Please post your suggestion to overcome the stagnant loss\nmodel used: resnext 50 32 4d facebook \nImages= Tiled (224)\nCV 0.8 or less
", "messages": [ { "id": 844546, "postDate": "2020-05-12T17:07:57.857Z", "content": "Please post your suggestion to overcome the stagnant loss\nmodel used: resnext 50 32 4d facebook \nImages= Tiled (224)\nCV 0.8 or less
", "rawMarkdown": "Please post your suggestion to overcome the stagnant loss\nmodel used: resnext 50 32 4d facebook \nImages= Tiled (224)\nCV 0.8 or less", "votes": 3 }, { "id": 845812, "postDate": "2020-05-13T12:48:37.670Z", "content": "There are some things you can try:\nStart with:\n1) Augment images by rotate them\n2) Check some image visualization after tiles, be sure that you catch almost all the tissues in the tiles you select \n3) Use TTA\n4) Change head structure
", "rawMarkdown": "There are some things you can try:\nStart with:\n1) Augment images by rotate them\n2) Check some image visualization after tiles, be sure that you catch almost all the tissues in the tiles you select \n3) Use TTA\n4) Change head structure", "votes": 1, "replies": [ { "id": 845878, "postDate": "2020-05-13T13:25:13.247Z", "content": "@vladvdv \nthanks for tips\n1) I currently use mix concat and avg in head. \nany suggestion for better changes
\n\n2) how about loss i use MSE only for regression is there any better option.
", "rawMarkdown": "@vladvdv \nthanks for tips\n1) I currently use mix concat and avg in head. \nany suggestion for better changes\n\n2) how about loss i use MSE only for regression is there any better option.\n" }, { "id": 845938, "postDate": "2020-05-13T13:50:25.403Z", "content": "I've tried smooth L1 loss which acts like L2 with small losses and L1 with big losses. I've got slightly better results
", "rawMarkdown": "I've tried smooth L1 loss which acts like L2 with small losses and L1 with big losses. I've got slightly better results" }, { "id": 861655, "postDate": "2020-05-26T06:38:47.120Z", "content": "L1 loss and its other variants works worst for me . not sure why
", "rawMarkdown": " L1 loss and its other variants works worst for me . not sure why\n" } ] }, { "id": 845438, "postDate": "2020-05-13T07:57:12.100Z", "content": "Same issue. I experience over-fitting from the 30th epoch\nI'm thinking on combining classification and regression in the same model. The final loss would be something like CE + MSE\nBTW how many tiles do you use per slide ? I'm using a tiling pretty close to @iafoss kernel but I'm thinking on trying bigger size
", "rawMarkdown": "Same issue. I experience over-fitting from the 30th epoch\nI'm thinking on combining classification and regression in the same model. The final loss would be something like CE + MSE\nBTW how many tiles do you use per slide ? I'm using a tiling pretty close to @iafoss kernel but I'm thinking on trying bigger size ", "votes": 2, "replies": [ { "id": 845468, "postDate": "2020-05-13T08:12:50.160Z", "content": "m using regression method and using 16 tiles . \n 1)Same thing happens to me i start seeing drastic overfitting once loss comes below 0.8 MSE loss\n2) what is best cv qkp you got ,is it matching with your lb score ?
", "rawMarkdown": "m using regression method and using 16 tiles . \n 1)Same thing happens to me i start seeing drastic overfitting once loss comes below 0.8 MSE loss\n2) what is best cv qkp you got ,is it matching with your lb score ?\n" }, { "id": 845505, "postDate": "2020-05-13T08:39:23.043Z", "content": "Best scores I've got are with classification : 0.80 val => 0.84 LB with TTA\nRegression scores are closer to LB than classification ones\nNote that I'm using single fold due to GPU quota
", "rawMarkdown": "Best scores I've got are with classification : 0.80 val => 0.84 LB with TTA\nRegression scores are closer to LB than classification ones\nNote that I'm using single fold due to GPU quota", "votes": 1 }, { "id": 845565, "postDate": "2020-05-13T09:30:15.710Z", "content": "any idea how many images could be there in the test set used for calculating lb.
", "rawMarkdown": "any idea how many images could be there in the test set used for calculating lb.\n" } ] }, { "id": 846255, "postDate": "2020-05-13T16:30:38.560Z", "content": "For me going to level 1 images improved my score by a lot, otherwise i was stuck at 0.78 lb with regression
", "rawMarkdown": "For me going to level 1 images improved my score by a lot, otherwise i was stuck at 0.78 lb with regression" }, { "id": 845419, "postDate": "2020-05-13T07:50:35.400Z", "content": "Well, you have to be more precise if you want an pertinent opinion.\nStart with what approach are you using, classification, regression ?\nThen with how many tiles of that dimensions are you using ? And from what image level ? Did you visualize the tiles for some images ? They seem to fit all the important tissue part ?\nWhat is the structure of the network head that you designed ?
", "rawMarkdown": "Well, you have to be more precise if you want an pertinent opinion.\nStart with what approach are you using, classification, regression ?\nThen with how many tiles of that dimensions are you using ? And from what image level ? Did you visualize the tiles for some images ? They seem to fit all the important tissue part ?\nWhat is the structure of the network head that you designed ? ", "replies": [ { "id": 845471, "postDate": "2020-05-13T08:13:53.680Z", "content": "Yes.. i have wrote more details. I use level 2 images,as level 3 as per public kernel score looks not so useful.
", "rawMarkdown": "Yes.. i have wrote more details. I use level 2 images,as level 3 as per public kernel score looks not so useful.\n" } ] } ], "comments": [ { "id": 845812, "author_name": "Vlad Vaduva", "author_url": "", "post_date": "2020-05-13T12:48:37.670000", "content": "There are some things you can try:\nStart with:\n1) Augment images by rotate them\n2) Check some image visualization after tiles, be sure that you catch almost all the tissues in the tiles you select \n3) Use TTA\n4) Change head structure
", "votes": 1, "replies": [ { "id": 845878, "author_name": "Jaideep", "author_url": "", "post_date": "2020-05-13T13:25:13.247000", "content": "@vladvdv \nthanks for tips\n1) I currently use mix concat and avg in head. \nany suggestion for better changes
\n\n2) how about loss i use MSE only for regression is there any better option.
", "votes": 0, "replies": [] }, { "id": 845938, "author_name": "Alex", "author_url": "", "post_date": "2020-05-13T13:50:25.403000", "content": "I've tried smooth L1 loss which acts like L2 with small losses and L1 with big losses. I've got slightly better results
", "votes": 0, "replies": [] }, { "id": 861655, "author_name": "Jaideep", "author_url": "", "post_date": "2020-05-26T06:38:47.120000", "content": "L1 loss and its other variants works worst for me . not sure why
", "votes": 0, "replies": [] } ] }, { "id": 845438, "author_name": "Alex", "author_url": "", "post_date": "2020-05-13T07:57:12.100000", "content": "Same issue. I experience over-fitting from the 30th epoch\nI'm thinking on combining classification and regression in the same model. The final loss would be something like CE + MSE\nBTW how many tiles do you use per slide ? I'm using a tiling pretty close to @iafoss kernel but I'm thinking on trying bigger size
", "votes": 2, "replies": [ { "id": 845468, "author_name": "Jaideep", "author_url": "", "post_date": "2020-05-13T08:12:50.160000", "content": "m using regression method and using 16 tiles . \n 1)Same thing happens to me i start seeing drastic overfitting once loss comes below 0.8 MSE loss\n2) what is best cv qkp you got ,is it matching with your lb score ?
", "votes": 0, "replies": [] }, { "id": 845505, "author_name": "Alex", "author_url": "", "post_date": "2020-05-13T08:39:23.043000", "content": "Best scores I've got are with classification : 0.80 val => 0.84 LB with TTA\nRegression scores are closer to LB than classification ones\nNote that I'm using single fold due to GPU quota
", "votes": 1, "replies": [] }, { "id": 845565, "author_name": "Jaideep", "author_url": "", "post_date": "2020-05-13T09:30:15.710000", "content": "any idea how many images could be there in the test set used for calculating lb.
", "votes": 0, "replies": [] } ] }, { "id": 846255, "author_name": "Yann Majewski", "author_url": "", "post_date": "2020-05-13T16:30:38.560000", "content": "For me going to level 1 images improved my score by a lot, otherwise i was stuck at 0.78 lb with regression
", "votes": 0, "replies": [] }, { "id": 845419, "author_name": "Vlad Vaduva", "author_url": "", "post_date": "2020-05-13T07:50:35.400000", "content": "Well, you have to be more precise if you want an pertinent opinion.\nStart with what approach are you using, classification, regression ?\nThen with how many tiles of that dimensions are you using ? And from what image level ? Did you visualize the tiles for some images ? They seem to fit all the important tissue part ?\nWhat is the structure of the network head that you designed ?
", "votes": 0, "replies": [ { "id": 845471, "author_name": "Jaideep", "author_url": "", "post_date": "2020-05-13T08:13:53.680000", "content": "Yes.. i have wrote more details. I use level 2 images,as level 3 as per public kernel score looks not so useful.
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "844546": "Please post your suggestion to overcome the stagnant loss\nmodel used: resnext 50 32 4d facebook \nImages= Tiled (224)\nCV 0.8 or less", "845812": "There are some things you can try:\nStart with:\n1) Augment images by rotate them\n2) Check some image visualization after tiles, be sure that you catch almost all the tissues in the tiles you select \n3) Use TTA\n4) Change head structure", "845438": "Same issue. I experience over-fitting from the 30th epoch\nI'm thinking on combining classification and regression in the same model. The final loss would be something like CE + MSE\nBTW how many tiles do you use per slide ? I'm using a tiling pretty close to @iafoss kernel but I'm thinking on trying bigger size ", "846255": "For me going to level 1 images improved my score by a lot, otherwise i was stuck at 0.78 lb with regression", "845419": "Well, you have to be more precise if you want an pertinent opinion.\nStart with what approach are you using, classification, regression ?\nThen with how many tiles of that dimensions are you using ? And from what image level ? Did you visualize the tiles for some images ? They seem to fit all the important tissue part ?\nWhat is the structure of the network head that you designed ? " } }