{ "id": 464746, "title": "Why does the model not converge for WSI image classification in multi-instance processing? The training acc is always equal to about 0.4", "url": "/competitions/UBC-OCEAN/discussion/464746", "author_name": "Metavers", "post_date": "2024-01-01T10:05:09.038000", "votes": -4, "comment_count": 6, "views": 0, "content": "

I first used CNN to extract feature maps. The obtained more than ten pictures were formed into a batch as a package, and then HW and B were pulled into a sequence. Input to encoder to learn which tiles are most important, but it does not converge and I don’t know what went wrong.\"\"\"\"

", "messages": [ { "id": 2582022, "postDate": "2024-01-01T10:05:09.037Z", "content": "

I first used CNN to extract feature maps. The obtained more than ten pictures were formed into a batch as a package, and then HW and B were pulled into a sequence. Input to encoder to learn which tiles are most important, but it does not converge and I don’t know what went wrong.\"\"\"\"

", "rawMarkdown": "I first used CNN to extract feature maps. The obtained more than ten pictures were formed into a batch as a package, and then HW and B were pulled into a sequence. Input to encoder to learn which tiles are most important, but it does not converge and I don’t know what went wrong.![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F16846649%2F4b641f72e5b1510142df738a7c3712d2%2F_20240101180417.png?generation=1704103500373088&alt=media)![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F16846649%2F71358bfdcd4ad3b0fb5d1bc83f77394b%2F_20240101180422.png?generation=1704103505073155&alt=media)", "votes": -4 }, { "id": 2582034, "postDate": "2024-01-01T10:15:25.850Z", "content": "

There is no problem with positional encoding, separator is used as the separator for each batch of images

", "rawMarkdown": "There is no problem with positional encoding, separator is used as the separator for each batch of images", "votes": -3, "replies": [ { "id": 2583837, "postDate": "2024-01-02T14:10:24.573Z", "content": "

为啥你认为在病理切片里位置编码是需要的？相反，我认为因为在预处理过程中不同patch的位置关系已经完全丢失，加入位置编码是有害的

", "rawMarkdown": "为啥你认为在病理切片里位置编码是需要的？相反，我认为因为在预处理过程中不同patch的位置关系已经完全丢失，加入位置编码是有害的", "votes": -1, "replies": [ { "id": 2585036, "postDate": "2024-01-03T10:13:02.963Z", "content": "

但是没有位置编码也一样啊 ，就是收敛不了 所以我严重怀疑我的聚合方法是不是出错了

", "rawMarkdown": "但是没有位置编码也一样啊 ，就是收敛不了 所以我严重怀疑我的聚合方法是不是出错了 ", "votes": -1 } ] } ] }, { "id": 2582023, "postDate": "2024-01-01T10:06:19.767Z", "content": "

大佬求教Ask the boss for advice

", "rawMarkdown": "大佬求教Ask the boss for advice", "votes": -3 }, { "id": 2585372, "postDate": "2024-01-03T14:28:26.227Z", "content": "

我好像知道为什么收敛不了了，用多实例去做的基本都没有用缩小后的图像的，而且是256256的未被缩小的切片去做的，也是就是大的分辨率可以观察到细胞核结构的分辨率去做的，而不是以20482048压缩成512512px去做的识别\nI seem to know why it cannot converge. Basically, the reduced images are not used when using multiple instances, and they are done with 256256 non-reduced slices, which means that the structure of the cell nucleus can be observed with a large resolution. resolution, rather than compressing 20482048 into 512512px for recognition

", "rawMarkdown": "我好像知道为什么收敛不了了，用多实例去做的基本都没有用缩小后的图像的，而且是256*256的未被缩小的切片去做的，也是就是大的分辨率可以观察到细胞核结构的分辨率去做的，而不是以2048*2048压缩成512*512px去做的识别\nI seem to know why it cannot converge. Basically, the reduced images are not used when using multiple instances, and they are done with 256*256 non-reduced slices, which means that the structure of the cell nucleus can be observed with a large resolution. resolution, rather than compressing 2048*2048 into 512*512px for recognition\n" }, { "id": 2585180, "postDate": "2024-01-03T12:24:28.727Z", "content": "

Why click and thumbs down? If you have the skills, tell me what the problem is. Why? Are you not used to Chinese comments made by Chinese people?点踩干嘛，你有本事你说说问题出在哪了，咋滴看不惯中国人中文注释吗

", "rawMarkdown": "Why click and thumbs down? If you have the skills, tell me what the problem is. Why? Are you not used to Chinese comments made by Chinese people?点踩干嘛，你有本事你说说问题出在哪了，咋滴看不惯中国人中文注释吗" } ], "comments": [ { "id": 2582034, "author_name": "Metavers", "author_url": "", "post_date": "2024-01-01T10:15:25.850000", "content": "

There is no problem with positional encoding, separator is used as the separator for each batch of images

", "votes": -3, "replies": [ { "id": 2583837, "author_name": "ForcewithMe", "author_url": "", "post_date": "2024-01-02T14:10:24.573000", "content": "

为啥你认为在病理切片里位置编码是需要的？相反，我认为因为在预处理过程中不同patch的位置关系已经完全丢失，加入位置编码是有害的

", "votes": -1, "replies": [ { "id": 2585036, "author_name": "Metavers", "author_url": "", "post_date": "2024-01-03T10:13:02.963000", "content": "

但是没有位置编码也一样啊 ，就是收敛不了 所以我严重怀疑我的聚合方法是不是出错了

", "votes": -1, "replies": [] } ] } ] }, { "id": 2582023, "author_name": "Metavers", "author_url": "", "post_date": "2024-01-01T10:06:19.767000", "content": "

大佬求教Ask the boss for advice

", "votes": -3, "replies": [] }, { "id": 2585372, "author_name": "Metavers", "author_url": "", "post_date": "2024-01-03T14:28:26.227000", "content": "

我好像知道为什么收敛不了了，用多实例去做的基本都没有用缩小后的图像的，而且是256256的未被缩小的切片去做的，也是就是大的分辨率可以观察到细胞核结构的分辨率去做的，而不是以20482048压缩成512512px去做的识别\nI seem to know why it cannot converge. Basically, the reduced images are not used when using multiple instances, and they are done with 256256 non-reduced slices, which means that the structure of the cell nucleus can be observed with a large resolution. resolution, rather than compressing 20482048 into 512512px for recognition

", "votes": 0, "replies": [] }, { "id": 2585180, "author_name": "Metavers", "author_url": "", "post_date": "2024-01-03T12:24:28.727000", "content": "

Why click and thumbs down? If you have the skills, tell me what the problem is. Why? Are you not used to Chinese comments made by Chinese people?点踩干嘛，你有本事你说说问题出在哪了，咋滴看不惯中国人中文注释吗

", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "2582022": "I first used CNN to extract feature maps. The obtained more than ten pictures were formed into a batch as a package, and then HW and B were pulled into a sequence. Input to encoder to learn which tiles are most important, but it does not converge and I don’t know what went wrong.![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F16846649%2F4b641f72e5b1510142df738a7c3712d2%2F_20240101180417.png?generation=1704103500373088&alt=media)![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F16846649%2F71358bfdcd4ad3b0fb5d1bc83f77394b%2F_20240101180422.png?generation=1704103505073155&alt=media)", "2582034": "There is no problem with positional encoding, separator is used as the separator for each batch of images", "2582023": "大佬求教Ask the boss for advice", "2585372": "我好像知道为什么收敛不了了，用多实例去做的基本都没有用缩小后的图像的，而且是256*256的未被缩小的切片去做的，也是就是大的分辨率可以观察到细胞核结构的分辨率去做的，而不是以2048*2048压缩成512*512px去做的识别\nI seem to know why it cannot converge. Basically, the reduced images are not used when using multiple instances, and they are done with 256*256 non-reduced slices, which means that the structure of the cell nucleus can be observed with a large resolution. resolution, rather than compressing 2048*2048 into 512*512px for recognition\n", "2585180": "Why click and thumbs down? If you have the skills, tell me what the problem is. Why? Are you not used to Chinese comments made by Chinese people?点踩干嘛，你有本事你说说问题出在哪了，咋滴看不惯中国人中文注释吗" } }