As the competition is over, I think it's the time to ask this question.\nI tried to use U-net in this competition, since I used it to do binary classification and segmentation before. However I faced a lot of difficulties to do a multiple classification and segmentation. \nMy solution is to transfer the input mask into a [-1, 224, 224, 47] np.array, code like this: sub_mask = cv2.resize(sub_mask, (SIZE, SIZE), interpolation=cv2.INTER_NEAREST)\n mask[:, :, int(label)] = sub_mask, where 47 means 47 types(including the background), then when doing inference, I don't know how to choose the channel, and actually \nI think the channel is not matching the label.\nSo, as a beginner, I am really curious if anyone who used U-net in this competition and realized the multiple classification and segmentation. \nI will appreciate it if you can share the code or even just give some hints. Thank you.
Additional: I'm using keras and tf.
", "messages": [ { "id": "549851", "postDate": "06/11/2019 04:57:36", "content": "As the competition is over, I think it's the time to ask this question.\nI tried to use U-net in this competition, since I used it to do binary classification and segmentation before. However I faced a lot of difficulties to do a multiple classification and segmentation. \nMy solution is to transfer the input mask into a [-1, 224, 224, 47] np.array, code like this: sub_mask = cv2.resize(sub_mask, (SIZE, SIZE), interpolation=cv2.INTER_NEAREST)\n mask[:, :, int(label)] = sub_mask, where 47 means 47 types(including the background), then when doing inference, I don't know how to choose the channel, and actually \nI think the channel is not matching the label.\nSo, as a beginner, I am really curious if anyone who used U-net in this competition and realized the multiple classification and segmentation. \nI will appreciate it if you can share the code or even just give some hints. Thank you.
Additional: I'm using keras and tf.
", "rawMarkdown": "As the competition is over, I think it's the time to ask this question.\nI tried to use U-net in this competition, since I used it to do binary classification and segmentation before. However I faced a lot of difficulties to do a multiple classification and segmentation. \nMy solution is to transfer the input mask into a [-1, 224, 224, 47] np.array, code like this: `sub_mask = cv2.resize(sub_mask, (SIZE, SIZE), interpolation=cv2.INTER_NEAREST)\n mask[:, :, int(label)] = sub_mask`, where 47 means 47 types(including the background), then when doing inference, I don't know how to choose the channel, and actually \nI think the channel is not matching the label.\nSo, as a beginner, I am really curious if anyone who used U-net in this competition and realized the multiple classification and segmentation. \nI will appreciate it if you can share the code or even just give some hints. Thank you.\n\nAdditional: I'm using keras and tf.", "votes": null }, { "id": "549865", "postDate": "06/11/2019 05:20:23", "content": "I was using a lot of Unet with different encoders (actually more than 100, 2-3almost for each class) \nStarted with them to check submission, they had no good score. But they could be used to refine an outputted mask
", "rawMarkdown": "I was using a lot of Unet with different encoders (actually more than 100, 2-3almost for each class) \nStarted with them to check submission, they had no good score. But they could be used to refine an outputted mask", "votes": null }, { "id": "549879", "postDate": "06/11/2019 05:32:03", "content": "Sorry for that, Could you explain what do you mean by > actually more than 100, 2-3almost for each class? Does that mean you regard it as a 47 bi-classification & segmentation?
", "rawMarkdown": "Sorry for that, Could you explain what do you mean by > actually more than 100, 2-3almost for each class? Does that mean you regard it as a 47 bi-classification & segmentation?", "votes": null }, { "id": "549884", "postDate": "06/11/2019 05:39:39", "content": "You are correct\nModel for each class, different models as well
", "rawMarkdown": "You are correct\nModel for each class, different models as well", "votes": null }, { "id": "549887", "postDate": "06/11/2019 05:44:29", "content": "I used U-Net at the beginning of this competition. \ninput: 768 * 768 * 3, output: 768 * 768 * 46. \nOutput is activated by sigmoid and trained by BCELoss because one pixel may has two or more class.\nAfter some postprocessing, my U-Net best score is 0.10007(private)/0.10582(public), so I decided to use mask-rcnn.
", "rawMarkdown": "I used U-Net at the beginning of this competition. \ninput: 768 * 768 * 3, output: 768 * 768 * 46. \nOutput is activated by sigmoid and trained by BCELoss because one pixel may has two or more class.\nAfter some postprocessing, my U-Net best score is 0.10007(private)/0.10582(public), so I decided to use mask-rcnn.", "votes": null }, { "id": "549917", "postDate": "06/11/2019 06:25:24", "content": "Seems that demands excellent GPU...
", "rawMarkdown": "Seems that demands excellent GPU...", "votes": null }, { "id": "549922", "postDate": "06/11/2019 06:28:16", "content": "DGX-1 in my control
", "rawMarkdown": "DGX-1 in my control", "votes": null }, { "id": "549928", "postDate": "06/11/2019 06:34:58", "content": "If one pixel may has two or more class, why did you choose sigmoid & bce instead of softmax & cce? And, as you have 46 categories, where is the background?\nWell, and a little beg(perhaps offensive), since the competition is over, could you please share your code of the Unet? We have the similar idea but I got mad for I can not come out any output. I would appreciate it if you did. Thank you again.
", "rawMarkdown": "If one pixel may has two or more class, why did you choose sigmoid & bce instead of softmax & cce? And, as you have 46 categories, where is the background?\nWell, and a little beg(perhaps offensive), since the competition is over, could you please share your code of the Unet? We have the similar idea but I got mad for I can not come out any output. I would appreciate it if you did. Thank you again.", "votes": null }, { "id": "550016", "postDate": "06/11/2019 08:06:02", "content": "If you use sigmoid & bce, each output value has 0 to 1 value. And you can thresholding each value(for example, threshold=0.5). So one pixel can have multi class. If values in all channels are under threshold, you can consider this pixel as background.
\n\nAnd I'm sorry but I have many work to do, and my code needs many refactoring, so I cannot premise that I can share my code.
\n\nInstead, following repository may helps you:\nhttps://github.com/milesial/Pytorch-UNet
\n\nI didn't use this repository but this repository uses sigmoid & bce.
", "rawMarkdown": "If you use sigmoid & bce, each output value has 0 to 1 value. And you can thresholding each value(for example, threshold=0.5). So one pixel can have multi class. If values in all channels are under threshold, you can consider this pixel as background.\n\nAnd I'm sorry but I have many work to do, and my code needs many refactoring, so I cannot premise that I can share my code.\n\nInstead, following repository may helps you:\nhttps://github.com/milesial/Pytorch-UNet\n\nI didn't use this repository but this repository uses sigmoid & bce.", "votes": null }, { "id": "550023", "postDate": "06/11/2019 08:20:22", "content": "Thanks for sharing!
", "rawMarkdown": "Thanks for sharing!", "votes": null }, { "id": "550681", "postDate": "06/11/2019 23:09:45", "content": "Not actually, If you are doing mask refinement. training 5 folds DenseNet121->FPN (Not exactly Unet, but still segmentation) takes 1h-24h on two 1080Ti(depending from class), with average of 2hours per 5 folds per class. A lot of classes have to few predicted classes to be trained. So if you have 1 1080 it will take ~100-200hours to train all them. Not a zero time, but there is nothing that can not be done with one gpu or kaggle/collab kernels.
", "rawMarkdown": "Not actually, If you are doing mask refinement. training 5 folds DenseNet121->FPN (Not exactly Unet, but still segmentation) takes 1h-24h on two 1080Ti(depending from class), with average of 2hours per 5 folds per class. A lot of classes have to few predicted classes to be trained. So if you have 1 1080 it will take ~100-200hours to train all them. Not a zero time, but there is nothing that can not be done with one gpu or kaggle/collab kernels.", "votes": null } ], "comments": [ { "id": 549865, "author_name": "venheads", "author_url": "", "post_date": "06/11/2019 05:20:23", "content": "I was using a lot of Unet with different encoders (actually more than 100, 2-3almost for each class) \nStarted with them to check submission, they had no good score. But they could be used to refine an outputted mask
", "votes": null, "replies": [ { "id": 549879, "author_name": "leonshangguan", "author_url": "", "post_date": "06/11/2019 05:32:03", "content": "Sorry for that, Could you explain what do you mean by > actually more than 100, 2-3almost for each class? Does that mean you regard it as a 47 bi-classification & segmentation?
", "votes": null, "replies": [] }, { "id": 549884, "author_name": "venheads", "author_url": "", "post_date": "06/11/2019 05:39:39", "content": "You are correct\nModel for each class, different models as well
", "votes": null, "replies": [] }, { "id": 549917, "author_name": "leonshangguan", "author_url": "", "post_date": "06/11/2019 06:25:24", "content": "Seems that demands excellent GPU...
", "votes": null, "replies": [] }, { "id": 549922, "author_name": "venheads", "author_url": "", "post_date": "06/11/2019 06:28:16", "content": "DGX-1 in my control
", "votes": null, "replies": [] }, { "id": 550681, "author_name": "realityfaker", "author_url": "", "post_date": "06/11/2019 23:09:45", "content": "Not actually, If you are doing mask refinement. training 5 folds DenseNet121->FPN (Not exactly Unet, but still segmentation) takes 1h-24h on two 1080Ti(depending from class), with average of 2hours per 5 folds per class. A lot of classes have to few predicted classes to be trained. So if you have 1 1080 it will take ~100-200hours to train all them. Not a zero time, but there is nothing that can not be done with one gpu or kaggle/collab kernels.
", "votes": null, "replies": [] } ] }, { "id": 549887, "author_name": "rskmoi", "author_url": "", "post_date": "06/11/2019 05:44:29", "content": "I used U-Net at the beginning of this competition. \ninput: 768 * 768 * 3, output: 768 * 768 * 46. \nOutput is activated by sigmoid and trained by BCELoss because one pixel may has two or more class.\nAfter some postprocessing, my U-Net best score is 0.10007(private)/0.10582(public), so I decided to use mask-rcnn.
", "votes": null, "replies": [ { "id": 549928, "author_name": "leonshangguan", "author_url": "", "post_date": "06/11/2019 06:34:58", "content": "If one pixel may has two or more class, why did you choose sigmoid & bce instead of softmax & cce? And, as you have 46 categories, where is the background?\nWell, and a little beg(perhaps offensive), since the competition is over, could you please share your code of the Unet? We have the similar idea but I got mad for I can not come out any output. I would appreciate it if you did. Thank you again.
", "votes": null, "replies": [] }, { "id": 550016, "author_name": "rskmoi", "author_url": "", "post_date": "06/11/2019 08:06:02", "content": "If you use sigmoid & bce, each output value has 0 to 1 value. And you can thresholding each value(for example, threshold=0.5). So one pixel can have multi class. If values in all channels are under threshold, you can consider this pixel as background.
\n\nAnd I'm sorry but I have many work to do, and my code needs many refactoring, so I cannot premise that I can share my code.
\n\nInstead, following repository may helps you:\nhttps://github.com/milesial/Pytorch-UNet
\n\nI didn't use this repository but this repository uses sigmoid & bce.
", "votes": null, "replies": [] }, { "id": 550023, "author_name": "leonshangguan", "author_url": "", "post_date": "06/11/2019 08:20:22", "content": "Thanks for sharing!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "549851": "As the competition is over, I think it's the time to ask this question.\nI tried to use U-net in this competition, since I used it to do binary classification and segmentation before. However I faced a lot of difficulties to do a multiple classification and segmentation. \nMy solution is to transfer the input mask into a [-1, 224, 224, 47] np.array, code like this: `sub_mask = cv2.resize(sub_mask, (SIZE, SIZE), interpolation=cv2.INTER_NEAREST)\n mask[:, :, int(label)] = sub_mask`, where 47 means 47 types(including the background), then when doing inference, I don't know how to choose the channel, and actually \nI think the channel is not matching the label.\nSo, as a beginner, I am really curious if anyone who used U-net in this competition and realized the multiple classification and segmentation. \nI will appreciate it if you can share the code or even just give some hints. Thank you.\n\nAdditional: I'm using keras and tf.", "549865": "I was using a lot of Unet with different encoders (actually more than 100, 2-3almost for each class) \nStarted with them to check submission, they had no good score. But they could be used to refine an outputted mask", "549879": "Sorry for that, Could you explain what do you mean by > actually more than 100, 2-3almost for each class? Does that mean you regard it as a 47 bi-classification & segmentation?", "549884": "You are correct\nModel for each class, different models as well", "549887": "I used U-Net at the beginning of this competition. \ninput: 768 * 768 * 3, output: 768 * 768 * 46. \nOutput is activated by sigmoid and trained by BCELoss because one pixel may has two or more class.\nAfter some postprocessing, my U-Net best score is 0.10007(private)/0.10582(public), so I decided to use mask-rcnn.", "549917": "Seems that demands excellent GPU...", "549922": "DGX-1 in my control", "549928": "If one pixel may has two or more class, why did you choose sigmoid & bce instead of softmax & cce? And, as you have 46 categories, where is the background?\nWell, and a little beg(perhaps offensive), since the competition is over, could you please share your code of the Unet? We have the similar idea but I got mad for I can not come out any output. I would appreciate it if you did. Thank you again.", "550016": "If you use sigmoid & bce, each output value has 0 to 1 value. And you can thresholding each value(for example, threshold=0.5). So one pixel can have multi class. If values in all channels are under threshold, you can consider this pixel as background.\n\nAnd I'm sorry but I have many work to do, and my code needs many refactoring, so I cannot premise that I can share my code.\n\nInstead, following repository may helps you:\nhttps://github.com/milesial/Pytorch-UNet\n\nI didn't use this repository but this repository uses sigmoid & bce.", "550023": "Thanks for sharing!", "550681": "Not actually, If you are doing mask refinement. training 5 folds DenseNet121->FPN (Not exactly Unet, but still segmentation) takes 1h-24h on two 1080Ti(depending from class), with average of 2hours per 5 folds per class. A lot of classes have to few predicted classes to be trained. So if you have 1 1080 it will take ~100-200hours to train all them. Not a zero time, but there is nothing that can not be done with one gpu or kaggle/collab kernels." }, "source": "meta" }