{
  "id": 217180,
  "title": "Queries regarding TF 2 Object Detection API",
  "url": "/competitions/vinbigdata-chest-xray-abnormalities-detection/discussion/217180",
  "author_name": "Akshit Bhalla",
  "post_date": "2021-02-05T16:47:16.090000",
  "votes": 3,
  "comment_count": 0,
  "views": 0,
  "content": "<p>I'm learning to build models with the TF 2 Object Detection API. The <a href=\"https://tensorflow-object-detection-api-tutorial.readthedocs.io/en/latest/training.html\">official tutorial</a> helped me get started but does not cover many topics clearly. Additionally, there are some things that are hard to accomplish on Kaggle. I'm putting them down here.</p>\n<ol>\n<li>According to the official tutorial, we can obtain mAP on the validation dataset by simultaneously running an evaluation script alongside the training script. On a regular machine, this is done by opening another terminal. How can one accomplish this on Kaggle?</li>\n<li>How do I obtain mAP on the test (not validation) dataset after training the model? The official tutorial makes no mention.</li>\n<li>How do I use TensorBoard on Kaggle? I was told this can be done with NGROK but I'm not sure how.</li>\n<li>I would like to plot a <a href=\"https://miro.medium.com/max/760/1*kmAon_ut7ZU3h0XkGbqR6g.png\">precision-recall curve</a> and if possible, precision and recall against threshold. How can this be done (preferably staying within TF ecosystem)?</li>\n<li>How do I make the entire pipeline reproducible? Usually, randomness is controlled using a seed but no such option seems to be available. I would like to run everything in such a way that I get the same error, results, etc each time.</li>\n</ol>\n<p>Help from my fellow Kagglers is much appreciated. I looked up the official tutorial, several blogs, and some answers on StackOverflow but they did not help much. </p>\n<p>Do comment or reach out to me for clarifications, if any. </p>\n<p>Thanks a lot in advance. </p>",
  "messages": [
    {
      "id": 1187712,
      "postDate": "2021-02-05T16:47:16.090Z",
      "content": "<p>I'm learning to build models with the TF 2 Object Detection API. The <a href=\"https://tensorflow-object-detection-api-tutorial.readthedocs.io/en/latest/training.html\">official tutorial</a> helped me get started but does not cover many topics clearly. Additionally, there are some things that are hard to accomplish on Kaggle. I'm putting them down here.</p>\n<ol>\n<li>According to the official tutorial, we can obtain mAP on the validation dataset by simultaneously running an evaluation script alongside the training script. On a regular machine, this is done by opening another terminal. How can one accomplish this on Kaggle?</li>\n<li>How do I obtain mAP on the test (not validation) dataset after training the model? The official tutorial makes no mention.</li>\n<li>How do I use TensorBoard on Kaggle? I was told this can be done with NGROK but I'm not sure how.</li>\n<li>I would like to plot a <a href=\"https://miro.medium.com/max/760/1*kmAon_ut7ZU3h0XkGbqR6g.png\">precision-recall curve</a> and if possible, precision and recall against threshold. How can this be done (preferably staying within TF ecosystem)?</li>\n<li>How do I make the entire pipeline reproducible? Usually, randomness is controlled using a seed but no such option seems to be available. I would like to run everything in such a way that I get the same error, results, etc each time.</li>\n</ol>\n<p>Help from my fellow Kagglers is much appreciated. I looked up the official tutorial, several blogs, and some answers on StackOverflow but they did not help much. </p>\n<p>Do comment or reach out to me for clarifications, if any. </p>\n<p>Thanks a lot in advance. </p>",
      "rawMarkdown": "I'm learning to build models with the TF 2 Object Detection API. The <a href=\"https://tensorflow-object-detection-api-tutorial.readthedocs.io/en/latest/training.html\">official tutorial</a> helped me get started but does not cover many topics clearly. Additionally, there are some things that are hard to accomplish on Kaggle. I'm putting them down here.\n1. According to the official tutorial, we can obtain mAP on the validation dataset by simultaneously running an evaluation script alongside the training script. On a regular machine, this is done by opening another terminal. How can one accomplish this on Kaggle?\n2. How do I obtain mAP on the test (not validation) dataset after training the model? The official tutorial makes no mention.\n3. How do I use TensorBoard on Kaggle? I was told this can be done with NGROK but I'm not sure how.\n4. I would like to plot a <a href=\"https://miro.medium.com/max/760/1*kmAon_ut7ZU3h0XkGbqR6g.png\">precision-recall curve</a> and if possible, precision and recall against threshold. How can this be done (preferably staying within TF ecosystem)?\n5. How do I make the entire pipeline reproducible? Usually, randomness is controlled using a seed but no such option seems to be available. I would like to run everything in such a way that I get the same error, results, etc each time.\n\nHelp from my fellow Kagglers is much appreciated. I looked up the official tutorial, several blogs, and some answers on StackOverflow but they did not help much. \n\nDo comment or reach out to me for clarifications, if any. \n\nThanks a lot in advance. ",
      "votes": 3
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1187712": "I'm learning to build models with the TF 2 Object Detection API. The <a href=\"https://tensorflow-object-detection-api-tutorial.readthedocs.io/en/latest/training.html\">official tutorial</a> helped me get started but does not cover many topics clearly. Additionally, there are some things that are hard to accomplish on Kaggle. I'm putting them down here.\n1. According to the official tutorial, we can obtain mAP on the validation dataset by simultaneously running an evaluation script alongside the training script. On a regular machine, this is done by opening another terminal. How can one accomplish this on Kaggle?\n2. How do I obtain mAP on the test (not validation) dataset after training the model? The official tutorial makes no mention.\n3. How do I use TensorBoard on Kaggle? I was told this can be done with NGROK but I'm not sure how.\n4. I would like to plot a <a href=\"https://miro.medium.com/max/760/1*kmAon_ut7ZU3h0XkGbqR6g.png\">precision-recall curve</a> and if possible, precision and recall against threshold. How can this be done (preferably staying within TF ecosystem)?\n5. How do I make the entire pipeline reproducible? Usually, randomness is controlled using a seed but no such option seems to be available. I would like to run everything in such a way that I get the same error, results, etc each time.\n\nHelp from my fellow Kagglers is much appreciated. I looked up the official tutorial, several blogs, and some answers on StackOverflow but they did not help much. \n\nDo comment or reach out to me for clarifications, if any. \n\nThanks a lot in advance. "
  }
}