{
  "id": 166697,
  "title": "Fast.ai (v. 1) starter with pretty good LB (0.926): EffNetB0 5-Fold with external data",
  "url": "/competitions/siim-isic-melanoma-classification/discussion/166697",
  "author_name": "",
  "post_date": "2020-07-13T18:10:16.841738800Z",
  "votes": 8,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I have noticed that fast.ai library not very popular in this competition. So I't nice to have a <a href=\"https://www.kaggle.com/volcanoflash/melanoma-siim-isic-2020-fast-ai-efficientnetb0\">good starter</a> for everyone who have troubles with it. </p>\n\n<p><strong>Some additional details:</strong>\nNo ensembling\nOne model: Efficient Net B0\n256x256 size\n5 folds\nBS: 32\nTTA\nOversampling x5 for minority class\nFocal Loss</p>",
  "messages": [
    {
      "id": "928058",
      "postDate": "07/13/2020 18:10:16",
      "content": "<p>I have noticed that fast.ai library not very popular in this competition. So I't nice to have a <a href=\"https://www.kaggle.com/volcanoflash/melanoma-siim-isic-2020-fast-ai-efficientnetb0\">good starter</a> for everyone who have troubles with it. </p>\n\n<p><strong>Some additional details:</strong>\nNo ensembling\nOne model: Efficient Net B0\n256x256 size\n5 folds\nBS: 32\nTTA\nOversampling x5 for minority class\nFocal Loss</p>",
      "rawMarkdown": "I have noticed that fast.ai library not very popular in this competition. So I't nice to have a [good starter](https://www.kaggle.com/volcanoflash/melanoma-siim-isic-2020-fast-ai-efficientnetb0) for everyone who have troubles with it. \n\n**Some additional details:**\nNo ensembling\nOne model: Efficient Net B0\n256x256 size\n5 folds\nBS: 32\nTTA\nOversampling x5 for minority class\nFocal Loss",
      "votes": null
    },
    {
      "id": "965506",
      "postDate": "08/10/2020 17:21:45",
      "content": "<p>Nice <a href=\"/volcanoflash\">@volcanoflash</a> , thanks for sharing.\nI've used some your code to use effnetb0 in my setup (same dataset). I have also implemented 2 steps (1 frozen, 1 unfrozen) training with several epochs each + training on worst training results.</p>\n\n<p>I've noticed the folds' loss and AUC don't improve that much considering the amount of time spent training. \nI get:\n- fold 0: initial AUC after 1 epoch: .794. After 2 hours of training, it reaches .896\n- fold 1: initial AUC after 1 epoch: .894. After 2 hours of training, it reaches .897\n- fold 2: initial AUC after 1 epoch: .933. After 2 hours of training, it reaches .935\n- fold 3: initial AUC after 1 epoch: .932. After 2 hours of training, it reaches .934\n- fold 4: initial AUC after 1 epoch: .845. After 2 hours of training, it reaches .902</p>\n\n<p>So this takes close to 10 hours and the difference between the 1st epoch and last is not that great.</p>\n\n<p>In your pinned version 15 of the notebook, I also see these logs for example. Here also the training loss decreases but the validation loss and auroc don't change much.\n&gt; &gt; epoch train_loss  valid_loss  auroc   time\n0   0.088257    0.079054    0.911207    22:03\n1   0.066991    0.065442    0.911773    21:30\n2   0.053480    0.074863    0.914240    21:28</p>\n\n<p>I'm guessing there are ways to improve this as it feels like a lot of training time for just a few points in some folds.\nI'm wondering if you noticed the same thing and how you went from there.</p>",
      "rawMarkdown": "Nice @volcanoflash , thanks for sharing.\nI've used some your code to use effnetb0 in my setup (same dataset). I have also implemented 2 steps (1 frozen, 1 unfrozen) training with several epochs each + training on worst training results.\n\nI've noticed the folds' loss and AUC don't improve that much considering the amount of time spent training. \nI get:\n- fold 0: initial AUC after 1 epoch: .794. After 2 hours of training, it reaches .896\n- fold 1: initial AUC after 1 epoch: .894. After 2 hours of training, it reaches .897\n- fold 2: initial AUC after 1 epoch: .933. After 2 hours of training, it reaches .935\n- fold 3: initial AUC after 1 epoch: .932. After 2 hours of training, it reaches .934\n- fold 4: initial AUC after 1 epoch: .845. After 2 hours of training, it reaches .902\n\nSo this takes close to 10 hours and the difference between the 1st epoch and last is not that great.\n\nIn your pinned version 15 of the notebook, I also see these logs for example. Here also the training loss decreases but the validation loss and auroc don't change much.\n&gt; &gt; epoch\ttrain_loss\tvalid_loss\tauroc\ttime\n0\t0.088257\t0.079054\t0.911207\t22:03\n1\t0.066991\t0.065442\t0.911773\t21:30\n2\t0.053480\t0.074863\t0.914240\t21:28\n\nI'm guessing there are ways to improve this as it feels like a lot of training time for just a few points in some folds.\nI'm wondering if you noticed the same thing and how you went from there.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 965506,
      "author_name": "hamonk",
      "author_url": "",
      "post_date": "08/10/2020 17:21:45",
      "content": "<p>Nice <a href=\"/volcanoflash\">@volcanoflash</a> , thanks for sharing.\nI've used some your code to use effnetb0 in my setup (same dataset). I have also implemented 2 steps (1 frozen, 1 unfrozen) training with several epochs each + training on worst training results.</p>\n\n<p>I've noticed the folds' loss and AUC don't improve that much considering the amount of time spent training. \nI get:\n- fold 0: initial AUC after 1 epoch: .794. After 2 hours of training, it reaches .896\n- fold 1: initial AUC after 1 epoch: .894. After 2 hours of training, it reaches .897\n- fold 2: initial AUC after 1 epoch: .933. After 2 hours of training, it reaches .935\n- fold 3: initial AUC after 1 epoch: .932. After 2 hours of training, it reaches .934\n- fold 4: initial AUC after 1 epoch: .845. After 2 hours of training, it reaches .902</p>\n\n<p>So this takes close to 10 hours and the difference between the 1st epoch and last is not that great.</p>\n\n<p>In your pinned version 15 of the notebook, I also see these logs for example. Here also the training loss decreases but the validation loss and auroc don't change much.\n&gt; &gt; epoch train_loss  valid_loss  auroc   time\n0   0.088257    0.079054    0.911207    22:03\n1   0.066991    0.065442    0.911773    21:30\n2   0.053480    0.074863    0.914240    21:28</p>\n\n<p>I'm guessing there are ways to improve this as it feels like a lot of training time for just a few points in some folds.\nI'm wondering if you noticed the same thing and how you went from there.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "928058": "I have noticed that fast.ai library not very popular in this competition. So I't nice to have a [good starter](https://www.kaggle.com/volcanoflash/melanoma-siim-isic-2020-fast-ai-efficientnetb0) for everyone who have troubles with it. \n\n**Some additional details:**\nNo ensembling\nOne model: Efficient Net B0\n256x256 size\n5 folds\nBS: 32\nTTA\nOversampling x5 for minority class\nFocal Loss",
    "965506": "Nice @volcanoflash , thanks for sharing.\nI've used some your code to use effnetb0 in my setup (same dataset). I have also implemented 2 steps (1 frozen, 1 unfrozen) training with several epochs each + training on worst training results.\n\nI've noticed the folds' loss and AUC don't improve that much considering the amount of time spent training. \nI get:\n- fold 0: initial AUC after 1 epoch: .794. After 2 hours of training, it reaches .896\n- fold 1: initial AUC after 1 epoch: .894. After 2 hours of training, it reaches .897\n- fold 2: initial AUC after 1 epoch: .933. After 2 hours of training, it reaches .935\n- fold 3: initial AUC after 1 epoch: .932. After 2 hours of training, it reaches .934\n- fold 4: initial AUC after 1 epoch: .845. After 2 hours of training, it reaches .902\n\nSo this takes close to 10 hours and the difference between the 1st epoch and last is not that great.\n\nIn your pinned version 15 of the notebook, I also see these logs for example. Here also the training loss decreases but the validation loss and auroc don't change much.\n&gt; &gt; epoch\ttrain_loss\tvalid_loss\tauroc\ttime\n0\t0.088257\t0.079054\t0.911207\t22:03\n1\t0.066991\t0.065442\t0.911773\t21:30\n2\t0.053480\t0.074863\t0.914240\t21:28\n\nI'm guessing there are ways to improve this as it feels like a lot of training time for just a few points in some folds.\nI'm wondering if you noticed the same thing and how you went from there."
  },
  "source": "meta"
}