{
  "id": 47730,
  "title": "Newbie solution. Best private score 0.90908.",
  "url": "/competitions/tensorflow-speech-recognition-challenge/writeups/great-shu-newbie-solution-best-private-score-0-909",
  "author_name": "",
  "post_date": "2018-01-18T12:20:06.023Z",
  "votes": 40,
  "comment_count": 17,
  "views": 0,
  "content": "<p>Hi, thank you all. This is my first time to Kaggle. I learned a lot in this competition. Congrats to Heng CherKeng, Ryan Sun, and See!</p>\n\n<p>My best private score is 0.90908.  What I chose as final submissions scores 0.90790 :).</p>\n\n<p>Here is my solution. A very simple solution... Code comes from <a href=\"https://www.tensorflow.org/tutorials/audio_recognition\">Google tutorial</a> mostly.</p>\n\n<ul>\n<li><p>I used vgg11 with an input of 40 * 98 log mel spectrogram. I got 0.86 private LB. After log mel + mfcc ensemble, I got 0.88 private LB.</p></li>\n<li><p>I used mixup, I got 0.89 private LB. After ensemble, I got 0.909 private LB, my best LB score. <a href=\"https://arxiv.org/abs/1710.09412\">paper link</a></p></li>\n<li>knowledge distillation to train a smaller model. The small model scores 0.90 private LB, which runs very fast. <a href=\"https://arxiv.org/abs/1503.02531\">paper link</a></li>\n</ul>\n\n<p><strong>Code</strong></p>\n\n<p>mixup:</p>\n\n<pre><code>def get_data(args):\n  weight = np.random.beta(alpha, alpha, batch_size)\n  x_weight = weight.reshape(batch_size, 1)\n  y_weight = weight.reshape(batch_size, 1)\n  index = np.random.permutation(batch_size)\n  x1, x2 = data, data[index]\n  x = x1 * x_weight + x2 * (1 - x_weight)\n  y1, y2 = labels, labels[index]\n  y = y1 * y_weight + y2 * (1 - y_weight)\n  return x, y\n</code></pre>\n\n<p>alpha is set to 2</p>\n\n<p>Knowledge distillation code <a href=\"https://github.com/chengshengchan/model_compression\">here</a>.</p>",
  "messages": [
    {
      "id": "270340",
      "postDate": "01/18/2018 05:30:47",
      "content": "<p>Hi, thank you all. This is my first time to Kaggle. I learned a lot in this competition. Congrats to Heng CherKeng, Ryan Sun, and See!</p>\n\n<p>My best private score is 0.90908.  What I chose as final submissions scores 0.90790 :).</p>\n\n<p>Here is my solution. A very simple solution... Code comes from <a href=\"https://www.tensorflow.org/tutorials/audio_recognition\">Google tutorial</a> mostly.</p>\n\n<ul>\n<li><p>I used vgg11 with an input of 40 * 98 log mel spectrogram. I got 0.86 private LB. After log mel + mfcc ensemble, I got 0.88 private LB.</p></li>\n<li><p>I used mixup, I got 0.89 private LB. After ensemble, I got 0.909 private LB, my best LB score. <a href=\"https://arxiv.org/abs/1710.09412\">paper link</a></p></li>\n<li>knowledge distillation to train a smaller model. The small model scores 0.90 private LB, which runs very fast. <a href=\"https://arxiv.org/abs/1503.02531\">paper link</a></li>\n</ul>\n\n<p><strong>Code</strong></p>\n\n<p>mixup:</p>\n\n<pre><code>def get_data(args):\n  weight = np.random.beta(alpha, alpha, batch_size)\n  x_weight = weight.reshape(batch_size, 1)\n  y_weight = weight.reshape(batch_size, 1)\n  index = np.random.permutation(batch_size)\n  x1, x2 = data, data[index]\n  x = x1 * x_weight + x2 * (1 - x_weight)\n  y1, y2 = labels, labels[index]\n  y = y1 * y_weight + y2 * (1 - y_weight)\n  return x, y\n</code></pre>\n\n<p>alpha is set to 2</p>\n\n<p>Knowledge distillation code <a href=\"https://github.com/chengshengchan/model_compression\">here</a>.</p>",
      "rawMarkdown": "Hi, thank you all. This is my first time to Kaggle. I learned a lot in this competition. Congrats to Heng CherKeng, Ryan Sun, and See!\n\nMy best private score is 0.90908.  What I chose as final submissions scores 0.90790 :).\n\nHere is my solution. A very simple solution... Code comes from [Google tutorial](https://www.tensorflow.org/tutorials/audio_recognition) mostly.\n\n- I used vgg11 with an input of 40 * 98 log mel spectrogram. I got 0.86 private LB. After log mel + mfcc ensemble, I got 0.88 private LB.\n\n- I used mixup, I got 0.89 private LB. After ensemble, I got 0.909 private LB, my best LB score. [paper link](https://arxiv.org/abs/1710.09412)\n- knowledge distillation to train a smaller model. The small model scores 0.90 private LB, which runs very fast. [paper link](https://arxiv.org/abs/1503.02531)\n\n**Code**\n\nmixup:\n\n    def get_data(args):\n      weight = np.random.beta(alpha, alpha, batch_size)\n      x_weight = weight.reshape(batch_size, 1)\n      y_weight = weight.reshape(batch_size, 1)\n      index = np.random.permutation(batch_size)\n      x1, x2 = data, data[index]\n      x = x1 * x_weight + x2 * (1 - x_weight)\n      y1, y2 = labels, labels[index]\n      y = y1 * y_weight + y2 * (1 - y_weight)\n      return x, y\n\nalpha is set to 2\n\nKnowledge distillation code [here](https://github.com/chengshengchan/model_compression).",
      "votes": null
    },
    {
      "id": "270359",
      "postDate": "01/18/2018 06:00:54",
      "content": "<p>In mixup code, you're just adding two waves file with weight?</p>",
      "rawMarkdown": "In mixup code, you're just adding two waves file with weight?",
      "votes": null
    },
    {
      "id": "270386",
      "postDate": "01/18/2018 07:08:32",
      "content": "<p>Yes.</p>",
      "rawMarkdown": "Yes.",
      "votes": null
    },
    {
      "id": "270472",
      "postDate": "01/18/2018 09:31:38",
      "content": "<p>Almost forget it. Before mixup two waves, I used data augmentation, which is the same as the tutorial,\n <a href=\"https://paste.ubuntu.com/26409540/\">code here, line 67-78</a>. </p>",
      "rawMarkdown": "Almost forget it. Before mixup two waves, I used data augmentation, which is the same as the tutorial,\n [code here, line 67-78](https://paste.ubuntu.com/26409540/).",
      "votes": null
    },
    {
      "id": "270675",
      "postDate": "01/18/2018 18:05:52",
      "content": "<p>Wow.. This is quite mind-opening. I read the paper also, but I thought it cannot possibly work. \nThanks for sharing!</p>",
      "rawMarkdown": "Wow.. This is quite mind-opening. I read the paper also, but I thought it cannot possibly work. \nThanks for sharing!",
      "votes": null
    },
    {
      "id": "270732",
      "postDate": "01/18/2018 19:46:38",
      "content": "<p>Wow, great to see how effective mixup was. I tried the version from the blog post that was posted on the discussion forum...which I now realize was only valid for two-class classification. Needless to say I didn't see an improvement and am now kicking myself haha</p>\n\n<p>Really like the idea of inducing a bias towards linear interpolation between class boundaries. Good to know it worked so well, thanks for the post!</p>",
      "rawMarkdown": "Wow, great to see how effective mixup was. I tried the version from the blog post that was posted on the discussion forum...which I now realize was only valid for two-class classification. Needless to say I didn't see an improvement and am now kicking myself haha\n\nReally like the idea of inducing a bias towards linear interpolation between class boundaries. Good to know it worked so well, thanks for the post!",
      "votes": null
    },
    {
      "id": "270943",
      "postDate": "01/19/2018 09:54:36",
      "content": "<p>Thank you for sharing simple but powerful approach! It is very impressive mixup can raise 1%.</p>",
      "rawMarkdown": "Thank you for sharing simple but powerful approach! It is very impressive mixup can raise 1%.",
      "votes": null
    },
    {
      "id": "270985",
      "postDate": "01/19/2018 11:34:33",
      "content": "<p>Awesome use of mixup. How did you decide on alpha=2?</p>",
      "rawMarkdown": "Awesome use of mixup. How did you decide on alpha=2?",
      "votes": null
    },
    {
      "id": "271088",
      "postDate": "01/19/2018 16:54:07",
      "content": "<p>You violated <strong>Rule 2nd: No private sharing outside teams</strong></p>\n\n<p>Don't steal the idea from me and \"share\" it.</p>\n\n<p>Shame on you</p>",
      "rawMarkdown": "You violated **Rule 2nd: No private sharing outside teams**\n\nDon't steal the idea from me and \"share\" it.\n\nShame on you",
      "votes": null
    },
    {
      "id": "271194",
      "postDate": "01/19/2018 21:33:18",
      "content": "<p>It seems that both teams have violated the rules. You can try to contact Kaggle team here (<a href=\"https://www.kaggle.com/compliance\">https://www.kaggle.com/compliance</a>). This link may help you to report both teams. And both teams will be removed from leaderboard. Or you can just create a new topic for your situation in the discussion.</p>",
      "rawMarkdown": "It seems that both teams have violated the rules. You can try to contact Kaggle team here (https://www.kaggle.com/compliance). This link may help you to report both teams. And both teams will be removed from leaderboard. Or you can just create a new topic for your situation in the discussion.",
      "votes": null
    },
    {
      "id": "271307",
      "postDate": "01/20/2018 05:22:53",
      "content": "<p>log mel feature: <a href=\"https://www.kaggle.com/c/tensorflow-speech-recognition-challenge/discussion/45362\">https://www.kaggle.com/c/tensorflow-speech-recognition-challenge/discussion/45362</a></p>\n\n<p>mixup: <a href=\"https://www.kaggle.com/c/tensorflow-speech-recognition-challenge/discussion/47134\">https://www.kaggle.com/c/tensorflow-speech-recognition-challenge/discussion/47134</a> </p>\n\n<p>knowledge distillation: <a href=\"https://www.kaggle.com/c/tensorflow-speech-recognition-challenge/discussion/46945\">https://www.kaggle.com/c/tensorflow-speech-recognition-challenge/discussion/46945</a></p>\n\n<p>I just mix the methods above up. -.-</p>",
      "rawMarkdown": "log mel feature: https://www.kaggle.com/c/tensorflow-speech-recognition-challenge/discussion/45362\n\nmixup: https://www.kaggle.com/c/tensorflow-speech-recognition-challenge/discussion/47134 \n\nknowledge distillation: https://www.kaggle.com/c/tensorflow-speech-recognition-challenge/discussion/46945\n\nI just mix the methods above up. -.-",
      "votes": null
    },
    {
      "id": "271310",
      "postDate": "01/20/2018 05:38:06",
      "content": "<p>Set alpha to 2 is inspired by a blog. And this is my first attempt. I find it works well for me. I tried 2, 4, 6. And '2' works best. If you want to do more experiment, I think you can try to set alpha slower, such as 0.2, 0.4, and so no.</p>",
      "rawMarkdown": "Set alpha to 2 is inspired by a blog. And this is my first attempt. I find it works well for me. I tried 2, 4, 6. And '2' works best. If you want to do more experiment, I think you can try to set alpha slower, such as 0.2, 0.4, and so no.",
      "votes": null
    },
    {
      "id": "271328",
      "postDate": "01/20/2018 07:14:23",
      "content": "<p>Your motivation for speaking all these is suspicious to me,  especially when I consider your extremely abusive team name in the LB, which I am trying to ignore. All the method we use was inspired by the Discussion and Kernel board on Kaggle, and it is totally legitimate for us to share any of ours on this platform according to rules. \nFurther, I suggest you to contact official compliance team if you think our results are illegal instead of making such unconstructive under this topic.</p>",
      "rawMarkdown": "Your motivation for speaking all these is suspicious to me,  especially when I consider your extremely abusive team name in the LB, which I am trying to ignore. All the method we use was inspired by the Discussion and Kernel board on Kaggle, and it is totally legitimate for us to share any of ours on this platform according to rules. \nFurther, I suggest you to contact official compliance team if you think our results are illegal instead of making such unconstructive under this topic.",
      "votes": null
    },
    {
      "id": "271435",
      "postDate": "01/20/2018 12:57:44",
      "content": "<p>It's a 3% increase! Thank you for sharing OP.</p>",
      "rawMarkdown": "It's a 3% increase! Thank you for sharing OP.",
      "votes": null
    },
    {
      "id": "271536",
      "postDate": "01/20/2018 17:32:00",
      "content": "<p>Another point worth noting is, if you want to do ensemble between mixuped model with none mixup model, you should add a proportion to model, instead of taking the average. \nFor examples, if model A used mixup, model B didn't, C is the result. When ensemble, C=0.5*B + A. You can also try other numbers. 0.5 works well for my model.\nThe reason is, after mixup, the score of each class will be slower, which is not difficult to understand if you know the formula of mixup. If you just take the average, B will lead the ensemble process, which may get you bad result.</p>",
      "rawMarkdown": "Another point worth noting is, if you want to do ensemble between mixuped model with none mixup model, you should add a proportion to model, instead of taking the average. \nFor examples, if model A used mixup, model B didn't, C is the result. When ensemble, C=0.5*B + A. You can also try other numbers. 0.5 works well for my model.\nThe reason is, after mixup, the score of each class will be slower, which is not difficult to understand if you know the formula of mixup. If you just take the average, B will lead the ensemble process, which may get you bad result.",
      "votes": null
    },
    {
      "id": "280413",
      "postDate": "02/09/2018 23:11:07",
      "content": "",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "283710",
      "postDate": "02/15/2018 21:32:57",
      "content": "<p>Anyone been able to reproduce this? I have not.</p>",
      "rawMarkdown": "Anyone been able to reproduce this? I have not.",
      "votes": null
    },
    {
      "id": "286987",
      "postDate": "02/23/2018 06:22:48",
      "content": "<p>The 40 * 98 log mel spectrogram: 40 means time bins, and 98 means frequence bins. Is it right?\nAfter log mel + mfcc ensemble, Is the new feature 2 * 40 * 98？\nThanks！</p>",
      "rawMarkdown": "The 40 * 98 log mel spectrogram: 40 means time bins, and 98 means frequence bins. Is it right?\nAfter log mel + mfcc ensemble, Is the new feature 2 * 40 * 98？\nThanks！",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 270359,
      "author_name": "jandjenter",
      "author_url": "",
      "post_date": "01/18/2018 06:00:54",
      "content": "<p>In mixup code, you're just adding two waves file with weight?</p>",
      "votes": null,
      "replies": [
        {
          "id": 270386,
          "author_name": "jihaoliu",
          "author_url": "",
          "post_date": "01/18/2018 07:08:32",
          "content": "<p>Yes.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 270472,
          "author_name": "jihaoliu",
          "author_url": "",
          "post_date": "01/18/2018 09:31:38",
          "content": "<p>Almost forget it. Before mixup two waves, I used data augmentation, which is the same as the tutorial,\n <a href=\"https://paste.ubuntu.com/26409540/\">code here, line 67-78</a>. </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 270675,
          "author_name": "jandjenter",
          "author_url": "",
          "post_date": "01/18/2018 18:05:52",
          "content": "<p>Wow.. This is quite mind-opening. I read the paper also, but I thought it cannot possibly work. \nThanks for sharing!</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 270732,
      "author_name": "omalleyt",
      "author_url": "",
      "post_date": "01/18/2018 19:46:38",
      "content": "<p>Wow, great to see how effective mixup was. I tried the version from the blog post that was posted on the discussion forum...which I now realize was only valid for two-class classification. Needless to say I didn't see an improvement and am now kicking myself haha</p>\n\n<p>Really like the idea of inducing a bias towards linear interpolation between class boundaries. Good to know it worked so well, thanks for the post!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 270943,
      "author_name": "soonhwankwon",
      "author_url": "",
      "post_date": "01/19/2018 09:54:36",
      "content": "<p>Thank you for sharing simple but powerful approach! It is very impressive mixup can raise 1%.</p>",
      "votes": null,
      "replies": [
        {
          "id": 271435,
          "author_name": "nimitz14",
          "author_url": "",
          "post_date": "01/20/2018 12:57:44",
          "content": "<p>It's a 3% increase! Thank you for sharing OP.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 270985,
      "author_name": "antorsae",
      "author_url": "",
      "post_date": "01/19/2018 11:34:33",
      "content": "<p>Awesome use of mixup. How did you decide on alpha=2?</p>",
      "votes": null,
      "replies": [
        {
          "id": 271310,
          "author_name": "jihaoliu",
          "author_url": "",
          "post_date": "01/20/2018 05:38:06",
          "content": "<p>Set alpha to 2 is inspired by a blog. And this is my first attempt. I find it works well for me. I tried 2, 4, 6. And '2' works best. If you want to do more experiment, I think you can try to set alpha slower, such as 0.2, 0.4, and so no.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 271088,
      "author_name": "delphi",
      "author_url": "",
      "post_date": "01/19/2018 16:54:07",
      "content": "<p>You violated <strong>Rule 2nd: No private sharing outside teams</strong></p>\n\n<p>Don't steal the idea from me and \"share\" it.</p>\n\n<p>Shame on you</p>",
      "votes": null,
      "replies": [
        {
          "id": 271194,
          "author_name": "cczaixian",
          "author_url": "",
          "post_date": "01/19/2018 21:33:18",
          "content": "<p>It seems that both teams have violated the rules. You can try to contact Kaggle team here (<a href=\"https://www.kaggle.com/compliance\">https://www.kaggle.com/compliance</a>). This link may help you to report both teams. And both teams will be removed from leaderboard. Or you can just create a new topic for your situation in the discussion.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 271307,
          "author_name": "jihaoliu",
          "author_url": "",
          "post_date": "01/20/2018 05:22:53",
          "content": "<p>log mel feature: <a href=\"https://www.kaggle.com/c/tensorflow-speech-recognition-challenge/discussion/45362\">https://www.kaggle.com/c/tensorflow-speech-recognition-challenge/discussion/45362</a></p>\n\n<p>mixup: <a href=\"https://www.kaggle.com/c/tensorflow-speech-recognition-challenge/discussion/47134\">https://www.kaggle.com/c/tensorflow-speech-recognition-challenge/discussion/47134</a> </p>\n\n<p>knowledge distillation: <a href=\"https://www.kaggle.com/c/tensorflow-speech-recognition-challenge/discussion/46945\">https://www.kaggle.com/c/tensorflow-speech-recognition-challenge/discussion/46945</a></p>\n\n<p>I just mix the methods above up. -.-</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 271328,
          "author_name": "bicbrv",
          "author_url": "",
          "post_date": "01/20/2018 07:14:23",
          "content": "<p>Your motivation for speaking all these is suspicious to me,  especially when I consider your extremely abusive team name in the LB, which I am trying to ignore. All the method we use was inspired by the Discussion and Kernel board on Kaggle, and it is totally legitimate for us to share any of ours on this platform according to rules. \nFurther, I suggest you to contact official compliance team if you think our results are illegal instead of making such unconstructive under this topic.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 271536,
      "author_name": "jihaoliu",
      "author_url": "",
      "post_date": "01/20/2018 17:32:00",
      "content": "<p>Another point worth noting is, if you want to do ensemble between mixuped model with none mixup model, you should add a proportion to model, instead of taking the average. \nFor examples, if model A used mixup, model B didn't, C is the result. When ensemble, C=0.5*B + A. You can also try other numbers. 0.5 works well for my model.\nThe reason is, after mixup, the score of each class will be slower, which is not difficult to understand if you know the formula of mixup. If you just take the average, B will lead the ensemble process, which may get you bad result.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 280413,
      "author_name": "nimitz14",
      "author_url": "",
      "post_date": "02/09/2018 23:11:07",
      "content": "",
      "votes": null,
      "replies": []
    },
    {
      "id": 283710,
      "author_name": "nimitz14",
      "author_url": "",
      "post_date": "02/15/2018 21:32:57",
      "content": "<p>Anyone been able to reproduce this? I have not.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 286987,
      "author_name": "zhengge",
      "author_url": "",
      "post_date": "02/23/2018 06:22:48",
      "content": "<p>The 40 * 98 log mel spectrogram: 40 means time bins, and 98 means frequence bins. Is it right?\nAfter log mel + mfcc ensemble, Is the new feature 2 * 40 * 98？\nThanks！</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "270340": "Hi, thank you all. This is my first time to Kaggle. I learned a lot in this competition. Congrats to Heng CherKeng, Ryan Sun, and See!\n\nMy best private score is 0.90908.  What I chose as final submissions scores 0.90790 :).\n\nHere is my solution. A very simple solution... Code comes from [Google tutorial](https://www.tensorflow.org/tutorials/audio_recognition) mostly.\n\n- I used vgg11 with an input of 40 * 98 log mel spectrogram. I got 0.86 private LB. After log mel + mfcc ensemble, I got 0.88 private LB.\n\n- I used mixup, I got 0.89 private LB. After ensemble, I got 0.909 private LB, my best LB score. [paper link](https://arxiv.org/abs/1710.09412)\n- knowledge distillation to train a smaller model. The small model scores 0.90 private LB, which runs very fast. [paper link](https://arxiv.org/abs/1503.02531)\n\n**Code**\n\nmixup:\n\n    def get_data(args):\n      weight = np.random.beta(alpha, alpha, batch_size)\n      x_weight = weight.reshape(batch_size, 1)\n      y_weight = weight.reshape(batch_size, 1)\n      index = np.random.permutation(batch_size)\n      x1, x2 = data, data[index]\n      x = x1 * x_weight + x2 * (1 - x_weight)\n      y1, y2 = labels, labels[index]\n      y = y1 * y_weight + y2 * (1 - y_weight)\n      return x, y\n\nalpha is set to 2\n\nKnowledge distillation code [here](https://github.com/chengshengchan/model_compression).",
    "270359": "In mixup code, you're just adding two waves file with weight?",
    "270386": "Yes.",
    "270472": "Almost forget it. Before mixup two waves, I used data augmentation, which is the same as the tutorial,\n [code here, line 67-78](https://paste.ubuntu.com/26409540/).",
    "270675": "Wow.. This is quite mind-opening. I read the paper also, but I thought it cannot possibly work. \nThanks for sharing!",
    "270732": "Wow, great to see how effective mixup was. I tried the version from the blog post that was posted on the discussion forum...which I now realize was only valid for two-class classification. Needless to say I didn't see an improvement and am now kicking myself haha\n\nReally like the idea of inducing a bias towards linear interpolation between class boundaries. Good to know it worked so well, thanks for the post!",
    "270943": "Thank you for sharing simple but powerful approach! It is very impressive mixup can raise 1%.",
    "270985": "Awesome use of mixup. How did you decide on alpha=2?",
    "271088": "You violated **Rule 2nd: No private sharing outside teams**\n\nDon't steal the idea from me and \"share\" it.\n\nShame on you",
    "271194": "It seems that both teams have violated the rules. You can try to contact Kaggle team here (https://www.kaggle.com/compliance). This link may help you to report both teams. And both teams will be removed from leaderboard. Or you can just create a new topic for your situation in the discussion.",
    "271307": "log mel feature: https://www.kaggle.com/c/tensorflow-speech-recognition-challenge/discussion/45362\n\nmixup: https://www.kaggle.com/c/tensorflow-speech-recognition-challenge/discussion/47134 \n\nknowledge distillation: https://www.kaggle.com/c/tensorflow-speech-recognition-challenge/discussion/46945\n\nI just mix the methods above up. -.-",
    "271310": "Set alpha to 2 is inspired by a blog. And this is my first attempt. I find it works well for me. I tried 2, 4, 6. And '2' works best. If you want to do more experiment, I think you can try to set alpha slower, such as 0.2, 0.4, and so no.",
    "271328": "Your motivation for speaking all these is suspicious to me,  especially when I consider your extremely abusive team name in the LB, which I am trying to ignore. All the method we use was inspired by the Discussion and Kernel board on Kaggle, and it is totally legitimate for us to share any of ours on this platform according to rules. \nFurther, I suggest you to contact official compliance team if you think our results are illegal instead of making such unconstructive under this topic.",
    "271435": "It's a 3% increase! Thank you for sharing OP.",
    "271536": "Another point worth noting is, if you want to do ensemble between mixuped model with none mixup model, you should add a proportion to model, instead of taking the average. \nFor examples, if model A used mixup, model B didn't, C is the result. When ensemble, C=0.5*B + A. You can also try other numbers. 0.5 works well for my model.\nThe reason is, after mixup, the score of each class will be slower, which is not difficult to understand if you know the formula of mixup. If you just take the average, B will lead the ensemble process, which may get you bad result.",
    "280413": "",
    "283710": "Anyone been able to reproduce this? I have not.",
    "286987": "The 40 * 98 log mel spectrogram: 40 means time bins, and 98 means frequence bins. Is it right?\nAfter log mel + mfcc ensemble, Is the new feature 2 * 40 * 98？\nThanks！"
  },
  "source": "meta"
}