{
  "id": 173262,
  "title": "Removing grayscale images boosts public LB score",
  "url": "/competitions/landmark-retrieval-2020/discussion/173262",
  "author_name": "",
  "post_date": "2020-08-08T14:28:45.891668400Z",
  "votes": 12,
  "comment_count": 5,
  "views": 0,
  "content": "<p>Have you noticed that provided index set contains many gray scale images ?</p>\n\n<p>Since I suspect they are just distractors, I built a model which predict zero vectors for gray scales images by tf.cond.\n<a href=\"https://www.tensorflow.org/api_docs/python/tf/cond\">https://www.tensorflow.org/api_docs/python/tf/cond</a></p>\n\n<p>This postprocessing boosted my public LB score a little.\nEven though this simple way may not useful for private LB, removing distractors will help us in some way. How do you think this idea ?</p>",
  "messages": [
    {
      "id": "962901",
      "postDate": "08/08/2020 14:28:45",
      "content": "<p>Have you noticed that provided index set contains many gray scale images ?</p>\n\n<p>Since I suspect they are just distractors, I built a model which predict zero vectors for gray scales images by tf.cond.\n<a href=\"https://www.tensorflow.org/api_docs/python/tf/cond\">https://www.tensorflow.org/api_docs/python/tf/cond</a></p>\n\n<p>This postprocessing boosted my public LB score a little.\nEven though this simple way may not useful for private LB, removing distractors will help us in some way. How do you think this idea ?</p>",
      "rawMarkdown": "Have you noticed that provided index set contains many gray scale images ?\n\nSince I suspect they are just distractors, I built a model which predict zero vectors for gray scales images by tf.cond.\nhttps://www.tensorflow.org/api_docs/python/tf/cond\n\nThis postprocessing boosted my public LB score a little.\nEven though this simple way may not useful for private LB, removing distractors will help us in some way. How do you think this idea ?",
      "votes": null
    },
    {
      "id": "963994",
      "postDate": "08/09/2020 13:54:26",
      "content": "<p>Thanks for the information. I really appreciate your efforts. I just wanna ask one thing. if you don't mind, would you reveal your hardware specs? GPU, RAM, and train time, just to compare with mine.this is first time to handle this much of data. so to have an idea of approx train time.</p>",
      "rawMarkdown": "Thanks for the information. I really appreciate your efforts. I just wanna ask one thing. if you don't mind, would you reveal your hardware specs? GPU, RAM, and train time, just to compare with mine.this is first time to handle this much of data. so to have an idea of approx train time.",
      "votes": null
    },
    {
      "id": "966558",
      "postDate": "08/11/2020 14:17:34",
      "content": "<p>I'm using one GPU (TITAN RTX) and 256 GB RAM.</p>",
      "rawMarkdown": "I'm using one GPU (TITAN RTX) and 256 GB RAM.",
      "votes": null
    },
    {
      "id": "966700",
      "postDate": "08/11/2020 16:06:18",
      "content": "<p>fantastic…is your framework able to use entire RAM? I am using TF/Keras, not able to use entire RAM efficiently.do you have any idea how to improve it?</p>",
      "rawMarkdown": "fantastic...is your framework able to use entire RAM? I am using TF/Keras, not able to use entire RAM efficiently.do you have any idea how to improve it?",
      "votes": null
    },
    {
      "id": "967695",
      "postDate": "08/12/2020 12:40:27",
      "content": "<p>No, the actual usage of RAM is about 20GB, because I'm loading images from SSD.</p>",
      "rawMarkdown": "No, the actual usage of RAM is about 20GB, because I'm loading images from SSD.",
      "votes": null
    },
    {
      "id": "971003",
      "postDate": "08/15/2020 04:57:56",
      "content": "<p>Tried that, didn't work in my case (lost 0.004 public score). How exactly did you altered the output embeddings in that case? (I just added a coordinate to separate the point clouds)</p>",
      "rawMarkdown": "Tried that, didn't work in my case (lost 0.004 public score). How exactly did you altered the output embeddings in that case? (I just added a coordinate to separate the point clouds)",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 971003,
      "author_name": "chankhavu",
      "author_url": "",
      "post_date": "08/15/2020 04:57:56",
      "content": "<p>Tried that, didn't work in my case (lost 0.004 public score). How exactly did you altered the output embeddings in that case? (I just added a coordinate to separate the point clouds)</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 963994,
      "author_name": "udaygurugubelli",
      "author_url": "",
      "post_date": "08/09/2020 13:54:26",
      "content": "<p>Thanks for the information. I really appreciate your efforts. I just wanna ask one thing. if you don't mind, would you reveal your hardware specs? GPU, RAM, and train time, just to compare with mine.this is first time to handle this much of data. so to have an idea of approx train time.</p>",
      "votes": null,
      "replies": [
        {
          "id": 966558,
          "author_name": "toshik",
          "author_url": "",
          "post_date": "08/11/2020 14:17:34",
          "content": "<p>I'm using one GPU (TITAN RTX) and 256 GB RAM.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 966700,
          "author_name": "udaygurugubelli",
          "author_url": "",
          "post_date": "08/11/2020 16:06:18",
          "content": "<p>fantastic…is your framework able to use entire RAM? I am using TF/Keras, not able to use entire RAM efficiently.do you have any idea how to improve it?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 967695,
          "author_name": "toshik",
          "author_url": "",
          "post_date": "08/12/2020 12:40:27",
          "content": "<p>No, the actual usage of RAM is about 20GB, because I'm loading images from SSD.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "962901": "Have you noticed that provided index set contains many gray scale images ?\n\nSince I suspect they are just distractors, I built a model which predict zero vectors for gray scales images by tf.cond.\nhttps://www.tensorflow.org/api_docs/python/tf/cond\n\nThis postprocessing boosted my public LB score a little.\nEven though this simple way may not useful for private LB, removing distractors will help us in some way. How do you think this idea ?",
    "963994": "Thanks for the information. I really appreciate your efforts. I just wanna ask one thing. if you don't mind, would you reveal your hardware specs? GPU, RAM, and train time, just to compare with mine.this is first time to handle this much of data. so to have an idea of approx train time.",
    "966558": "I'm using one GPU (TITAN RTX) and 256 GB RAM.",
    "966700": "fantastic...is your framework able to use entire RAM? I am using TF/Keras, not able to use entire RAM efficiently.do you have any idea how to improve it?",
    "967695": "No, the actual usage of RAM is about 20GB, because I'm loading images from SSD.",
    "971003": "Tried that, didn't work in my case (lost 0.004 public score). How exactly did you altered the output embeddings in that case? (I just added a coordinate to separate the point clouds)"
  },
  "source": "meta"
}