{
  "id": 418296,
  "title": "What I learned in this competition",
  "url": "/competitions/benetech-making-graphs-accessible/discussion/418296",
  "author_name": "",
  "post_date": "2023-06-19T21:53:28.890546300Z",
  "votes": 15,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Teaming - I had the pleasure of teaming with <a href=\"https://www.kaggle.com/alejopaullier\" target=\"_blank\">@alejopaullier</a> who showed me a lot of new things and gave a ton of interesting ideas that were surely the majority of the reason we placed well</p>\n<p>Training models with GPU - This was my first time needing 36GB of CPU so there was definitely a learning curve here. Image models take far more space than simpler tabular data so this is a must. <a href=\"https://www.kaggle.com/alejopaullier\" target=\"_blank\">@alejopaullier</a>’s idea to use AMP and reduce precision definitely helped here</p>\n<p>Working with jsons - In order to make the exploratory analysis and evaluate models on CV being able to navigate jsons was a needed and I hadn't had to do this frequently yet</p>\n<p>Cross-validation with image models - I had actually never done this in this manner so being able to do cross validation and use the benetech metric with the annotation files was interesting </p>\n<p>Tricks for training image models - Image models are famously quite beefy so learning more about the training process, precision, batching, and hyperparameters effects on GPU and memory go a long way</p>\n<p>Using matcha and donut - Nicholas’ donut model was a great starting place and gave me the chance to learn more about the generation step. My original matcha model was not great because I was trying to run it locally. <a href=\"https://www.kaggle.com/alejopaullier\" target=\"_blank\">@alejopaullier</a>’s introduction to his code on google colab made things much more clear </p>\n<p>All and all there have been many ups and downs with the drama of this competition but I have learned oh so much and I hope you all did too!  </p>",
  "messages": [
    {
      "id": "2309767",
      "postDate": "06/19/2023 21:53:28",
      "content": "<p>Teaming - I had the pleasure of teaming with <a href=\"https://www.kaggle.com/alejopaullier\" target=\"_blank\">@alejopaullier</a> who showed me a lot of new things and gave a ton of interesting ideas that were surely the majority of the reason we placed well</p>\n<p>Training models with GPU - This was my first time needing 36GB of CPU so there was definitely a learning curve here. Image models take far more space than simpler tabular data so this is a must. <a href=\"https://www.kaggle.com/alejopaullier\" target=\"_blank\">@alejopaullier</a>’s idea to use AMP and reduce precision definitely helped here</p>\n<p>Working with jsons - In order to make the exploratory analysis and evaluate models on CV being able to navigate jsons was a needed and I hadn't had to do this frequently yet</p>\n<p>Cross-validation with image models - I had actually never done this in this manner so being able to do cross validation and use the benetech metric with the annotation files was interesting </p>\n<p>Tricks for training image models - Image models are famously quite beefy so learning more about the training process, precision, batching, and hyperparameters effects on GPU and memory go a long way</p>\n<p>Using matcha and donut - Nicholas’ donut model was a great starting place and gave me the chance to learn more about the generation step. My original matcha model was not great because I was trying to run it locally. <a href=\"https://www.kaggle.com/alejopaullier\" target=\"_blank\">@alejopaullier</a>’s introduction to his code on google colab made things much more clear </p>\n<p>All and all there have been many ups and downs with the drama of this competition but I have learned oh so much and I hope you all did too!  </p>",
      "rawMarkdown": "Teaming - I had the pleasure of teaming with @alejopaullier who showed me a lot of new things and gave a ton of interesting ideas that were surely the majority of the reason we placed well\n\nTraining models with GPU - This was my first time needing 36GB of CPU so there was definitely a learning curve here. Image models take far more space than simpler tabular data so this is a must. @alejopaullier’s idea to use AMP and reduce precision definitely helped here\n\nWorking with jsons - In order to make the exploratory analysis and evaluate models on CV being able to navigate jsons was a needed and I hadn't had to do this frequently yet\n\nCross-validation with image models - I had actually never done this in this manner so being able to do cross validation and use the benetech metric with the annotation files was interesting \n\nTricks for training image models - Image models are famously quite beefy so learning more about the training process, precision, batching, and hyperparameters effects on GPU and memory go a long way\n\nUsing matcha and donut - Nicholas’ donut model was a great starting place and gave me the chance to learn more about the generation step. My original matcha model was not great because I was trying to run it locally. @alejopaullier’s introduction to his code on google colab made things much more clear \n\nAll and all there have been many ups and downs with the drama of this competition but I have learned oh so much and I hope you all did too!",
      "votes": null
    },
    {
      "id": "2309892",
      "postDate": "06/20/2023 00:54:14",
      "content": "<p>It was a pleasure teaming up <a href=\"https://www.kaggle.com/cody11null\" target=\"_blank\">@cody11null</a> we learned a lot!</p>",
      "rawMarkdown": "It was a pleasure teaming up @cody11null we learned a lot!",
      "votes": null
    },
    {
      "id": "2310098",
      "postDate": "06/20/2023 05:34:56",
      "content": "<p><a href=\"https://www.kaggle.com/cody11null\" target=\"_blank\">@cody11null</a> It's great that you were able to improve your skills so well!</p>",
      "rawMarkdown": "cody11null It's great that you were able to improve your skills so well!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2309892,
      "author_name": "alejopaullier",
      "author_url": "",
      "post_date": "06/20/2023 00:54:14",
      "content": "<p>It was a pleasure teaming up <a href=\"https://www.kaggle.com/cody11null\" target=\"_blank\">@cody11null</a> we learned a lot!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 2310098,
      "author_name": "serangu",
      "author_url": "",
      "post_date": "06/20/2023 05:34:56",
      "content": "<p><a href=\"https://www.kaggle.com/cody11null\" target=\"_blank\">@cody11null</a> It's great that you were able to improve your skills so well!</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2309767": "Teaming - I had the pleasure of teaming with @alejopaullier who showed me a lot of new things and gave a ton of interesting ideas that were surely the majority of the reason we placed well\n\nTraining models with GPU - This was my first time needing 36GB of CPU so there was definitely a learning curve here. Image models take far more space than simpler tabular data so this is a must. @alejopaullier’s idea to use AMP and reduce precision definitely helped here\n\nWorking with jsons - In order to make the exploratory analysis and evaluate models on CV being able to navigate jsons was a needed and I hadn't had to do this frequently yet\n\nCross-validation with image models - I had actually never done this in this manner so being able to do cross validation and use the benetech metric with the annotation files was interesting \n\nTricks for training image models - Image models are famously quite beefy so learning more about the training process, precision, batching, and hyperparameters effects on GPU and memory go a long way\n\nUsing matcha and donut - Nicholas’ donut model was a great starting place and gave me the chance to learn more about the generation step. My original matcha model was not great because I was trying to run it locally. @alejopaullier’s introduction to his code on google colab made things much more clear \n\nAll and all there have been many ups and downs with the drama of this competition but I have learned oh so much and I hope you all did too!",
    "2309892": "It was a pleasure teaming up @cody11null we learned a lot!",
    "2310098": "cody11null It's great that you were able to improve your skills so well!"
  },
  "source": "meta"
}