For me, I'm using Mask R-CNN for this competion, but I'm stuck at LB 0.318 has been on for several days and can't find a good breakthrough. (single fold only so far.)
\nlibrary: mmdetection
\ntrain/val data: 1 fold of 5 stratifiedkfold
\nmodel: default Mask R-CNN settings
Struggling, train with only one 3090 GPU.
\nGPU limited. Time limited. Can NOT explore more.
", "messages": [ { "id": "1616050", "postDate": "12/13/2021 05:00:24", "content": "For me, I'm using Mask R-CNN for this competion, but I'm stuck at LB 0.318 has been on for several days and can't find a good breakthrough. (single fold only so far.)
\nlibrary: mmdetection
\ntrain/val data: 1 fold of 5 stratifiedkfold
\nmodel: default Mask R-CNN settings
Struggling, train with only one 3090 GPU.
\nGPU limited. Time limited. Can NOT explore more.
", "rawMarkdown": "## Let's talk about the model you used in this competion~\n\nFor me, I'm using Mask R-CNN for this competion, but I'm stuck at LB 0.318 has been on for several days and can't find a good breakthrough. (single fold only so far.)\n\n### **Implementation details.**\nlibrary: mmdetection\ntrain/val data: 1 fold of 5 stratifiedkfold\nmodel: default Mask R-CNN settings\n\n### **What I learned from many Mask R-CNN experiments.**\n1. Heavy augmentations may hurt performance.\n2. The optimal thresholds for CV and LB may be different.\n3. Similar CV may have different LB. e.g., 0.2825/0.315; 0.2811/0.318; 0.2836/0.317 (CV/LB)\n4. multi-scale TTA may hurt performance and fluctuate (public LB↑ private LB↓ / public LB↓ private LB↑?).\n\n\n\n**Struggling,** train with only one 3090 GPU.\n\nGPU limited. Time limited. Can NOT explore more.", "votes": null }, { "id": "1616061", "postDate": "12/13/2021 05:20:03", "content": "Same Mask RCNN, and my strategy is almost like you. In my humble opinion, it seems that Post Processing in this game is key to your score. Color augmentation like random brightness are not proper here, and I was wondering any top competitors are using other backbones than ResNet😲. In my local CV and LB, basically they are positive related and TTA is not that powerful for me: (
", "rawMarkdown": "Same Mask RCNN, and my strategy is almost like you. In my humble opinion, it seems that Post Processing in this game is key to your score. Color augmentation like random brightness are not proper here, and I was wondering any top competitors are using other backbones than ResNet😲. In my local CV and LB, basically they are positive related and TTA is not that powerful for me: (", "votes": null }, { "id": "1616064", "postDate": "12/13/2021 05:33:09", "content": "\n\nit seems that Post Processing in this game is key to your score.
\n
Exactly. It is necessary to find an effective post processing method.
", "rawMarkdown": "> it seems that Post Processing in this game is key to your score.\n\nExactly. It is necessary to find an effective post processing method.", "votes": null }, { "id": "1616503", "postDate": "12/13/2021 13:08:39", "content": "i use cascade mask rcnn on detectron2 and got ~0.299-0.301 CV on all folds but LB was very varied :/ I think I need a strong post process to \"normalise\" these results so that my ensemble may get stronger. I'm open to teaming up if anyone wants to contact me
", "rawMarkdown": "i use cascade mask rcnn on detectron2 and got ~0.299-0.301 CV on all folds but LB was very varied :/ I think I need a strong post process to \"normalise\" these results so that my ensemble may get stronger. I'm open to teaming up if anyone wants to contact me", "votes": null }, { "id": "1616595", "postDate": "12/13/2021 14:25:38", "content": "It's also exciting to use only a single model to win the competition.🔥
", "rawMarkdown": "It's also exciting to use only a single model to win the competition.🔥", "votes": null }, { "id": "1616617", "postDate": "12/13/2021 14:41:24", "content": "haha if by exciting you mean 'nerve wracking because of fear of LB shakeup' sure 😬
", "rawMarkdown": "haha if by exciting you mean 'nerve wracking because of fear of LB shakeup' sure 😬", "votes": null }, { "id": "1622070", "postDate": "12/18/2021 10:34:49", "content": "Can I ask you, what is post processing ?
", "rawMarkdown": "Can I ask you, what is post processing ?", "votes": null }, { "id": "1622072", "postDate": "12/18/2021 10:34:59", "content": "Can I ask you, what is post processing ?
", "rawMarkdown": "Can I ask you, what is post processing ?", "votes": null }, { "id": "1622520", "postDate": "12/18/2021 20:31:25", "content": "就是后处理 你预测完了之后的处理手段
", "rawMarkdown": "就是后处理 你预测完了之后的处理手段", "votes": null }, { "id": "1622750", "postDate": "12/19/2021 05:53:40", "content": "Nerve wracking would be an understatement. Since we are also pretty much stucked, I'm really concern about how many people are holding their submission till the very last moment..
", "rawMarkdown": "Nerve wracking would be an understatement. Since we are also pretty much stucked, I'm really concern about how many people are holding their submission till the very last moment..", "votes": null }, { "id": "1623420", "postDate": "12/19/2021 21:08:54", "content": "I first used UNet and stuck at 0.15~0.16.
\nThere is a shared notebook using UNet and watershed hits 0.27 which is impressive. Cellpose is also UNet based with flow processing and hits 0.307 (https://www.kaggle.com/slawekbiel/cellpose-inference-307-lb).
\nThen I turned to Mask RCNN and reached 0.3. I guess its capability would be higher than 0.33.
", "rawMarkdown": "I first used UNet and stuck at 0.15~0.16.\n\nThere is a shared notebook using UNet and watershed hits 0.27 which is impressive. Cellpose is also UNet based with flow processing and hits 0.307 (https://www.kaggle.com/slawekbiel/cellpose-inference-307-lb).\n\nThen I turned to Mask RCNN and reached 0.3. I guess its capability would be higher than 0.33.", "votes": null }, { "id": "1624809", "postDate": "12/21/2021 08:53:02", "content": "我不懂后处理，是训练完了之后处理手段吗？就像改变detectron2中预测阈值这些吗？
", "rawMarkdown": "我不懂后处理，是训练完了之后处理手段吗？就像改变detectron2中预测阈值这些吗？", "votes": null }, { "id": "1624826", "postDate": "12/21/2021 09:15:52", "content": "Hi @blueboy97 , do you use the metric to calculate CV from here ? https://www.kaggle.com/theoviel/competition-metric-map-iou/notebook
", "rawMarkdown": "Hi @blueboy97 , do you use the metric to calculate CV from here ? https://www.kaggle.com/theoviel/competition-metric-map-iou/notebook", "votes": null }, { "id": "1624862", "postDate": "12/21/2021 10:10:44", "content": "Yes, that's it.
", "rawMarkdown": "Yes, that's it.", "votes": null }, { "id": "1625068", "postDate": "12/21/2021 13:49:55", "content": "严格来说是预测之后处理的手段 对你说的那个算一种
", "rawMarkdown": "严格来说是预测之后处理的手段 对你说的那个算一种", "votes": null } ], "comments": [ { "id": 1616061, "author_name": "charonwangg", "author_url": "", "post_date": "12/13/2021 05:20:03", "content": "Same Mask RCNN, and my strategy is almost like you. In my humble opinion, it seems that Post Processing in this game is key to your score. Color augmentation like random brightness are not proper here, and I was wondering any top competitors are using other backbones than ResNet😲. In my local CV and LB, basically they are positive related and TTA is not that powerful for me: (
", "votes": null, "replies": [ { "id": 1616064, "author_name": "blueboy97", "author_url": "", "post_date": "12/13/2021 05:33:09", "content": "\n\nit seems that Post Processing in this game is key to your score.
\n
Exactly. It is necessary to find an effective post processing method.
", "votes": null, "replies": [] }, { "id": 1622070, "author_name": "henini", "author_url": "", "post_date": "12/18/2021 10:34:49", "content": "Can I ask you, what is post processing ?
", "votes": null, "replies": [] }, { "id": 1622072, "author_name": "henini", "author_url": "", "post_date": "12/18/2021 10:34:59", "content": "Can I ask you, what is post processing ?
", "votes": null, "replies": [] }, { "id": 1622520, "author_name": "charonwangg", "author_url": "", "post_date": "12/18/2021 20:31:25", "content": "就是后处理 你预测完了之后的处理手段
", "votes": null, "replies": [] }, { "id": 1624809, "author_name": "henini", "author_url": "", "post_date": "12/21/2021 08:53:02", "content": "我不懂后处理，是训练完了之后处理手段吗？就像改变detectron2中预测阈值这些吗？
", "votes": null, "replies": [] }, { "id": 1625068, "author_name": "charonwangg", "author_url": "", "post_date": "12/21/2021 13:49:55", "content": "严格来说是预测之后处理的手段 对你说的那个算一种
", "votes": null, "replies": [] } ] }, { "id": 1616503, "author_name": "ferlockx", "author_url": "", "post_date": "12/13/2021 13:08:39", "content": "i use cascade mask rcnn on detectron2 and got ~0.299-0.301 CV on all folds but LB was very varied :/ I think I need a strong post process to \"normalise\" these results so that my ensemble may get stronger. I'm open to teaming up if anyone wants to contact me
", "votes": null, "replies": [ { "id": 1616595, "author_name": "blueboy97", "author_url": "", "post_date": "12/13/2021 14:25:38", "content": "It's also exciting to use only a single model to win the competition.🔥
", "votes": null, "replies": [] }, { "id": 1616617, "author_name": "ferlockx", "author_url": "", "post_date": "12/13/2021 14:41:24", "content": "haha if by exciting you mean 'nerve wracking because of fear of LB shakeup' sure 😬
", "votes": null, "replies": [] }, { "id": 1622750, "author_name": "woprime", "author_url": "", "post_date": "12/19/2021 05:53:40", "content": "Nerve wracking would be an understatement. Since we are also pretty much stucked, I'm really concern about how many people are holding their submission till the very last moment..
", "votes": null, "replies": [] } ] }, { "id": 1623420, "author_name": "junxhuang", "author_url": "", "post_date": "12/19/2021 21:08:54", "content": "I first used UNet and stuck at 0.15~0.16.
\nThere is a shared notebook using UNet and watershed hits 0.27 which is impressive. Cellpose is also UNet based with flow processing and hits 0.307 (https://www.kaggle.com/slawekbiel/cellpose-inference-307-lb).
\nThen I turned to Mask RCNN and reached 0.3. I guess its capability would be higher than 0.33.
", "votes": null, "replies": [] }, { "id": 1624826, "author_name": "forgakki", "author_url": "", "post_date": "12/21/2021 09:15:52", "content": "Hi @blueboy97 , do you use the metric to calculate CV from here ? https://www.kaggle.com/theoviel/competition-metric-map-iou/notebook
", "votes": null, "replies": [ { "id": 1624862, "author_name": "blueboy97", "author_url": "", "post_date": "12/21/2021 10:10:44", "content": "Yes, that's it.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1616050": "## Let's talk about the model you used in this competion~\n\nFor me, I'm using Mask R-CNN for this competion, but I'm stuck at LB 0.318 has been on for several days and can't find a good breakthrough. (single fold only so far.)\n\n### **Implementation details.**\nlibrary: mmdetection\ntrain/val data: 1 fold of 5 stratifiedkfold\nmodel: default Mask R-CNN settings\n\n### **What I learned from many Mask R-CNN experiments.**\n1. Heavy augmentations may hurt performance.\n2. The optimal thresholds for CV and LB may be different.\n3. Similar CV may have different LB. e.g., 0.2825/0.315; 0.2811/0.318; 0.2836/0.317 (CV/LB)\n4. multi-scale TTA may hurt performance and fluctuate (public LB↑ private LB↓ / public LB↓ private LB↑?).\n\n\n\n**Struggling,** train with only one 3090 GPU.\n\nGPU limited. Time limited. Can NOT explore more.", "1616061": "Same Mask RCNN, and my strategy is almost like you. In my humble opinion, it seems that Post Processing in this game is key to your score. Color augmentation like random brightness are not proper here, and I was wondering any top competitors are using other backbones than ResNet😲. In my local CV and LB, basically they are positive related and TTA is not that powerful for me: (", "1616064": "> it seems that Post Processing in this game is key to your score.\n\nExactly. It is necessary to find an effective post processing method.", "1616503": "i use cascade mask rcnn on detectron2 and got ~0.299-0.301 CV on all folds but LB was very varied :/ I think I need a strong post process to \"normalise\" these results so that my ensemble may get stronger. I'm open to teaming up if anyone wants to contact me", "1616595": "It's also exciting to use only a single model to win the competition.🔥", "1616617": "haha if by exciting you mean 'nerve wracking because of fear of LB shakeup' sure 😬", "1622070": "Can I ask you, what is post processing ?", "1622072": "Can I ask you, what is post processing ?", "1622520": "就是后处理 你预测完了之后的处理手段", "1622750": "Nerve wracking would be an understatement. Since we are also pretty much stucked, I'm really concern about how many people are holding their submission till the very last moment..", "1623420": "I first used UNet and stuck at 0.15~0.16.\n\nThere is a shared notebook using UNet and watershed hits 0.27 which is impressive. Cellpose is also UNet based with flow processing and hits 0.307 (https://www.kaggle.com/slawekbiel/cellpose-inference-307-lb).\n\nThen I turned to Mask RCNN and reached 0.3. I guess its capability would be higher than 0.33.", "1624809": "我不懂后处理，是训练完了之后处理手段吗？就像改变detectron2中预测阈值这些吗？", "1624826": "Hi @blueboy97 , do you use the metric to calculate CV from here ? https://www.kaggle.com/theoviel/competition-metric-map-iou/notebook", "1624862": "Yes, that's it.", "1625068": "严格来说是预测之后处理的手段 对你说的那个算一种" }, "source": "meta" }