{
  "id": 238595,
  "title": "My submission",
  "url": "/competitions/digit-recognizer/discussion/238595",
  "author_name": "sangstar",
  "post_date": "2021-05-12T18:05:16.827000",
  "votes": 1,
  "comment_count": 3,
  "views": null,
  "content": "<p>Hello! My work for this submission is here: <a href=\"https://github.com/sangstar/kaggle-digit-recognizer\" target=\"_blank\">https://github.com/sangstar/kaggle-digit-recognizer</a></p>\n<p>Feel free to do with it as you wish, and feel free to give me any tips!</p>",
  "messages": [
    {
      "id": 1304588,
      "postDate": "2021-05-12T18:05:16.827Z",
      "content": "<p>Hello! My work for this submission is here: <a href=\"https://github.com/sangstar/kaggle-digit-recognizer\" target=\"_blank\">https://github.com/sangstar/kaggle-digit-recognizer</a></p>\n<p>Feel free to do with it as you wish, and feel free to give me any tips!</p>",
      "rawMarkdown": "Hello! My work for this submission is here: https://github.com/sangstar/kaggle-digit-recognizer\n\nFeel free to do with it as you wish, and feel free to give me any tips!\n\n",
      "votes": 1
    },
    {
      "id": 1373687,
      "postDate": "2021-07-02T17:08:49.657Z",
      "content": "<p>Hi,<br>\nI'm not in Data Science for long but I hope that I can help a bit. <br>\nFirst, I got a glance at your code on github and it is a clear and diligent notebook.<br>\nWhat I would suggest is : <br>\n1/ if your accuracy on the training set is 99.28%, it is unlikely that you will get a higher score on the test set. So you should try to improve your score on the training set. How to do this? It's hard to say but training longer or increasing the complexity of your NN should help. Once you are satisfied with the training accuracy, check if the model generalizes well. If not, you can either add regularizations techniques (e.g. increasing your dropout) or increasing your training set (using data augmentation)<br>\n2/ A good practice also is to split your training set in 2 parts : training set + validation set. Allocate 80 or 90% of the images to the training set and the remaining to the validation set. Train your model on the training set and checks how well it generalize on the validation set. You can then train several NN and pick the one that gives the best result on the validation set.<br>\nGood luck!</p>",
      "rawMarkdown": "Hi,\nI'm not in Data Science for long but I hope that I can help a bit. \nFirst, I got a glance at your code on github and it is a clear and diligent notebook.\nWhat I would suggest is : \n1/ if your accuracy on the training set is 99.28%, it is unlikely that you will get a higher score on the test set. So you should try to improve your score on the training set. How to do this? It's hard to say but training longer or increasing the complexity of your NN should help. Once you are satisfied with the training accuracy, check if the model generalizes well. If not, you can either add regularizations techniques (e.g. increasing your dropout) or increasing your training set (using data augmentation)\n2/ A good practice also is to split your training set in 2 parts : training set + validation set. Allocate 80 or 90% of the images to the training set and the remaining to the validation set. Train your model on the training set and checks how well it generalize on the validation set. You can then train several NN and pick the one that gives the best result on the validation set.\nGood luck!\n"
    },
    {
      "id": 1310598,
      "postDate": "2021-05-16T18:46:12.877Z",
      "content": "<p>Nice work :-) I read your programming code on GitHub. How about stacking conv2d layers? Usually, next to conv2d is maxpooling, though. Here's my programming code. I don't have such a higher score, but stacking two layers like conv2d next to conv2d improved my code's accuracy for MNIST.   </p>\n<p><a href=\"https://www.kaggle.com/nohrud/cnn-for-mnist-in-a-simple-way\" target=\"_blank\">https://www.kaggle.com/nohrud/cnn-for-mnist-in-a-simple-way</a></p>",
      "rawMarkdown": "Nice work :-) I read your programming code on GitHub. How about stacking conv2d layers? Usually, next to conv2d is maxpooling, though. Here's my programming code. I don't have such a higher score, but stacking two layers like conv2d next to conv2d improved my code's accuracy for MNIST.   \n\nhttps://www.kaggle.com/nohrud/cnn-for-mnist-in-a-simple-way"
    },
    {
      "id": 1306643,
      "postDate": "2021-05-14T00:32:08.060Z",
      "content": "<p>Looks great! Maybe one more step for evaluation / visualization?</p>\n<p>i'd appreciate if you check out <a href=\"https://www.kaggle.com/sinanenocak/digit-recognizer-convolutional-neural-network\" target=\"_blank\">my submission</a> as well!</p>",
      "rawMarkdown": "Looks great! Maybe one more step for evaluation / visualization?\n\ni'd appreciate if you check out [my submission](https://www.kaggle.com/sinanenocak/digit-recognizer-convolutional-neural-network) as well!"
    }
  ],
  "comments": [
    {
      "id": 1373687,
      "author_name": "Arnaud",
      "author_url": "",
      "post_date": "2021-07-02T17:08:49.657000",
      "content": "<p>Hi,<br>\nI'm not in Data Science for long but I hope that I can help a bit. <br>\nFirst, I got a glance at your code on github and it is a clear and diligent notebook.<br>\nWhat I would suggest is : <br>\n1/ if your accuracy on the training set is 99.28%, it is unlikely that you will get a higher score on the test set. So you should try to improve your score on the training set. How to do this? It's hard to say but training longer or increasing the complexity of your NN should help. Once you are satisfied with the training accuracy, check if the model generalizes well. If not, you can either add regularizations techniques (e.g. increasing your dropout) or increasing your training set (using data augmentation)<br>\n2/ A good practice also is to split your training set in 2 parts : training set + validation set. Allocate 80 or 90% of the images to the training set and the remaining to the validation set. Train your model on the training set and checks how well it generalize on the validation set. You can then train several NN and pick the one that gives the best result on the validation set.<br>\nGood luck!</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1310598,
      "author_name": "NohrUd",
      "author_url": "",
      "post_date": "2021-05-16T18:46:12.877000",
      "content": "<p>Nice work :-) I read your programming code on GitHub. How about stacking conv2d layers? Usually, next to conv2d is maxpooling, though. Here's my programming code. I don't have such a higher score, but stacking two layers like conv2d next to conv2d improved my code's accuracy for MNIST.   </p>\n<p><a href=\"https://www.kaggle.com/nohrud/cnn-for-mnist-in-a-simple-way\" target=\"_blank\">https://www.kaggle.com/nohrud/cnn-for-mnist-in-a-simple-way</a></p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1306643,
      "author_name": "Sinan Şenocak",
      "author_url": "",
      "post_date": "2021-05-14T00:32:08.060000",
      "content": "<p>Looks great! Maybe one more step for evaluation / visualization?</p>\n<p>i'd appreciate if you check out <a href=\"https://www.kaggle.com/sinanenocak/digit-recognizer-convolutional-neural-network\" target=\"_blank\">my submission</a> as well!</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1304588": "Hello! My work for this submission is here: https://github.com/sangstar/kaggle-digit-recognizer\n\nFeel free to do with it as you wish, and feel free to give me any tips!\n\n",
    "1373687": "Hi,\nI'm not in Data Science for long but I hope that I can help a bit. \nFirst, I got a glance at your code on github and it is a clear and diligent notebook.\nWhat I would suggest is : \n1/ if your accuracy on the training set is 99.28%, it is unlikely that you will get a higher score on the test set. So you should try to improve your score on the training set. How to do this? It's hard to say but training longer or increasing the complexity of your NN should help. Once you are satisfied with the training accuracy, check if the model generalizes well. If not, you can either add regularizations techniques (e.g. increasing your dropout) or increasing your training set (using data augmentation)\n2/ A good practice also is to split your training set in 2 parts : training set + validation set. Allocate 80 or 90% of the images to the training set and the remaining to the validation set. Train your model on the training set and checks how well it generalize on the validation set. You can then train several NN and pick the one that gives the best result on the validation set.\nGood luck!\n",
    "1310598": "Nice work :-) I read your programming code on GitHub. How about stacking conv2d layers? Usually, next to conv2d is maxpooling, though. Here's my programming code. I don't have such a higher score, but stacking two layers like conv2d next to conv2d improved my code's accuracy for MNIST.   \n\nhttps://www.kaggle.com/nohrud/cnn-for-mnist-in-a-simple-way",
    "1306643": "Looks great! Maybe one more step for evaluation / visualization?\n\ni'd appreciate if you check out [my submission](https://www.kaggle.com/sinanenocak/digit-recognizer-convolutional-neural-network) as well!"
  }
}