{
  "id": 11531,
  "title": "Show filters from the 1st convolution layer",
  "url": "/competitions/datasciencebowl/discussion/11531",
  "author_name": "",
  "post_date": "2015-01-16T10:56:22.267Z",
  "votes": null,
  "comment_count": 12,
  "views": 4129,
  "content": "<p>I trained a DCNN model and observed the first convolution layers. Can anyone share it ?</p>",
  "messages": [
    {
      "id": "62786",
      "postDate": "01/16/2015 10:56:22",
      "content": "<p>I trained a DCNN model and observed the first convolution layers. Can anyone share it ?</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "62788",
      "postDate": "01/16/2015 12:01:28",
      "content": "<p>Why they are in colour? They seem strangely familiar but not this dataset.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "62790",
      "postDate": "01/16/2015 13:05:49",
      "content": "",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "62791",
      "postDate": "01/16/2015 13:09:07",
      "content": "<p>Yes, you are right. it uses pretrained model, although it isn't permitted.</p>\n<p>The following is conv1 trained using provided data:</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "62808",
      "postDate": "01/17/2015 00:13:51",
      "content": "<p>Are the colours supposed to represent pixel intensities?&nbsp;<br><br>I'm curious how those colours were generated considering the dataset is provided in grayscale.&nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "62819",
      "postDate": "01/17/2015 08:39:46",
      "content": "<p>I'd guess if you took a CNN originally designed for&nbsp;a colour dataset, and trained it on a grey image expanded to RGB, then rendered as if it was RGB-based, then you might&nbsp;see three overlaid random greyscale filters in each cell. I think BigFish's second post looks like it could be that.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "62942",
      "postDate": "01/20/2015 23:09:01",
      "content": "<p>Here's one of mine. All my filters are 3x3</p>\n<p>(obviously, epoch_065. The 000 one is after just one training epoch, and was accidentally added - but I can't seem to remove attachments)</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "62945",
      "postDate": "01/21/2015 01:10:55",
      "content": "<p>Here are a few of my filters from mid-training...nothing to write home about.&nbsp; They're learning but I wouldn't mind if they didn't take their sweet time about it!&nbsp; I'm pretty sure that a good part of this black art of deep nets is ensuring your filter layers learn well...it means the information is getting back to them during the weight updates.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "65611",
      "postDate": "03/06/2015 16:26:05",
      "content": "<p>How do you guys generate such visualizations? Is there some convenient too in python for such layer visualization ?</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "65613",
      "postDate": "03/06/2015 17:00:37",
      "content": "<p>[quote=BigFish;62791]</p>\n<p>Yes, you are right. it uses pretrained model, although it isn't permitted.</p>\n<p>The following is conv1 trained using provided data:</p>\n<p>[/quote]</p>\n<p>So the publicly available pre-trained models like VGG and GoogLeNet are not allowed?</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "65628",
      "postDate": "03/06/2015 19:54:00",
      "content": "<p>[quote=Chengyao Shen;65613]</p>\n<p>[quote=BigFish;62791]</p>\n<p>Yes, you are right. it uses pretrained model, although it isn't permitted.</p>\n<p>The following is conv1 trained using provided data:</p>\n<p>[/quote]</p>\n<p>So the publicly available pre-trained models like VGG and GoogLeNet are not allowed?</p>\n<p>[/quote]<br>Not allowed.&nbsp;https://www.kaggle.com/c/datasciencebowl/forums/t/11251/using-external-data/60493#post60493<br><br>And, btw, VGG models are released for research, not commercial purpose only. Only BVLC models are open, however they are not allowed anyway.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "65661",
      "postDate": "03/07/2015 03:38:09",
      "content": "<p>[quote=old-ufo;65628]</p>\n<p>[quote=Chengyao Shen;65613]</p>\n<p>[quote=BigFish;62791]</p>\n<p>Yes, you are right. it uses pretrained model, although it isn't permitted.</p>\n<p>The following is conv1 trained using provided data:</p>\n<p>[/quote]</p>\n<p>So the publicly available pre-trained models like VGG and GoogLeNet are not allowed?</p>\n<p>[/quote]<br>Not allowed.&nbsp;https://www.kaggle.com/c/datasciencebowl/forums/t/11251/using-external-data/60493#post60493<br><br>And, btw, VGG models are released for research, not commercial purpose only. Only BVLC models are open, however they are not allowed anyway.</p>\n<p>[/quote]</p>\n\n<p>Thanks for the info:)</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "65668",
      "postDate": "03/07/2015 08:10:01",
      "content": "<p>One of my CNN models have these filters (5x5x128).</p>",
      "rawMarkdown": "",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 62788,
      "author_name": "paweljankiewicz",
      "author_url": "",
      "post_date": "01/16/2015 12:01:28",
      "content": "<p>Why they are in colour? They seem strangely familiar but not this dataset.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 62790,
      "author_name": "fishpam",
      "author_url": "",
      "post_date": "01/16/2015 13:05:49",
      "content": "",
      "votes": null,
      "replies": []
    },
    {
      "id": 62791,
      "author_name": "fishpam",
      "author_url": "",
      "post_date": "01/16/2015 13:09:07",
      "content": "<p>Yes, you are right. it uses pretrained model, although it isn't permitted.</p>\n<p>The following is conv1 trained using provided data:</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 62808,
      "author_name": "miroslaw",
      "author_url": "",
      "post_date": "01/17/2015 00:13:51",
      "content": "<p>Are the colours supposed to represent pixel intensities?&nbsp;<br><br>I'm curious how those colours were generated considering the dataset is provided in grayscale.&nbsp;</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 62819,
      "author_name": "slobo777",
      "author_url": "",
      "post_date": "01/17/2015 08:39:46",
      "content": "<p>I'd guess if you took a CNN originally designed for&nbsp;a colour dataset, and trained it on a grey image expanded to RGB, then rendered as if it was RGB-based, then you might&nbsp;see three overlaid random greyscale filters in each cell. I think BigFish's second post looks like it could be that.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 62942,
      "author_name": "jaythodenvanvelzen",
      "author_url": "",
      "post_date": "01/20/2015 23:09:01",
      "content": "<p>Here's one of mine. All my filters are 3x3</p>\n<p>(obviously, epoch_065. The 000 one is after just one training epoch, and was accidentally added - but I can't seem to remove attachments)</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 62945,
      "author_name": "zerozero",
      "author_url": "",
      "post_date": "01/21/2015 01:10:55",
      "content": "<p>Here are a few of my filters from mid-training...nothing to write home about.&nbsp; They're learning but I wouldn't mind if they didn't take their sweet time about it!&nbsp; I'm pretty sure that a good part of this black art of deep nets is ensuring your filter layers learn well...it means the information is getting back to them during the weight updates.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 65611,
      "author_name": "pengpai",
      "author_url": "",
      "post_date": "03/06/2015 16:26:05",
      "content": "<p>How do you guys generate such visualizations? Is there some convenient too in python for such layer visualization ?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 65613,
      "author_name": "farawayao",
      "author_url": "",
      "post_date": "03/06/2015 17:00:37",
      "content": "<p>[quote=BigFish;62791]</p>\n<p>Yes, you are right. it uses pretrained model, although it isn't permitted.</p>\n<p>The following is conv1 trained using provided data:</p>\n<p>[/quote]</p>\n<p>So the publicly available pre-trained models like VGG and GoogLeNet are not allowed?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 65628,
      "author_name": "oldufo",
      "author_url": "",
      "post_date": "03/06/2015 19:54:00",
      "content": "<p>[quote=Chengyao Shen;65613]</p>\n<p>[quote=BigFish;62791]</p>\n<p>Yes, you are right. it uses pretrained model, although it isn't permitted.</p>\n<p>The following is conv1 trained using provided data:</p>\n<p>[/quote]</p>\n<p>So the publicly available pre-trained models like VGG and GoogLeNet are not allowed?</p>\n<p>[/quote]<br>Not allowed.&nbsp;https://www.kaggle.com/c/datasciencebowl/forums/t/11251/using-external-data/60493#post60493<br><br>And, btw, VGG models are released for research, not commercial purpose only. Only BVLC models are open, however they are not allowed anyway.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 65661,
      "author_name": "farawayao",
      "author_url": "",
      "post_date": "03/07/2015 03:38:09",
      "content": "<p>[quote=old-ufo;65628]</p>\n<p>[quote=Chengyao Shen;65613]</p>\n<p>[quote=BigFish;62791]</p>\n<p>Yes, you are right. it uses pretrained model, although it isn't permitted.</p>\n<p>The following is conv1 trained using provided data:</p>\n<p>[/quote]</p>\n<p>So the publicly available pre-trained models like VGG and GoogLeNet are not allowed?</p>\n<p>[/quote]<br>Not allowed.&nbsp;https://www.kaggle.com/c/datasciencebowl/forums/t/11251/using-external-data/60493#post60493<br><br>And, btw, VGG models are released for research, not commercial purpose only. Only BVLC models are open, however they are not allowed anyway.</p>\n<p>[/quote]</p>\n\n<p>Thanks for the info:)</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 65668,
      "author_name": "toshik",
      "author_url": "",
      "post_date": "03/07/2015 08:10:01",
      "content": "<p>One of my CNN models have these filters (5x5x128).</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "62786": "",
    "62788": "",
    "62790": "",
    "62791": "",
    "62808": "",
    "62819": "",
    "62942": "",
    "62945": "",
    "65611": "",
    "65613": "",
    "65628": "",
    "65661": "",
    "65668": ""
  },
  "source": "meta"
}