{
  "id": 238611,
  "title": "Channel-wise Vs Spatial",
  "url": "/competitions/seti-breakthrough-listen/discussion/238611",
  "author_name": "Awsaf",
  "post_date": "2021-05-12T19:13:35.660000",
  "votes": 52,
  "comment_count": 4,
  "views": 0,
  "content": "<h2>Channel-wise Vs Spatial</h2>\n<p>For our given signal we can represent input signal in following 2 ways,</p>\n<ul>\n<li><strong>channel-wise</strong>: considering <code>cadence</code> as image channels. <a href=\"https://www.kaggle.com/awsaf49/seti-bl-tf-starter-tpu\" target=\"_blank\">Notebook|LB: 0.93+</a></li>\n<li><strong>spatial</strong>: considering them as spatial information. <a href=\"https://www.kaggle.com/awsaf49/seti-bl-spatial-info-tf-tpu\" target=\"_blank\">Notebook|LB:0.95+</a></li>\n</ul>\n<p>You can check the following image for better comprehension,<br>\n<a href=\"https://ibb.co/mF8wwTD\"><img src=\"https://i.ibb.co/JFQ44tB/channel-vs-spatial.png\" alt=\"channel-vs-spatial\"></a><br>\nIt turns out <strong>spatial</strong> represntation boost the result around <strong>2%</strong>  on <strong>CV</strong> and <strong>1%</strong> on <strong>LB</strong></p>\n<table>\n<thead>\n<tr>\n<th>Repr.</th>\n<th>CV</th>\n<th>LB</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td><strong>channel</strong></td>\n<td>0.950</td>\n<td>0.93</td>\n</tr>\n<tr>\n<td><strong>spatial</strong></td>\n<td>0.968</td>\n<td>0.95</td>\n</tr>\n</tbody>\n</table>",
  "messages": [
    {
      "id": 1304679,
      "postDate": "2021-05-12T19:13:35.660Z",
      "content": "<h2>Channel-wise Vs Spatial</h2>\n<p>For our given signal we can represent input signal in following 2 ways,</p>\n<ul>\n<li><strong>channel-wise</strong>: considering <code>cadence</code> as image channels. <a href=\"https://www.kaggle.com/awsaf49/seti-bl-tf-starter-tpu\" target=\"_blank\">Notebook|LB: 0.93+</a></li>\n<li><strong>spatial</strong>: considering them as spatial information. <a href=\"https://www.kaggle.com/awsaf49/seti-bl-spatial-info-tf-tpu\" target=\"_blank\">Notebook|LB:0.95+</a></li>\n</ul>\n<p>You can check the following image for better comprehension,<br>\n<a href=\"https://ibb.co/mF8wwTD\"><img src=\"https://i.ibb.co/JFQ44tB/channel-vs-spatial.png\" alt=\"channel-vs-spatial\"></a><br>\nIt turns out <strong>spatial</strong> represntation boost the result around <strong>2%</strong>  on <strong>CV</strong> and <strong>1%</strong> on <strong>LB</strong></p>\n<table>\n<thead>\n<tr>\n<th>Repr.</th>\n<th>CV</th>\n<th>LB</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td><strong>channel</strong></td>\n<td>0.950</td>\n<td>0.93</td>\n</tr>\n<tr>\n<td><strong>spatial</strong></td>\n<td>0.968</td>\n<td>0.95</td>\n</tr>\n</tbody>\n</table>",
      "rawMarkdown": "## Channel-wise Vs Spatial\nFor our given signal we can represent input signal in following 2 ways,\n* **channel-wise**: considering `cadence` as image channels. [Notebook|LB: 0.93+](https://www.kaggle.com/awsaf49/seti-bl-tf-starter-tpu)\n* **spatial**: considering them as spatial information. [Notebook|LB:0.95+](https://www.kaggle.com/awsaf49/seti-bl-spatial-info-tf-tpu)\n\nYou can check the following image for better comprehension,\n<a href=\"https://ibb.co/mF8wwTD\"><img src=\"https://i.ibb.co/JFQ44tB/channel-vs-spatial.png\" alt=\"channel-vs-spatial\" border=\"0\"></a>\nIt turns out **spatial** represntation boost the result around **2%**  on **CV** and **1%** on **LB**\n|Repr.| CV | LB |\n| --- | --- | --- |\n| **channel** | 0.950 | 0.93 |\n| **spatial** | 0.968 | 0.95 |\n",
      "votes": 52
    },
    {
      "id": 1308003,
      "postDate": "2021-05-14T21:20:39.173Z",
      "content": "<p>After 10 experiments, I think spatial is a little better (for now).</p>",
      "rawMarkdown": "After 10 experiments, I think spatial is a little better (for now).",
      "votes": 6,
      "replies": [
        {
          "id": 1338103,
          "postDate": "2021-06-06T07:20:42.543Z",
          "content": "<p>Thanks for sharing, this is very helpful!</p>",
          "rawMarkdown": "Thanks for sharing, this is very helpful!",
          "votes": 2
        }
      ]
    },
    {
      "id": 1340634,
      "postDate": "2021-06-08T06:52:22.003Z",
      "content": "<p>Sorry if this is a silly question, but why would a spatial-wise relation be any different from a channel-wise relation if you're just changing the representation of data? Has it something to do with (GPU) performance?</p>",
      "rawMarkdown": "Sorry if this is a silly question, but why would a spatial-wise relation be any different from a channel-wise relation if you're just changing the representation of data? Has it something to do with (GPU) performance?",
      "replies": [
        {
          "id": 1340642,
          "postDate": "2021-06-08T06:57:51.650Z",
          "content": "<p>adding unnecessary data is effecting the score i guess</p>",
          "rawMarkdown": "adding unnecessary data is effecting the score i guess"
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 1308003,
      "author_name": "sin",
      "author_url": "",
      "post_date": "2021-05-14T21:20:39.173000",
      "content": "<p>After 10 experiments, I think spatial is a little better (for now).</p>",
      "votes": 6,
      "replies": [
        {
          "id": 1338103,
          "author_name": "Old Monk",
          "author_url": "",
          "post_date": "2021-06-06T07:20:42.543000",
          "content": "<p>Thanks for sharing, this is very helpful!</p>",
          "votes": 2,
          "replies": []
        }
      ]
    },
    {
      "id": 1340634,
      "author_name": "Erik Kaufman",
      "author_url": "",
      "post_date": "2021-06-08T06:52:22.003000",
      "content": "<p>Sorry if this is a silly question, but why would a spatial-wise relation be any different from a channel-wise relation if you're just changing the representation of data? Has it something to do with (GPU) performance?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1340642,
          "author_name": "Awsaf",
          "author_url": "",
          "post_date": "2021-06-08T06:57:51.650000",
          "content": "<p>adding unnecessary data is effecting the score i guess</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1304679": "## Channel-wise Vs Spatial\nFor our given signal we can represent input signal in following 2 ways,\n* **channel-wise**: considering `cadence` as image channels. [Notebook|LB: 0.93+](https://www.kaggle.com/awsaf49/seti-bl-tf-starter-tpu)\n* **spatial**: considering them as spatial information. [Notebook|LB:0.95+](https://www.kaggle.com/awsaf49/seti-bl-spatial-info-tf-tpu)\n\nYou can check the following image for better comprehension,\n<a href=\"https://ibb.co/mF8wwTD\"><img src=\"https://i.ibb.co/JFQ44tB/channel-vs-spatial.png\" alt=\"channel-vs-spatial\" border=\"0\"></a>\nIt turns out **spatial** represntation boost the result around **2%**  on **CV** and **1%** on **LB**\n|Repr.| CV | LB |\n| --- | --- | --- |\n| **channel** | 0.950 | 0.93 |\n| **spatial** | 0.968 | 0.95 |\n",
    "1308003": "After 10 experiments, I think spatial is a little better (for now).",
    "1340634": "Sorry if this is a silly question, but why would a spatial-wise relation be any different from a channel-wise relation if you're just changing the representation of data? Has it something to do with (GPU) performance?"
  }
}