{ "id": 407119, "title": "Should I use another architecture(refinenet) to improve my performance?", "url": "/competitions/vesuvius-challenge-ink-detection/discussion/407119", "author_name": "", "post_date": "2023-05-05T05:32:14.553170600Z", "votes": null, "comment_count": 3, "views": 0, "content": "

In the latest notebook (LB:0.41) he uses Unet with the resnext50_32x4d model to build his CNN classifier. I tried several backbone networks such as efficientnet_b0~b5 and tried different number of image channels, but it didn't improve the performance.
\nA few hours ago, I found that there are many architectures for Image Segmentation, such as FCN, DeepLab, RefineNet.
\nshould I try this?（I am a beginner in this field.）
\nrefinenet_github_code
\nrefinenet_essay
\nsmp_architectures_API

", "messages": [ { "id": "2246378", "postDate": "05/05/2023 05:32:14", "content": "

In the latest notebook (LB:0.41) he uses Unet with the resnext50_32x4d model to build his CNN classifier. I tried several backbone networks such as efficientnet_b0~b5 and tried different number of image channels, but it didn't improve the performance.
\nA few hours ago, I found that there are many architectures for Image Segmentation, such as FCN, DeepLab, RefineNet.
\nshould I try this?（I am a beginner in this field.）
\nrefinenet_github_code
\nrefinenet_essay
\nsmp_architectures_API

", "rawMarkdown": "In the latest notebook (LB:0.41) he uses Unet with the resnext50_32x4d model to build his CNN classifier. I tried several backbone networks such as efficientnet_b0~b5 and tried different number of image channels, but it didn't improve the performance.\nA few hours ago, I found that there are many architectures for Image Segmentation, such as FCN, DeepLab, RefineNet.\nshould I try this?（I am a beginner in this field.）\n[refinenet_github_code](rehttps://github.com/DrSleep/light-weight-refinenetfinenet)\n[refinenet_essay](http://bmvc2018.org/contents/papers/0494.pdf)\n[smp_architectures_API](https://segmentation-modelspytorch.readthedocs.io/en/latest/docs/api.html#unet)", "votes": null }, { "id": "2246837", "postDate": "05/05/2023 13:56:55", "content": "

I haven't tried other architectures yet, but I think Unet should be good enough. Although I've got stuck with LB 0.4~0.45 for many days, too. Hope others will share some insights.

", "rawMarkdown": "I haven't tried other architectures yet, but I think Unet should be good enough. Although I've got stuck with LB 0.4~0.45 for many days, too. Hope others will share some insights.", "votes": null }, { "id": "2247447", "postDate": "05/06/2023 02:31:58", "content": "

I tried different encoders with that Unet notebook and achieved a slight gain LB 0.47 with the Mix Visual Transformer encoder. https://smp.readthedocs.io/en/latest/encoders.html

", "rawMarkdown": "I tried different encoders with that Unet notebook and achieved a slight gain LB 0.47 with the Mix Visual Transformer encoder. https://smp.readthedocs.io/en/latest/encoders.html", "votes": null }, { "id": "2253220", "postDate": "05/10/2023 03:22:06", "content": "

I have tried segformer in huggingface, but i got worse cv than using unet.

", "rawMarkdown": "I have tried segformer in huggingface, but i got worse cv than using unet.", "votes": null } ], "comments": [ { "id": 2246837, "author_name": "wushidiguo", "author_url": "", "post_date": "05/05/2023 13:56:55", "content": "

I haven't tried other architectures yet, but I think Unet should be good enough. Although I've got stuck with LB 0.4~0.45 for many days, too. Hope others will share some insights.

", "votes": null, "replies": [ { "id": 2253220, "author_name": "ynhuhu", "author_url": "", "post_date": "05/10/2023 03:22:06", "content": "

I have tried segformer in huggingface, but i got worse cv than using unet.

", "votes": null, "replies": [] } ] }, { "id": 2247447, "author_name": "samuelgroenjes", "author_url": "", "post_date": "05/06/2023 02:31:58", "content": "

I tried different encoders with that Unet notebook and achieved a slight gain LB 0.47 with the Mix Visual Transformer encoder. https://smp.readthedocs.io/en/latest/encoders.html

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2246378": "In the latest notebook (LB:0.41) he uses Unet with the resnext50_32x4d model to build his CNN classifier. I tried several backbone networks such as efficientnet_b0~b5 and tried different number of image channels, but it didn't improve the performance.\nA few hours ago, I found that there are many architectures for Image Segmentation, such as FCN, DeepLab, RefineNet.\nshould I try this?（I am a beginner in this field.）\n[refinenet_github_code](rehttps://github.com/DrSleep/light-weight-refinenetfinenet)\n[refinenet_essay](http://bmvc2018.org/contents/papers/0494.pdf)\n[smp_architectures_API](https://segmentation-modelspytorch.readthedocs.io/en/latest/docs/api.html#unet)", "2246837": "I haven't tried other architectures yet, but I think Unet should be good enough. Although I've got stuck with LB 0.4~0.45 for many days, too. Hope others will share some insights.", "2247447": "I tried different encoders with that Unet notebook and achieved a slight gain LB 0.47 with the Mix Visual Transformer encoder. https://smp.readthedocs.io/en/latest/encoders.html", "2253220": "I have tried segformer in huggingface, but i got worse cv than using unet." }, "source": "meta" }