![\"enter](\"https://storage.googleapis.com/kaggle-forum-message-attachments/311223/9045/0f1f896d9ae5a04752d3239c690402c022db4d72c0d2c087d73380896f72c466.png\")
\n ![\"enter](\"https://storage.googleapis.com/kaggle-forum-message-attachments/311223/9049/c8e79ff4ac55f4b772057de28e539727b7f4f2a3de73bf7a082a0ace86d609eb.png\")
\n ![\"enter](\"https://storage.googleapis.com/kaggle-forum-message-attachments/311223/9056/44afae184c89e6ba55985b4d341acc1ae1e8b6ef96312064e0e6e630e022b078.png\")
argumentation i used.
\n\nthe ensembled results
\n\n\n \n
argumentation i used.
\n\nthe ensembled results
\n\n\n \n
What is your LB on color only? I got 0.063 LB with color set and without stain normalization
", "rawMarkdown": "What is your LB on color only? I got 0.063 LB with color set and without stain normalization", "votes": null }, { "id": "311281", "postDate": "04/09/2018 18:38:07", "content": "Wow, we did very badly on this image, so I guess most of the colour images are in the private LB then. @Heng, I remember you trained external colour images too?
", "rawMarkdown": "Wow, we did very badly on this image, so I guess most of the colour images are in the private LB then. @Heng, I remember you trained external colour images too?", "votes": null }, { "id": "311288", "postDate": "04/09/2018 18:51:44", "content": "Very Nice !
\n\nA small remark, don't you think that the small yellow particles in the second row are also nuclei ? \nI think H&E DAB stains have been used in this case, and the color are a bit distorted.
", "rawMarkdown": "Very Nice ! \n\nA small remark, don't you think that the small yellow particles in the second row are also nuclei ? \nI think H&E DAB stains have been used in this case, and the color are a bit distorted.", "votes": null }, { "id": "311396", "postDate": "04/10/2018 01:26:00", "content": "the results shown here is trained without external data. Currently, external data training is still on going
", "rawMarkdown": "the results shown here is trained without external data. Currently, external data training is still on going", "votes": null }, { "id": "311609", "postDate": "04/10/2018 11:07:26", "content": "Impressive!
", "rawMarkdown": "Impressive!", "votes": null }, { "id": "311936", "postDate": "04/10/2018 23:44:27", "content": "Although, I much wonder how your nets perform on different colors. Inverted colors, different color images, etc. Because, they said that the test set is gonna be hell lotta different. And I'm getting extremely different results whether I train on different synthetic colors, or not. If I don't I get a cool score on data similar to what I have at my disposal. Much better than if I do heavy color augmentation. But if I tweak that data, like inverting it, then the results suck and my networks that were trained on heavy augmentation, although they suck in general, are incomparably better on that kind of data.
", "rawMarkdown": "Although, I much wonder how your nets perform on different colors. Inverted colors, different color images, etc. Because, they said that the test set is gonna be hell lotta different. And I'm getting extremely different results whether I train on different synthetic colors, or not. If I don't I get a cool score on data similar to what I have at my disposal. Much better than if I do heavy color augmentation. But if I tweak that data, like inverting it, then the results suck and my networks that were trained on heavy augmentation, although they suck in general, are incomparably better on that kind of data.", "votes": null }, { "id": "311946", "postDate": "04/11/2018 00:37:03", "content": "here is the results of the difficult images using external train data (single model)
\n\n![\"enter](\"https://storage.googleapis.com/kaggle-forum-message-attachments/311946/9073/259b35151d4a7a5ffdd7ab7f171b142db8cfe40beeee67277fac6adca4d042c4.png\")
![\"enter](\"https://storage.googleapis.com/kaggle-forum-message-attachments/311946/9074/17b9bf4356db24967c4677b8376ac38f826de73a88b93a8d73a8b452e399cdff.png\")
The color result looks fine. Have you merger gray and color into 1 set and train it? Do you convert color to gray ? I trained on single model but i got only 0.063 lb on color set
", "rawMarkdown": "The color result looks fine. Have you merger gray and color into 1 set and train it? Do you convert color to gray ? I trained on single model but i got only 0.063 lb on color set", "votes": null }, { "id": "311984", "postDate": "04/11/2018 02:45:16", "content": "Yeah, I also wasn't able to save my nets from annotating that round thing at the bottom. The annotations look really great, hovewer I wouldn't exactly call this a difficult image, as it's not really all that different from the training samples. How does it process this image with inverted colors, for example? I don't know what the actual testing data is gonna be, but I've found quite a bunch of scary images on the net just using image search, so I went hard for the color augmentation. Though with this precision, I'm guessing your net might fare fine even on those and nucleuses at vastly different scale. You've done your homework ;)
", "rawMarkdown": "Yeah, I also wasn't able to save my nets from annotating that round thing at the bottom. The annotations look really great, hovewer I wouldn't exactly call this a difficult image, as it's not really all that different from the training samples. How does it process this image with inverted colors, for example? I don't know what the actual testing data is gonna be, but I've found quite a bunch of scary images on the net just using image search, so I went hard for the color augmentation. Though with this precision, I'm guessing your net might fare fine even on those and nucleuses at vastly different scale. You've done your homework ;)", "votes": null }, { "id": "311986", "postDate": "04/11/2018 02:53:34", "content": "If you create synthetic data like it you can get rid of the dot. But the private lb is a different data set, so it may not be worth to overfit public lb
", "rawMarkdown": "If you create synthetic data like it you can get rid of the dot. But the private lb is a different data set, so it may not be worth to overfit public lb", "votes": null }, { "id": "311993", "postDate": "04/11/2018 03:05:58", "content": "Well, I see it more like a symptom saying something like that the context info doesn't travel that far, or isn't well understood by the net. If you don't care about that, you can actually take care of it during post-processing, I think, as it has a way different hue, saturation and brightness vs the rest of the pack. If you don't want to leave that job to the all-mighty net.
", "rawMarkdown": "Well, I see it more like a symptom saying something like that the context info doesn't travel that far, or isn't well understood by the net. If you don't care about that, you can actually take care of it during post-processing, I think, as it has a way different hue, saturation and brightness vs the rest of the pack. If you don't want to leave that job to the all-mighty net.", "votes": null }, { "id": "312090", "postDate": "04/11/2018 07:07:08", "content": "test results on external images (not in training).
\n\nFor the remaining time, i am testing as many images as i can, then hand correct the images and feed back to training.
\n\n![\"enter](\"https://storage.googleapis.com/kaggle-forum-message-attachments/312090/9077/00001.png\")
\n ![\"enter](\"https://storage.googleapis.com/kaggle-forum-message-attachments/312090/9079/00005.png\")
works on other color too! These testing images are not in training.\nI can tell the prediction is wrong with there is high percentage of uncertain pixels (e.g. pixels with prob <0.5 is high)
\n\n![\"enter](\"https://storage.googleapis.com/kaggle-forum-message-attachments/312093/9081/00006.png\")
\n ![\"enter](\"https://storage.googleapis.com/kaggle-forum-message-attachments/312093/9082/xx00.png\")
\n ![\"enter](\"https://storage.googleapis.com/kaggle-forum-message-attachments/312093/9083/xx01.png\")
what is your external images. These images do not have gt mask?
", "rawMarkdown": "what is your external images. These images do not have gt mask?", "votes": null }, { "id": "312130", "postDate": "04/11/2018 08:38:35", "content": "data similar to the round dot false positive
\n\n![\"enter](\"https://storage.googleapis.com/kaggle-forum-message-attachments/312130/9084/yy00.png\")
I'm not very familiar with the rules on manual annotation, so I'm curious about your process of fixing images and feeding them back to training. Doesn't manually annotating images fall under the category of external training data? And if so, it must be shared on the external data thread?
", "rawMarkdown": "I'm not very familiar with the rules on manual annotation, so I'm curious about your process of fixing images and feeding them back to training. Doesn't manually annotating images fall under the category of external training data? And if so, it must be shared on the external data thread?", "votes": null } ], "comments": [ { "id": 311241, "author_name": "john1231983", "author_url": "", "post_date": "04/09/2018 17:14:11", "content": "What is your LB on color only? I got 0.063 LB with color set and without stain normalization
", "votes": null, "replies": [] }, { "id": 311281, "author_name": "nicolefinnie", "author_url": "", "post_date": "04/09/2018 18:38:07", "content": "Wow, we did very badly on this image, so I guess most of the colour images are in the private LB then. @Heng, I remember you trained external colour images too?
", "votes": null, "replies": [ { "id": 311396, "author_name": "hengck23", "author_url": "", "post_date": "04/10/2018 01:26:00", "content": "the results shown here is trained without external data. Currently, external data training is still on going
", "votes": null, "replies": [] } ] }, { "id": 311288, "author_name": "jeandebleau", "author_url": "", "post_date": "04/09/2018 18:51:44", "content": "Very Nice !
\n\nA small remark, don't you think that the small yellow particles in the second row are also nuclei ? \nI think H&E DAB stains have been used in this case, and the color are a bit distorted.
", "votes": null, "replies": [] }, { "id": 311609, "author_name": "thenickname", "author_url": "", "post_date": "04/10/2018 11:07:26", "content": "Impressive!
", "votes": null, "replies": [ { "id": 311936, "author_name": "thenickname", "author_url": "", "post_date": "04/10/2018 23:44:27", "content": "Although, I much wonder how your nets perform on different colors. Inverted colors, different color images, etc. Because, they said that the test set is gonna be hell lotta different. And I'm getting extremely different results whether I train on different synthetic colors, or not. If I don't I get a cool score on data similar to what I have at my disposal. Much better than if I do heavy color augmentation. But if I tweak that data, like inverting it, then the results suck and my networks that were trained on heavy augmentation, although they suck in general, are incomparably better on that kind of data.
", "votes": null, "replies": [] } ] }, { "id": 311946, "author_name": "hengck23", "author_url": "", "post_date": "04/11/2018 00:37:03", "content": "here is the results of the difficult images using external train data (single model)
\n\n
The color result looks fine. Have you merger gray and color into 1 set and train it? Do you convert color to gray ? I trained on single model but i got only 0.063 lb on color set
", "votes": null, "replies": [] }, { "id": 311984, "author_name": "thenickname", "author_url": "", "post_date": "04/11/2018 02:45:16", "content": "Yeah, I also wasn't able to save my nets from annotating that round thing at the bottom. The annotations look really great, hovewer I wouldn't exactly call this a difficult image, as it's not really all that different from the training samples. How does it process this image with inverted colors, for example? I don't know what the actual testing data is gonna be, but I've found quite a bunch of scary images on the net just using image search, so I went hard for the color augmentation. Though with this precision, I'm guessing your net might fare fine even on those and nucleuses at vastly different scale. You've done your homework ;)
", "votes": null, "replies": [] }, { "id": 311986, "author_name": "hengck23", "author_url": "", "post_date": "04/11/2018 02:53:34", "content": "If you create synthetic data like it you can get rid of the dot. But the private lb is a different data set, so it may not be worth to overfit public lb
", "votes": null, "replies": [] }, { "id": 311993, "author_name": "thenickname", "author_url": "", "post_date": "04/11/2018 03:05:58", "content": "Well, I see it more like a symptom saying something like that the context info doesn't travel that far, or isn't well understood by the net. If you don't care about that, you can actually take care of it during post-processing, I think, as it has a way different hue, saturation and brightness vs the rest of the pack. If you don't want to leave that job to the all-mighty net.
", "votes": null, "replies": [] } ] }, { "id": 312090, "author_name": "hengck23", "author_url": "", "post_date": "04/11/2018 07:07:08", "content": "test results on external images (not in training).
\n\nFor the remaining time, i am testing as many images as i can, then hand correct the images and feed back to training.
\n\n\n
I'm not very familiar with the rules on manual annotation, so I'm curious about your process of fixing images and feeding them back to training. Doesn't manually annotating images fall under the category of external training data? And if so, it must be shared on the external data thread?
", "votes": null, "replies": [] } ] }, { "id": 312093, "author_name": "hengck23", "author_url": "", "post_date": "04/11/2018 07:13:32", "content": "works on other color too! These testing images are not in training.\nI can tell the prediction is wrong with there is high percentage of uncertain pixels (e.g. pixels with prob <0.5 is high)
\n\n\n \n
what is your external images. These images do not have gt mask?
", "votes": null, "replies": [] } ] }, { "id": 312130, "author_name": "hengck23", "author_url": "", "post_date": "04/11/2018 08:38:35", "content": "data similar to the round dot false positive
\n\n