{
  "id": 33263,
  "title": "Ways to improve score?",
  "url": "/competitions/digit-recognizer/discussion/33263",
  "author_name": "Pedro Nascimento",
  "post_date": "2017-05-19T20:02:25.340000",
  "votes": 0,
  "comment_count": 1,
  "views": null,
  "content": "<p>Hello everyone</p>\n\n<p>I ran what was pretty much the Deep MNIST tutorial from TensorFlow and got a 99.15% score and was trying to come up with ways to improve it.\nwith a batch size of 50 and 20000 iterations of training it took about 75 mins to train the network on my computer, so I'm not to keen on adding more parameters to my model.</p>\n\n<p>One thing I thought was to train the model on the original MNIST dataset since it contains more examples then the Kaggle training set, so it would give a better fit, with less overfitting, but I figured that some of the entries on the test set would be taken from the MNIST test set, in your opinion would this count as cheating ?</p>\n\n<p>So far I haven't really fiddled with other models, or trying to fine-tune any of the hyper-parameters on my model, so any clues on a good direction to move ahead would be welcome :D</p>\n\n<p>Cheers</p>",
  "messages": [
    {
      "id": 183986,
      "postDate": "2017-05-19T23:53:46.117Z",
      "content": "<p>Hello there!\nI think that was discussed a time ago and using the original dataset from mnist was considered \"cheating\" for the reasons you cited.</p>\n\n<p>I don't know exactly how you have implemented but 99.15% is already a really good score. You could try now add Batch Normalization layers or try to do image augmentation. When I'm training I'm using 95% of data to train and 5% to validate... I think it is a good idea try to shuffle this data and compare the results of multiple runs in different train/test data.</p>\n\n<p>Cheers,</p>",
      "rawMarkdown": "Hello there!\nI think that was discussed a time ago and using the original dataset from mnist was considered \"cheating\" for the reasons you cited.\n\nI don't know exactly how you have implemented but 99.15% is already a really good score. You could try now add Batch Normalization layers or try to do image augmentation. When I'm training I'm using 95% of data to train and 5% to validate... I think it is a good idea try to shuffle this data and compare the results of multiple runs in different train/test data.\n\nCheers,",
      "votes": 1
    },
    {
      "id": 183936,
      "postDate": "2017-05-19T20:02:25.340Z",
      "content": "<p>Hello everyone</p>\n\n<p>I ran what was pretty much the Deep MNIST tutorial from TensorFlow and got a 99.15% score and was trying to come up with ways to improve it.\nwith a batch size of 50 and 20000 iterations of training it took about 75 mins to train the network on my computer, so I'm not to keen on adding more parameters to my model.</p>\n\n<p>One thing I thought was to train the model on the original MNIST dataset since it contains more examples then the Kaggle training set, so it would give a better fit, with less overfitting, but I figured that some of the entries on the test set would be taken from the MNIST test set, in your opinion would this count as cheating ?</p>\n\n<p>So far I haven't really fiddled with other models, or trying to fine-tune any of the hyper-parameters on my model, so any clues on a good direction to move ahead would be welcome :D</p>\n\n<p>Cheers</p>",
      "rawMarkdown": "Hello everyone\n\nI ran what was pretty much the Deep MNIST tutorial from TensorFlow and got a 99.15% score and was trying to come up with ways to improve it.\nwith a batch size of 50 and 20000 iterations of training it took about 75 mins to train the network on my computer, so I'm not to keen on adding more parameters to my model.\n\nOne thing I thought was to train the model on the original MNIST dataset since it contains more examples then the Kaggle training set, so it would give a better fit, with less overfitting, but I figured that some of the entries on the test set would be taken from the MNIST test set, in your opinion would this count as cheating ?\n\nSo far I haven't really fiddled with other models, or trying to fine-tune any of the hyper-parameters on my model, so any clues on a good direction to move ahead would be welcome :D\n\nCheers"
    }
  ],
  "comments": [
    {
      "id": 183986,
      "author_name": "",
      "author_url": "",
      "post_date": "2017-05-19T23:53:46.117000",
      "content": "<p>Hello there!\nI think that was discussed a time ago and using the original dataset from mnist was considered \"cheating\" for the reasons you cited.</p>\n\n<p>I don't know exactly how you have implemented but 99.15% is already a really good score. You could try now add Batch Normalization layers or try to do image augmentation. When I'm training I'm using 95% of data to train and 5% to validate... I think it is a good idea try to shuffle this data and compare the results of multiple runs in different train/test data.</p>\n\n<p>Cheers,</p>",
      "votes": 1,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "183986": "Hello there!\nI think that was discussed a time ago and using the original dataset from mnist was considered \"cheating\" for the reasons you cited.\n\nI don't know exactly how you have implemented but 99.15% is already a really good score. You could try now add Batch Normalization layers or try to do image augmentation. When I'm training I'm using 95% of data to train and 5% to validate... I think it is a good idea try to shuffle this data and compare the results of multiple runs in different train/test data.\n\nCheers,",
    "183936": "Hello everyone\n\nI ran what was pretty much the Deep MNIST tutorial from TensorFlow and got a 99.15% score and was trying to come up with ways to improve it.\nwith a batch size of 50 and 20000 iterations of training it took about 75 mins to train the network on my computer, so I'm not to keen on adding more parameters to my model.\n\nOne thing I thought was to train the model on the original MNIST dataset since it contains more examples then the Kaggle training set, so it would give a better fit, with less overfitting, but I figured that some of the entries on the test set would be taken from the MNIST test set, in your opinion would this count as cheating ?\n\nSo far I haven't really fiddled with other models, or trying to fine-tune any of the hyper-parameters on my model, so any clues on a good direction to move ahead would be welcome :D\n\nCheers"
  }
}