{ "id": 168825, "title": "Which backbone is the best?", "url": "/competitions/prostate-cancer-grade-assessment/discussion/168825", "author_name": "seefun", "post_date": "2020-07-22T03:27:09.166000", "votes": 0, "comment_count": 1, "views": 0, "content": "

I have tried two ResNet-50 level backbone (ResNet50-gn & tresnet-m) with smooth L1 regression, I can only get 0.88LB but 0.91CV. (TTA, k folds ensemble and model blending are not able to give me great improvement in LB, but CV increased)
\nLimited by computing resources, I can’t run more experiments, each experiment takes me a week.
\nDoes the choice of backbone matter? Welcome everyone to share the experimental results.

", "messages": [ { "id": 939286, "postDate": "2020-07-22T06:24:21.243Z", "content": "
    \n
  1. efn-b0 here, also limited by computing resources. 5-folds splited but I only choose 2 of 5 folds(2 hightest local score fold) and average them can get 0.888 lb(no tta).
    2. \n
  2. I try re-sampling the data to train a b0 these days but didn't get improvement.(actually it hurt local score and lb)
    3. \n
  3. because there are 2 data-providers, I also try feed full-img(after resize, no tiles) to model(a efn-b4) and training with reverse-gradient(use a extra head with 1 output to do binary classification), want to retard the difference between them, but the loss of binary-head become bery huge(can reach 10^2 magnitude), so I give up this way.
    4. \n
\n
\n

No more experiments, hope my average-model can bring me the good luck. wanna shake 亿点点 too😜

", "rawMarkdown": "1. efn-b0 here, also limited by computing resources. 5-folds splited but I only choose 2 of 5 folds(2 hightest local score fold) and average them can get 0.888 lb(no tta).\n2. I try re-sampling the data to train a b0 these days but didn't get improvement.(actually it hurt local score and lb)\n3. because there are 2 data-providers, I also try feed full-img(after resize, no tiles) to model(a efn-b4) and training with reverse-gradient(use a extra head with 1 output to do binary classification), want to retard the difference between them, but the loss of binary-head become bery huge(can reach 10^2 magnitude), so I give up this way.\n***\nNo more experiments, hope my average-model can bring me the good luck. wanna shake 亿点点 too😜", "votes": 2 }, { "id": 939109, "postDate": "2020-07-22T03:27:09.167Z", "content": "

I have tried two ResNet-50 level backbone (ResNet50-gn & tresnet-m) with smooth L1 regression, I can only get 0.88LB but 0.91CV. (TTA, k folds ensemble and model blending are not able to give me great improvement in LB, but CV increased)
\nLimited by computing resources, I can’t run more experiments, each experiment takes me a week.
\nDoes the choice of backbone matter? Welcome everyone to share the experimental results.

", "rawMarkdown": "I have tried two ResNet-50 level backbone (ResNet50-gn & tresnet-m) with smooth L1 regression, I can only get 0.88LB but 0.91CV. (TTA, k folds ensemble and model blending are not able to give me great improvement in LB, but CV increased)\nLimited by computing resources, I can’t run more experiments, each experiment takes me a week.\nDoes the choice of backbone matter? Welcome everyone to share the experimental results." } ], "comments": [ { "id": 939286, "author_name": "Shiyuan Zeng", "author_url": "", "post_date": "2020-07-22T06:24:21.243000", "content": "
    \n
  1. efn-b0 here, also limited by computing resources. 5-folds splited but I only choose 2 of 5 folds(2 hightest local score fold) and average them can get 0.888 lb(no tta).
    2. \n
  2. I try re-sampling the data to train a b0 these days but didn't get improvement.(actually it hurt local score and lb)
    3. \n
  3. because there are 2 data-providers, I also try feed full-img(after resize, no tiles) to model(a efn-b4) and training with reverse-gradient(use a extra head with 1 output to do binary classification), want to retard the difference between them, but the loss of binary-head become bery huge(can reach 10^2 magnitude), so I give up this way.
    4. \n
\n
\n

No more experiments, hope my average-model can bring me the good luck. wanna shake 亿点点 too😜

", "votes": 2, "replies": [] } ], "raw_markdown_by_id": { "939286": "1. efn-b0 here, also limited by computing resources. 5-folds splited but I only choose 2 of 5 folds(2 hightest local score fold) and average them can get 0.888 lb(no tta).\n2. I try re-sampling the data to train a b0 these days but didn't get improvement.(actually it hurt local score and lb)\n3. because there are 2 data-providers, I also try feed full-img(after resize, no tiles) to model(a efn-b4) and training with reverse-gradient(use a extra head with 1 output to do binary classification), want to retard the difference between them, but the loss of binary-head become bery huge(can reach 10^2 magnitude), so I give up this way.\n***\nNo more experiments, hope my average-model can bring me the good luck. wanna shake 亿点点 too😜", "939109": "I have tried two ResNet-50 level backbone (ResNet50-gn & tresnet-m) with smooth L1 regression, I can only get 0.88LB but 0.91CV. (TTA, k folds ensemble and model blending are not able to give me great improvement in LB, but CV increased)\nLimited by computing resources, I can’t run more experiments, each experiment takes me a week.\nDoes the choice of backbone matter? Welcome everyone to share the experimental results." } }