Does anyone try to train separate model for each cell class? Is this working approach?
", "messages": [ { "id": 1601737, "postDate": "2021-12-01T14:48:58.953Z", "content": "Does anyone try to train separate model for each cell class? Is this working approach?
", "rawMarkdown": "Does anyone try to train separate model for each cell class? Is this working approach?", "votes": 8 }, { "id": 1603931, "postDate": "2021-12-03T01:03:23.750Z", "content": "I am doing that at the moment actually! At the moment the version I have developed has strong Iou and F1 scores (0.81, 0.90) on the validation dataset thresholded at 0.95, have yet to figure out how to ensemble the models, but training the classes separately gave promising results. I've come to realize that the data provided has vastly different masks for visually similar features, so the best way to mitigate this issue would be to train the models separately and then decide when model to use. I'd love to hear what others have done to combat these issues.
", "rawMarkdown": "I am doing that at the moment actually! At the moment the version I have developed has strong Iou and F1 scores (0.81, 0.90) on the validation dataset thresholded at 0.95, have yet to figure out how to ensemble the models, but training the classes separately gave promising results. I've come to realize that the data provided has vastly different masks for *visually* similar features, so the best way to mitigate this issue would be to train the models separately and then decide when model to use. I'd love to hear what others have done to combat these issues. \n", "votes": 1 }, { "id": 1610389, "postDate": "2021-12-07T07:11:41.857Z", "content": "Not yet train seperate models,I've been digging into feature mining,trying to provide sound reasons to build a spicific cnn architechture.
\nThe cort is not the hard one. The sysh5y challenges with its density and non-isometric and tiny shape.The astro is the boss which I'm not sure if the existing cnn things can even beat.
\nI figured out that although the cells are zoomed to be visible to human eyes in pictures.It does not mean the features are equivalently generalized with those coco 80 classes of daily objects. The objects are similar in a CV perspect however they differ like day and night physically just in the m vs nm scale.
\nThough the mask rcnn weights can be finetuned, the system itself can overfit too.Does anyone have similar thoughts or some more inspiring ideas?
Yes, I trained separate models
", "rawMarkdown": "Yes, I trained separate models", "isDeleted": true }, { "id": 1605630, "postDate": "2021-12-04T12:14:20.513Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 1603931, "author_name": "Krithik Ramesh", "author_url": "", "post_date": "2021-12-03T01:03:23.750000", "content": "I am doing that at the moment actually! At the moment the version I have developed has strong Iou and F1 scores (0.81, 0.90) on the validation dataset thresholded at 0.95, have yet to figure out how to ensemble the models, but training the classes separately gave promising results. I've come to realize that the data provided has vastly different masks for visually similar features, so the best way to mitigate this issue would be to train the models separately and then decide when model to use. I'd love to hear what others have done to combat these issues.
", "votes": 1, "replies": [] }, { "id": 1610389, "author_name": "Lupin", "author_url": "", "post_date": "2021-12-07T07:11:41.857000", "content": "Not yet train seperate models,I've been digging into feature mining,trying to provide sound reasons to build a spicific cnn architechture.
\nThe cort is not the hard one. The sysh5y challenges with its density and non-isometric and tiny shape.The astro is the boss which I'm not sure if the existing cnn things can even beat.
\nI figured out that although the cells are zoomed to be visible to human eyes in pictures.It does not mean the features are equivalently generalized with those coco 80 classes of daily objects. The objects are similar in a CV perspect however they differ like day and night physically just in the m vs nm scale.
\nThough the mask rcnn weights can be finetuned, the system itself can overfit too.Does anyone have similar thoughts or some more inspiring ideas?
Yes, I trained separate models
", "votes": 0, "replies": [] }, { "id": 1605630, "author_name": "", "author_url": "", "post_date": "2021-12-04T12:14:20.513000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1601737": "Does anyone try to train separate model for each cell class? Is this working approach?", "1603931": "I am doing that at the moment actually! At the moment the version I have developed has strong Iou and F1 scores (0.81, 0.90) on the validation dataset thresholded at 0.95, have yet to figure out how to ensemble the models, but training the classes separately gave promising results. I've come to realize that the data provided has vastly different masks for *visually* similar features, so the best way to mitigate this issue would be to train the models separately and then decide when model to use. I'd love to hear what others have done to combat these issues. \n", "1610389": "Not yet train seperate models,I've been digging into feature mining,trying to provide sound reasons to build a spicific cnn architechture.\nThe cort is not the hard one. The sysh5y challenges with its density and non-isometric and tiny shape.The astro is the boss which I'm not sure if the existing cnn things can even beat.\nI figured out that although the cells are zoomed to be visible to human eyes in pictures.It does not mean the features are equivalently generalized with those coco 80 classes of daily objects. The objects are similar in a CV perspect however they differ like day and night physically just in the m vs nm scale.\nThough the mask rcnn weights can be finetuned, the system itself can overfit too.Does anyone have similar thoughts or some more inspiring ideas?", "1605827": "Yes, I trained separate models", "1605630": "" } }