{
  "id": 202652,
  "title": "Pytorch vs Keras?",
  "url": "/competitions/cassava-leaf-disease-classification/discussion/202652",
  "author_name": "",
  "post_date": "2020-12-11T08:10:18.971931600Z",
  "votes": 3,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Hi everyone, I'm pretty new to this competition so I hope you will forgive me if this seems like a stupid question: Given that identical models were used, why does it seem like models utilizing Pytorch performs better on the LB than models using Keras? </p>\n<p>I understand that different hyperparameter tuning or image augmentation as well as variance may lead to different performance, but from what I can tell, public notebooks utilising Keras have had difficulty breaking past 0.901 score on LB while notebooks utilising Pytorch hasn't..</p>",
  "messages": [
    {
      "id": "1108984",
      "postDate": "12/11/2020 08:10:18",
      "content": "<p>Hi everyone, I'm pretty new to this competition so I hope you will forgive me if this seems like a stupid question: Given that identical models were used, why does it seem like models utilizing Pytorch performs better on the LB than models using Keras? </p>\n<p>I understand that different hyperparameter tuning or image augmentation as well as variance may lead to different performance, but from what I can tell, public notebooks utilising Keras have had difficulty breaking past 0.901 score on LB while notebooks utilising Pytorch hasn't..</p>",
      "rawMarkdown": "Hi everyone, I'm pretty new to this competition so I hope you will forgive me if this seems like a stupid question: Given that identical models were used, why does it seem like models utilizing Pytorch performs better on the LB than models using Keras? \n\nI understand that different hyperparameter tuning or image augmentation as well as variance may lead to different performance, but from what I can tell, public notebooks utilising Keras have had difficulty breaking past 0.901 score on LB while notebooks utilising Pytorch hasn't..",
      "votes": null
    },
    {
      "id": "1109035",
      "postDate": "12/11/2020 09:12:53",
      "content": "<p>Interesting, can you give the link of those notebooks?</p>",
      "rawMarkdown": "Interesting, can you give the link of those notebooks?",
      "votes": null
    },
    {
      "id": "1109774",
      "postDate": "12/12/2020 03:43:23",
      "content": "<p>Sure, here are some.</p>\n<p>Pytorch:<br>\n<a href=\"https://www.kaggle.com/piantic/no-tta-cassava-resnext50-32x4d-inference-lb0-903\" target=\"_blank\">https://www.kaggle.com/piantic/no-tta-cassava-resnext50-32x4d-inference-lb0-903</a><br>\n<a href=\"https://www.kaggle.com/mekhdigakhramanian/pytorch-efficientnet-baseline-inference-tta\" target=\"_blank\">https://www.kaggle.com/mekhdigakhramanian/pytorch-efficientnet-baseline-inference-tta</a><br>\n<a href=\"https://www.kaggle.com/underwearfitting/clean-inference-kernel-8xtta-lb902\" target=\"_blank\">https://www.kaggle.com/underwearfitting/clean-inference-kernel-8xtta-lb902</a></p>\n<p>Keras:<br>\n<a href=\"https://www.kaggle.com/itsuki9180/efficientnet-and-cutmixup-with-tpu-predict-phase\" target=\"_blank\">https://www.kaggle.com/itsuki9180/efficientnet-and-cutmixup-with-tpu-predict-phase</a><br>\n<a href=\"https://www.kaggle.com/maksymshkliarevskyi/cassava-leaf-disease-best-keras-cnn\" target=\"_blank\">https://www.kaggle.com/maksymshkliarevskyi/cassava-leaf-disease-best-keras-cnn</a><br>\n<a href=\"https://www.kaggle.com/vfomenko/keras-baseline\" target=\"_blank\">https://www.kaggle.com/vfomenko/keras-baseline</a></p>",
      "rawMarkdown": "Sure, here are some.\n\nPytorch:\nhttps://www.kaggle.com/piantic/no-tta-cassava-resnext50-32x4d-inference-lb0-903\nhttps://www.kaggle.com/mekhdigakhramanian/pytorch-efficientnet-baseline-inference-tta\nhttps://www.kaggle.com/underwearfitting/clean-inference-kernel-8xtta-lb902\n\nKeras:\nhttps://www.kaggle.com/itsuki9180/efficientnet-and-cutmixup-with-tpu-predict-phase\nhttps://www.kaggle.com/maksymshkliarevskyi/cassava-leaf-disease-best-keras-cnn\nhttps://www.kaggle.com/vfomenko/keras-baseline",
      "votes": null
    },
    {
      "id": "1110920",
      "postDate": "12/13/2020 07:45:46",
      "content": "<p>simple solution: Use the onr that is giving YOU the better results; <br>\nand to your doubt many pytorch users uses weights from someone's repo not released by pytorch itself. maybe the ones who have trained it to save the weights have used different machines  or for higher epochs<br>\n(I can be  wrong ) <br>\nOne such awesome PyTorch repo is<br>\n<a href=\"https://github.com/rwightman/pytorch-image-models\" target=\"_blank\">https://github.com/rwightman/pytorch-image-models</a></p>",
      "rawMarkdown": "simple solution: Use the onr that is giving YOU the better results; \nand to your doubt many pytorch users uses weights from someone's repo not released by pytorch itself. maybe the ones who have trained it to save the weights have used different machines  or for higher epochs\n(I can be  wrong ) \nOne such awesome PyTorch repo is\nhttps://github.com/rwightman/pytorch-image-models",
      "votes": null
    },
    {
      "id": "1110961",
      "postDate": "12/13/2020 08:41:32",
      "content": "<p>Yeah, I understand that. I just wonder if there are some key differences in the ways the two libraries are implemented that contributes to the performance difference.</p>",
      "rawMarkdown": "Yeah, I understand that. I just wonder if there are some key differences in the ways the two libraries are implemented that contributes to the performance difference.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1109035,
      "author_name": "andreaschandra",
      "author_url": "",
      "post_date": "12/11/2020 09:12:53",
      "content": "<p>Interesting, can you give the link of those notebooks?</p>",
      "votes": null,
      "replies": [
        {
          "id": 1109774,
          "author_name": "junyingsg",
          "author_url": "",
          "post_date": "12/12/2020 03:43:23",
          "content": "<p>Sure, here are some.</p>\n<p>Pytorch:<br>\n<a href=\"https://www.kaggle.com/piantic/no-tta-cassava-resnext50-32x4d-inference-lb0-903\" target=\"_blank\">https://www.kaggle.com/piantic/no-tta-cassava-resnext50-32x4d-inference-lb0-903</a><br>\n<a href=\"https://www.kaggle.com/mekhdigakhramanian/pytorch-efficientnet-baseline-inference-tta\" target=\"_blank\">https://www.kaggle.com/mekhdigakhramanian/pytorch-efficientnet-baseline-inference-tta</a><br>\n<a href=\"https://www.kaggle.com/underwearfitting/clean-inference-kernel-8xtta-lb902\" target=\"_blank\">https://www.kaggle.com/underwearfitting/clean-inference-kernel-8xtta-lb902</a></p>\n<p>Keras:<br>\n<a href=\"https://www.kaggle.com/itsuki9180/efficientnet-and-cutmixup-with-tpu-predict-phase\" target=\"_blank\">https://www.kaggle.com/itsuki9180/efficientnet-and-cutmixup-with-tpu-predict-phase</a><br>\n<a href=\"https://www.kaggle.com/maksymshkliarevskyi/cassava-leaf-disease-best-keras-cnn\" target=\"_blank\">https://www.kaggle.com/maksymshkliarevskyi/cassava-leaf-disease-best-keras-cnn</a><br>\n<a href=\"https://www.kaggle.com/vfomenko/keras-baseline\" target=\"_blank\">https://www.kaggle.com/vfomenko/keras-baseline</a></p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1110920,
      "author_name": "mrinath",
      "author_url": "",
      "post_date": "12/13/2020 07:45:46",
      "content": "<p>simple solution: Use the onr that is giving YOU the better results; <br>\nand to your doubt many pytorch users uses weights from someone's repo not released by pytorch itself. maybe the ones who have trained it to save the weights have used different machines  or for higher epochs<br>\n(I can be  wrong ) <br>\nOne such awesome PyTorch repo is<br>\n<a href=\"https://github.com/rwightman/pytorch-image-models\" target=\"_blank\">https://github.com/rwightman/pytorch-image-models</a></p>",
      "votes": null,
      "replies": [
        {
          "id": 1110961,
          "author_name": "junyingsg",
          "author_url": "",
          "post_date": "12/13/2020 08:41:32",
          "content": "<p>Yeah, I understand that. I just wonder if there are some key differences in the ways the two libraries are implemented that contributes to the performance difference.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1108984": "Hi everyone, I'm pretty new to this competition so I hope you will forgive me if this seems like a stupid question: Given that identical models were used, why does it seem like models utilizing Pytorch performs better on the LB than models using Keras? \n\nI understand that different hyperparameter tuning or image augmentation as well as variance may lead to different performance, but from what I can tell, public notebooks utilising Keras have had difficulty breaking past 0.901 score on LB while notebooks utilising Pytorch hasn't..",
    "1109035": "Interesting, can you give the link of those notebooks?",
    "1109774": "Sure, here are some.\n\nPytorch:\nhttps://www.kaggle.com/piantic/no-tta-cassava-resnext50-32x4d-inference-lb0-903\nhttps://www.kaggle.com/mekhdigakhramanian/pytorch-efficientnet-baseline-inference-tta\nhttps://www.kaggle.com/underwearfitting/clean-inference-kernel-8xtta-lb902\n\nKeras:\nhttps://www.kaggle.com/itsuki9180/efficientnet-and-cutmixup-with-tpu-predict-phase\nhttps://www.kaggle.com/maksymshkliarevskyi/cassava-leaf-disease-best-keras-cnn\nhttps://www.kaggle.com/vfomenko/keras-baseline",
    "1110920": "simple solution: Use the onr that is giving YOU the better results; \nand to your doubt many pytorch users uses weights from someone's repo not released by pytorch itself. maybe the ones who have trained it to save the weights have used different machines  or for higher epochs\n(I can be  wrong ) \nOne such awesome PyTorch repo is\nhttps://github.com/rwightman/pytorch-image-models",
    "1110961": "Yeah, I understand that. I just wonder if there are some key differences in the ways the two libraries are implemented that contributes to the performance difference."
  },
  "source": "meta"
}