{
  "id": 63071,
  "title": "Is pre-training effective?",
  "url": "/competitions/airbus-ship-detection/discussion/63071",
  "author_name": "",
  "post_date": "2018-08-11T13:13:55.783979500Z",
  "votes": 2,
  "comment_count": 2,
  "views": 0,
  "content": "<p>At the training, I want to know pre-training is effective or not. \nAnd if it's effective, how to pre-training? \nI thought when pre-training we should use gray-scaled original image as teacher data. \nI want to know whether or not there is another way to pre-train.</p>",
  "messages": [
    {
      "id": "368972",
      "postDate": "08/11/2018 13:13:55",
      "content": "<p>At the training, I want to know pre-training is effective or not. \nAnd if it's effective, how to pre-training? \nI thought when pre-training we should use gray-scaled original image as teacher data. \nI want to know whether or not there is another way to pre-train.</p>",
      "rawMarkdown": "At the training, I want to know pre-training is effective or not. \nAnd if it's effective, how to pre-training? \nI thought when pre-training we should use gray-scaled original image as teacher data. \nI want to know whether or not there is another way to pre-train.",
      "votes": null
    },
    {
      "id": "370568",
      "postDate": "08/15/2018 03:16:04",
      "content": "<p>pretraining is effective.  Except that the rules of this competition aren't clear if we can pretrain a model with external data.</p>\n\n<p>How to pretrain a model?  Pretraining is just training.  </p>\n\n<p>Normally you would just train a complete model normally or use a famous pretrained one and use just the convolutional layers to transform your data in to features and train a new network on the end of it.  That way you get the benefits of edge detection and etc from the pretrained model.  This is known as \"Transfer Learning\".</p>\n\n<p>Or you can use the pretrained model and let it train on newer data which would require you to allow it to back propagate to the filters and be in tune with the new data.  This is called \"Fine Tuning\" </p>\n\n<p>Just remember a pretrained model means that the filters have already been trained to detect some type of feature.</p>\n\n<p>Your statement \"I thought when pre-training we should use gray-scaled original image as teacher data.\", makes me want to explain what pre-training is.  Pre-training is just training like normal.  The only thing is that you are now using the trained filters (the feature extraction part) for extraction and then wanting to map those features to a new problem and update a logical layer.</p>",
      "rawMarkdown": "pretraining is effective.  Except that the rules of this competition aren't clear if we can pretrain a model with external data.\n\nHow to pretrain a model?  Pretraining is just training.  \n\nNormally you would just train a complete model normally or use a famous pretrained one and use just the convolutional layers to transform your data in to features and train a new network on the end of it.  That way you get the benefits of edge detection and etc from the pretrained model.  This is known as \"Transfer Learning\".\n\nOr you can use the pretrained model and let it train on newer data which would require you to allow it to back propagate to the filters and be in tune with the new data.  This is called \"Fine Tuning\" \n\nJust remember a pretrained model means that the filters have already been trained to detect some type of feature.\n\nYour statement \"I thought when pre-training we should use gray-scaled original image as teacher data.\", makes me want to explain what pre-training is.  Pre-training is just training like normal.  The only thing is that you are now using the trained filters (the feature extraction part) for extraction and then wanting to map those features to a new problem and update a logical layer.",
      "votes": null
    },
    {
      "id": "371352",
      "postDate": "08/16/2018 15:28:34",
      "content": "<p>Oh, I misunderstood.</p>\n\n<p>Thank you for your teaching. \nI will try some.</p>",
      "rawMarkdown": "Oh, I misunderstood.\n\nThank you for your teaching. \nI will try some.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 370568,
      "author_name": "bgdev5279",
      "author_url": "",
      "post_date": "08/15/2018 03:16:04",
      "content": "<p>pretraining is effective.  Except that the rules of this competition aren't clear if we can pretrain a model with external data.</p>\n\n<p>How to pretrain a model?  Pretraining is just training.  </p>\n\n<p>Normally you would just train a complete model normally or use a famous pretrained one and use just the convolutional layers to transform your data in to features and train a new network on the end of it.  That way you get the benefits of edge detection and etc from the pretrained model.  This is known as \"Transfer Learning\".</p>\n\n<p>Or you can use the pretrained model and let it train on newer data which would require you to allow it to back propagate to the filters and be in tune with the new data.  This is called \"Fine Tuning\" </p>\n\n<p>Just remember a pretrained model means that the filters have already been trained to detect some type of feature.</p>\n\n<p>Your statement \"I thought when pre-training we should use gray-scaled original image as teacher data.\", makes me want to explain what pre-training is.  Pre-training is just training like normal.  The only thing is that you are now using the trained filters (the feature extraction part) for extraction and then wanting to map those features to a new problem and update a logical layer.</p>",
      "votes": null,
      "replies": [
        {
          "id": 371352,
          "author_name": "sunshine87345",
          "author_url": "",
          "post_date": "08/16/2018 15:28:34",
          "content": "<p>Oh, I misunderstood.</p>\n\n<p>Thank you for your teaching. \nI will try some.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "368972": "At the training, I want to know pre-training is effective or not. \nAnd if it's effective, how to pre-training? \nI thought when pre-training we should use gray-scaled original image as teacher data. \nI want to know whether or not there is another way to pre-train.",
    "370568": "pretraining is effective.  Except that the rules of this competition aren't clear if we can pretrain a model with external data.\n\nHow to pretrain a model?  Pretraining is just training.  \n\nNormally you would just train a complete model normally or use a famous pretrained one and use just the convolutional layers to transform your data in to features and train a new network on the end of it.  That way you get the benefits of edge detection and etc from the pretrained model.  This is known as \"Transfer Learning\".\n\nOr you can use the pretrained model and let it train on newer data which would require you to allow it to back propagate to the filters and be in tune with the new data.  This is called \"Fine Tuning\" \n\nJust remember a pretrained model means that the filters have already been trained to detect some type of feature.\n\nYour statement \"I thought when pre-training we should use gray-scaled original image as teacher data.\", makes me want to explain what pre-training is.  Pre-training is just training like normal.  The only thing is that you are now using the trained filters (the feature extraction part) for extraction and then wanting to map those features to a new problem and update a logical layer.",
    "371352": "Oh, I misunderstood.\n\nThank you for your teaching. \nI will try some."
  },
  "source": "meta"
}