{
  "id": 71098,
  "title": "NN model and its performance",
  "url": "/competitions/quora-insincere-questions-classification/discussion/71098",
  "author_name": "",
  "post_date": "2018-11-10T05:37:55.875164Z",
  "votes": 1,
  "comment_count": 3,
  "views": 0,
  "content": "<p>I think besides many good public kernel, it is still worth to have a post to discuss different NN model and performance. Please share your network structure here and the LB score and validation score</p>",
  "messages": [
    {
      "id": "418557",
      "postDate": "11/10/2018 05:37:55",
      "content": "<p>I think besides many good public kernel, it is still worth to have a post to discuss different NN model and performance. Please share your network structure here and the LB score and validation score</p>",
      "rawMarkdown": "I think besides many good public kernel, it is still worth to have a post to discuss different NN model and performance. Please share your network structure here and the LB score and validation score",
      "votes": null
    },
    {
      "id": "418558",
      "postDate": "11/10/2018 05:40:42",
      "content": "<p>Here is my model with Glove embeddings:\nTwo layer of bidirectional  GRU, and globalMaxPooling concatenate with GlobalAvePooling, then directly dense layer with sigmoid activation\nLoval validation: 0.681, LB: 0.667</p>",
      "rawMarkdown": "Here is my model with Glove embeddings:\nTwo layer of bidirectional  GRU, and globalMaxPooling concatenate with GlobalAvePooling, then directly dense layer with sigmoid activation\nLoval validation: 0.681, LB: 0.667",
      "votes": null
    },
    {
      "id": "418560",
      "postDate": "11/10/2018 05:42:06",
      "content": "<p>It seems that it is very easy to overfitting, I tried several different model, but best F1 score appears just after 1 or 2 epoches training</p>",
      "rawMarkdown": "It seems that it is very easy to overfitting, I tried several different model, but best F1 score appears just after 1 or 2 epoches training",
      "votes": null
    },
    {
      "id": "418611",
      "postDate": "11/10/2018 09:09:49",
      "content": "<p>Also Inception like CNN seems to perform well. (0.67 LB) See <a href=\"https://www.kaggle.com/christofhenkel/inceptioncnn-with-flip\">https://www.kaggle.com/christofhenkel/inceptioncnn-with-flip</a> . Have not done a test how good it ensembles with RNN though.</p>",
      "rawMarkdown": "Also Inception like CNN seems to perform well. (0.67 LB) See https://www.kaggle.com/christofhenkel/inceptioncnn-with-flip . Have not done a test how good it ensembles with RNN though.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 418558,
      "author_name": "strideradu",
      "author_url": "",
      "post_date": "11/10/2018 05:40:42",
      "content": "<p>Here is my model with Glove embeddings:\nTwo layer of bidirectional  GRU, and globalMaxPooling concatenate with GlobalAvePooling, then directly dense layer with sigmoid activation\nLoval validation: 0.681, LB: 0.667</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 418560,
      "author_name": "strideradu",
      "author_url": "",
      "post_date": "11/10/2018 05:42:06",
      "content": "<p>It seems that it is very easy to overfitting, I tried several different model, but best F1 score appears just after 1 or 2 epoches training</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 418611,
      "author_name": "christofhenkel",
      "author_url": "",
      "post_date": "11/10/2018 09:09:49",
      "content": "<p>Also Inception like CNN seems to perform well. (0.67 LB) See <a href=\"https://www.kaggle.com/christofhenkel/inceptioncnn-with-flip\">https://www.kaggle.com/christofhenkel/inceptioncnn-with-flip</a> . Have not done a test how good it ensembles with RNN though.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "418557": "I think besides many good public kernel, it is still worth to have a post to discuss different NN model and performance. Please share your network structure here and the LB score and validation score",
    "418558": "Here is my model with Glove embeddings:\nTwo layer of bidirectional  GRU, and globalMaxPooling concatenate with GlobalAvePooling, then directly dense layer with sigmoid activation\nLoval validation: 0.681, LB: 0.667",
    "418560": "It seems that it is very easy to overfitting, I tried several different model, but best F1 score appears just after 1 or 2 epoches training",
    "418611": "Also Inception like CNN seems to perform well. (0.67 LB) See https://www.kaggle.com/christofhenkel/inceptioncnn-with-flip . Have not done a test how good it ensembles with RNN though."
  },
  "source": "meta"
}