Hey all, I reworked julian3833's notebook with pytorch lightning, but for some reason it gets 0.000 on submission. I don't think I changed the RLE code at all but if someone wants to work on this with me I'm happy to collaborate. Thanks!
\nhttps://www.kaggle.com/evancofsky/sartorius-starter-torch-mask-r-cnn-w-lightnin
", "messages": [ { "id": "1554207", "postDate": "10/22/2021 22:57:02", "content": "Hey all, I reworked julian3833's notebook with pytorch lightning, but for some reason it gets 0.000 on submission. I don't think I changed the RLE code at all but if someone wants to work on this with me I'm happy to collaborate. Thanks!
\nhttps://www.kaggle.com/evancofsky/sartorius-starter-torch-mask-r-cnn-w-lightnin
", "rawMarkdown": "Hey all, I reworked julian3833's notebook with pytorch lightning, but for some reason it gets 0.000 on submission. I don't think I changed the RLE code at all but if someone wants to work on this with me I'm happy to collaborate. Thanks!\n\nhttps://www.kaggle.com/evancofsky/sartorius-starter-torch-mask-r-cnn-w-lightnin", "votes": null }, { "id": "1554209", "postDate": "10/22/2021 23:05:44", "content": "try removing overlap ? https://www.kaggle.com/c/sartorius-cell-instance-segmentation/discussion/279790
", "rawMarkdown": "try removing overlap ? https://www.kaggle.com/c/sartorius-cell-instance-segmentation/discussion/279790", "votes": null }, { "id": "1554427", "postDate": "10/23/2021 04:49:05", "content": "Thank you! I’ll go through the overlap code and see if I can find the issue.
", "rawMarkdown": "Thank you! I’ll go through the overlap code and see if I can find the issue.", "votes": null }, { "id": "1554597", "postDate": "10/23/2021 08:32:21", "content": "Hello Evan, I think the reason for that is (the decoding function), it is decoding a single instance at a time, therefore your model is left with a mask that is mostly black and hence it predicts the same. Try to decode all the instances at once and I believe you'll see an improvement. Also, I'd love to collaborate with you in this competition so lmk :D
", "rawMarkdown": "Hello Evan, I think the reason for that is (the decoding function), it is decoding a single instance at a time, therefore your model is left with a mask that is mostly black and hence it predicts the same. Try to decode all the instances at once and I believe you'll see an improvement. Also, I'd love to collaborate with you in this competition so lmk :D", "votes": null }, { "id": "1557331", "postDate": "10/25/2021 15:44:41", "content": "Turns out there's a pretty weird bug in the flag in the notebook I have. I don't know if I did it or if it was in there, but I suspect I added it when I wasn't sure what it was doing and never set it back. I've updated it to always remove overlaps and I'm testing it now.
\nThanks for the tips!
", "rawMarkdown": "Turns out there's a pretty weird bug in the flag in the notebook I have. I don't know if I did it or if it was in there, but I suspect I added it when I wasn't sure what it was doing and never set it back. I've updated it to always remove overlaps and I'm testing it now.\n\nThanks for the tips!", "votes": null }, { "id": "1557873", "postDate": "10/25/2021 22:26:08", "content": "Yeah I'd definitely love to work with someone else. I'm hoping to see if I've fixed this issue today sometime, but I'm nearly out of GPU credits so I'm doing testing in colab now with limited epochs… If you want to join my \"team\" (one person, The UNIX Man haha) let me know, otherwise I could join yours.
", "rawMarkdown": "Yeah I'd definitely love to work with someone else. I'm hoping to see if I've fixed this issue today sometime, but I'm nearly out of GPU credits so I'm doing testing in colab now with limited epochs... If you want to join my \"team\" (one person, The UNIX Man haha) let me know, otherwise I could join yours.", "votes": null }, { "id": "1557877", "postDate": "10/25/2021 22:31:11", "content": "finger crossed =)
", "rawMarkdown": "finger crossed =)", "votes": null }, { "id": "1599392", "postDate": "11/29/2021 12:02:03", "content": "@evancofsky how did u resolve the issue.. i too face the same
", "rawMarkdown": "evancofsky how did u resolve the issue.. i too face the same", "votes": null }, { "id": "1599763", "postDate": "11/29/2021 19:31:58", "content": "@drhabib had a great write up on it https://www.kaggle.com/c/sartorius-cell-instance-segmentation/discussion/279790
", "rawMarkdown": "drhabib had a great write up on it https://www.kaggle.com/c/sartorius-cell-instance-segmentation/discussion/279790", "votes": null } ], "comments": [ { "id": 1554209, "author_name": "drhabib", "author_url": "", "post_date": "10/22/2021 23:05:44", "content": "try removing overlap ? https://www.kaggle.com/c/sartorius-cell-instance-segmentation/discussion/279790
", "votes": null, "replies": [ { "id": 1554427, "author_name": "evancofsky", "author_url": "", "post_date": "10/23/2021 04:49:05", "content": "Thank you! I’ll go through the overlap code and see if I can find the issue.
", "votes": null, "replies": [] }, { "id": 1557331, "author_name": "evancofsky", "author_url": "", "post_date": "10/25/2021 15:44:41", "content": "Turns out there's a pretty weird bug in the flag in the notebook I have. I don't know if I did it or if it was in there, but I suspect I added it when I wasn't sure what it was doing and never set it back. I've updated it to always remove overlaps and I'm testing it now.
\nThanks for the tips!
", "votes": null, "replies": [] }, { "id": 1557877, "author_name": "drhabib", "author_url": "", "post_date": "10/25/2021 22:31:11", "content": "finger crossed =)
", "votes": null, "replies": [] } ] }, { "id": 1554597, "author_name": "eddwait", "author_url": "", "post_date": "10/23/2021 08:32:21", "content": "Hello Evan, I think the reason for that is (the decoding function), it is decoding a single instance at a time, therefore your model is left with a mask that is mostly black and hence it predicts the same. Try to decode all the instances at once and I believe you'll see an improvement. Also, I'd love to collaborate with you in this competition so lmk :D
", "votes": null, "replies": [ { "id": 1557873, "author_name": "evancofsky", "author_url": "", "post_date": "10/25/2021 22:26:08", "content": "Yeah I'd definitely love to work with someone else. I'm hoping to see if I've fixed this issue today sometime, but I'm nearly out of GPU credits so I'm doing testing in colab now with limited epochs… If you want to join my \"team\" (one person, The UNIX Man haha) let me know, otherwise I could join yours.
", "votes": null, "replies": [] } ] }, { "id": 1599392, "author_name": "jaideepvalani", "author_url": "", "post_date": "11/29/2021 12:02:03", "content": "@evancofsky how did u resolve the issue.. i too face the same
", "votes": null, "replies": [ { "id": 1599763, "author_name": "evancofsky", "author_url": "", "post_date": "11/29/2021 19:31:58", "content": "@drhabib had a great write up on it https://www.kaggle.com/c/sartorius-cell-instance-segmentation/discussion/279790
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1554207": "Hey all, I reworked julian3833's notebook with pytorch lightning, but for some reason it gets 0.000 on submission. I don't think I changed the RLE code at all but if someone wants to work on this with me I'm happy to collaborate. Thanks!\n\nhttps://www.kaggle.com/evancofsky/sartorius-starter-torch-mask-r-cnn-w-lightnin", "1554209": "try removing overlap ? https://www.kaggle.com/c/sartorius-cell-instance-segmentation/discussion/279790", "1554427": "Thank you! I’ll go through the overlap code and see if I can find the issue.", "1554597": "Hello Evan, I think the reason for that is (the decoding function), it is decoding a single instance at a time, therefore your model is left with a mask that is mostly black and hence it predicts the same. Try to decode all the instances at once and I believe you'll see an improvement. Also, I'd love to collaborate with you in this competition so lmk :D", "1557331": "Turns out there's a pretty weird bug in the flag in the notebook I have. I don't know if I did it or if it was in there, but I suspect I added it when I wasn't sure what it was doing and never set it back. I've updated it to always remove overlaps and I'm testing it now.\n\nThanks for the tips!", "1557873": "Yeah I'd definitely love to work with someone else. I'm hoping to see if I've fixed this issue today sometime, but I'm nearly out of GPU credits so I'm doing testing in colab now with limited epochs... If you want to join my \"team\" (one person, The UNIX Man haha) let me know, otherwise I could join yours.", "1557877": "finger crossed =)", "1599392": "evancofsky how did u resolve the issue.. i too face the same", "1599763": "drhabib had a great write up on it https://www.kaggle.com/c/sartorius-cell-instance-segmentation/discussion/279790" }, "source": "meta" }