the case. I think that, because a leaf with one diseased corner of the leaf is still diseased, even if you replace 80% of the leaf with something from an image of healthy leaf.", "1187670": "For me cutmix works but you cant just add it to your original pipeline with hard regularization because you will underfit. I had to change many things to make it work", "1187687": "Cutmix alone did improve the performance of my model. I apply cutmix only with 50% probability on each batch", "1187927": "Heavy augmentation such as cutout, cutmix or snapmix requires resetting of training parameters. Adjust trainning epochs, lr and scheduler may help.", "1187935": "same for me, you can check [CutMix and MixUp help?](https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/214225) for reference.", "1189159": "Thanks, @angqx95!\nwhat model did you use?", "1191498": "i used effnetb4" }, "source": "meta" }