Hi, I am training 3D Unet from Monai with decent results. During training the validation loss is improving steadily.. But when I switch to other architectures based on resnet like SeResnetDS, the model is much worse and the validation loss / dice metric is unstable. But I have see people using them with success, what could I doing wrong. I try to decrease learning rate etc. but without succes. Thx
", "messages": [ { "id": "3089594", "postDate": "01/06/2025 09:09:32", "content": "Hi, I am training 3D Unet from Monai with decent results. During training the validation loss is improving steadily.. But when I switch to other architectures based on resnet like SeResnetDS, the model is much worse and the validation loss / dice metric is unstable. But I have see people using them with success, what could I doing wrong. I try to decrease learning rate etc. but without succes. Thx
", "rawMarkdown": "Hi, I am training 3D Unet from Monai with decent results. During training the validation loss is improving steadily.. But when I switch to other architectures based on resnet like SeResnetDS, the model is much worse and the validation loss / dice metric is unstable. But I have see people using them with success, what could I doing wrong. I try to decrease learning rate etc. but without succes. Thx", "votes": null }, { "id": "3089598", "postDate": "01/06/2025 09:15:27", "content": "Sounds a lot like what happens with Monai as well under the right conditions:
\nhttps://www.kaggle.com/competitions/czii-cryo-et-object-identification/discussion/551740
I try to avoid this issue by using FocalDiceloss instead of Tversky loss. But you are probably right that it has something to do with a lot of negative (background only) samples.
", "rawMarkdown": "I try to avoid this issue by using FocalDiceloss instead of Tversky loss. But you are probably right that it has something to do with a lot of negative (background only) samples.", "votes": null }, { "id": "3090158", "postDate": "01/07/2025 00:18:08", "content": "Try making your labels bigger. Or if you go back to Tversky loss you can adjust alpha and beta. Both of those may fix the instability…but unfortunately in my experience they also resulted in worse overall performance. …Though it's possible at least some of that was due to a separate bug.
", "rawMarkdown": "Try making your labels bigger. Or if you go back to Tversky loss you can adjust alpha and beta. Both of those may fix the instability...but unfortunately in my experience they also resulted in worse overall performance. ...Though it's possible at least some of that was due to a separate bug.", "votes": null }, { "id": "3090398", "postDate": "01/07/2025 08:19:25", "content": "Making labels bigger helps. I am now also trying to make them less deep and more wide, and add deep supervision.
", "rawMarkdown": "Making labels bigger helps. I am now also trying to make them less deep and more wide, and add deep supervision.", "votes": null } ], "comments": [ { "id": 3089598, "author_name": "davidlist", "author_url": "", "post_date": "01/06/2025 09:15:27", "content": "Sounds a lot like what happens with Monai as well under the right conditions:
\nhttps://www.kaggle.com/competitions/czii-cryo-et-object-identification/discussion/551740
I try to avoid this issue by using FocalDiceloss instead of Tversky loss. But you are probably right that it has something to do with a lot of negative (background only) samples.
", "votes": null, "replies": [ { "id": 3090158, "author_name": "davidlist", "author_url": "", "post_date": "01/07/2025 00:18:08", "content": "Try making your labels bigger. Or if you go back to Tversky loss you can adjust alpha and beta. Both of those may fix the instability…but unfortunately in my experience they also resulted in worse overall performance. …Though it's possible at least some of that was due to a separate bug.
", "votes": null, "replies": [ { "id": 3090398, "author_name": "tiborvansa", "author_url": "", "post_date": "01/07/2025 08:19:25", "content": "Making labels bigger helps. I am now also trying to make them less deep and more wide, and add deep supervision.
", "votes": null, "replies": [] } ] } ] } ] } ], "raw_markdown_by_id": { "3089594": "Hi, I am training 3D Unet from Monai with decent results. During training the validation loss is improving steadily.. But when I switch to other architectures based on resnet like SeResnetDS, the model is much worse and the validation loss / dice metric is unstable. But I have see people using them with success, what could I doing wrong. I try to decrease learning rate etc. but without succes. Thx", "3089598": "Sounds a lot like what happens with Monai as well under the right conditions:\n[https://www.kaggle.com/competitions/czii-cryo-et-object-identification/discussion/551740](https://www.kaggle.com/competitions/czii-cryo-et-object-identification/discussion/551740)", "3089685": "I try to avoid this issue by using FocalDiceloss instead of Tversky loss. But you are probably right that it has something to do with a lot of negative (background only) samples.", "3090158": "Try making your labels bigger. Or if you go back to Tversky loss you can adjust alpha and beta. Both of those may fix the instability...but unfortunately in my experience they also resulted in worse overall performance. ...Though it's possible at least some of that was due to a separate bug.", "3090398": "Making labels bigger helps. I am now also trying to make them less deep and more wide, and add deep supervision." }, "source": "meta" }