{
  "id": 25683,
  "title": "Training Accuracy stagnant at 50%",
  "url": "/competitions/dogs-vs-cats-redux-kernels-edition/discussion/25683",
  "author_name": "",
  "post_date": "2016-11-22T11:41:22.037Z",
  "votes": null,
  "comment_count": 9,
  "views": 425,
  "content": "<p>Hello everyone,</p>\n\n<p>I have created a VGG16 Net for this binary classification, but the accuracy stagnates at 50% and loss at ~0.69.</p>\n\n<p><a href=\"https://gist.github.com/anonymous/865df085c6bef6ba2d23e8038f76cba6\">https://gist.github.com/anonymous/865df085c6bef6ba2d23e8038f76cba6</a></p>\n\n<p>Can anybody tell me what's wrong with the network?</p>",
  "messages": [
    {
      "id": "146024",
      "postDate": "11/22/2016 11:41:22",
      "content": "<p>Hello everyone,</p>\n\n<p>I have created a VGG16 Net for this binary classification, but the accuracy stagnates at 50% and loss at ~0.69.</p>\n\n<p><a href=\"https://gist.github.com/anonymous/865df085c6bef6ba2d23e8038f76cba6\">https://gist.github.com/anonymous/865df085c6bef6ba2d23e8038f76cba6</a></p>\n\n<p>Can anybody tell me what's wrong with the network?</p>",
      "rawMarkdown": "Hello everyone,\r\n\r\nI have created a VGG16 Net for this binary classification, but the accuracy stagnates at 50% and loss at ~0.69.\r\n\r\nhttps://gist.github.com/anonymous/865df085c6bef6ba2d23e8038f76cba6\r\n\r\nCan anybody tell me what's wrong with the network?",
      "votes": null
    },
    {
      "id": "146077",
      "postDate": "11/22/2016 16:45:02",
      "content": "<p>I think this may be a fresh VGG16 net. You want a VGG16 net that includes its pre-trained weights.</p>",
      "rawMarkdown": "I think this may be a fresh VGG16 net. You want a VGG16 net that includes its pre-trained weights.",
      "votes": null
    },
    {
      "id": "147796",
      "postDate": "12/02/2016 04:20:32",
      "content": "<p>Not sure about tensorflow, but the loss function I used in keras was binary cross entropy. As suggested by  @Jeff finetuning pretrained VGG16 model is far better as compared to training from scratch.</p>",
      "rawMarkdown": "Not sure about tensorflow, but the loss function I used in keras was binary cross entropy. As suggested by  @Jeff finetuning pretrained VGG16 model is far better as compared to training from scratch.",
      "votes": null
    },
    {
      "id": "147804",
      "postDate": "12/02/2016 04:39:24",
      "content": "<p>Now that Anish is 23rd place, I'm pretty sure he somehow figured it out, Anjith ;-)</p>",
      "rawMarkdown": "Now that Anish is 23rd place, I'm pretty sure he somehow figured it out, Anjith ;-)",
      "votes": null
    },
    {
      "id": "147806",
      "postDate": "12/02/2016 04:44:40",
      "content": "<p>@Jeff din't see that :D :D</p>",
      "rawMarkdown": "Jeff din't see that :D :D",
      "votes": null
    },
    {
      "id": "147809",
      "postDate": "12/02/2016 04:55:50",
      "content": "<p>Sorry for the late reply! Thanks @Jeff. Using pre-trained network worked!\nBTW, I'm on 3rd place. :P</p>",
      "rawMarkdown": "Sorry for the late reply! Thanks @Jeff. Using pre-trained network worked!\r\nBTW, I'm on 3rd place. :P",
      "votes": null
    },
    {
      "id": "147810",
      "postDate": "12/02/2016 04:57:32",
      "content": "<p>Congrats, Anish. Did you try other network architectures? Inception? ResNet?</p>",
      "rawMarkdown": "Congrats, Anish. Did you try other network architectures? Inception? ResNet?",
      "votes": null
    },
    {
      "id": "147823",
      "postDate": "12/02/2016 05:30:18",
      "content": "<p>Yes! I used ResNet50 where I got 23rd rank. And then Inception which improved the rank to top 10</p>",
      "rawMarkdown": "Yes! I used ResNet50 where I got 23rd rank. And then Inception which improved the rank to top 10",
      "votes": null
    },
    {
      "id": "148018",
      "postDate": "12/02/2016 19:45:09",
      "content": "<p>That's awesome, Anish. How did were you able to tame Keras' ResNet50? I've tried lower and lower learning rates and waiting 32 epochs, but it doesn't seem to perform as quickly and reliably as VGG16.</p>",
      "rawMarkdown": "That's awesome, Anish. How did were you able to tame Keras' ResNet50? I've tried lower and lower learning rates and waiting 32 epochs, but it doesn't seem to perform as quickly and reliably as VGG16.",
      "votes": null
    },
    {
      "id": "148022",
      "postDate": "12/02/2016 20:02:48",
      "content": "<p>I used  a learning rate of 1e-4 and momentum 0.9... somewhat like in the paper..</p>",
      "rawMarkdown": "I used  a learning rate of 1e-4 and momentum 0.9... somewhat like in the paper..",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 146077,
      "author_name": "agent007",
      "author_url": "",
      "post_date": "11/22/2016 16:45:02",
      "content": "<p>I think this may be a fresh VGG16 net. You want a VGG16 net that includes its pre-trained weights.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 147796,
      "author_name": "anjithgeorge",
      "author_url": "",
      "post_date": "12/02/2016 04:20:32",
      "content": "<p>Not sure about tensorflow, but the loss function I used in keras was binary cross entropy. As suggested by  @Jeff finetuning pretrained VGG16 model is far better as compared to training from scratch.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 147804,
      "author_name": "agent007",
      "author_url": "",
      "post_date": "12/02/2016 04:39:24",
      "content": "<p>Now that Anish is 23rd place, I'm pretty sure he somehow figured it out, Anjith ;-)</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 147806,
      "author_name": "anjithgeorge",
      "author_url": "",
      "post_date": "12/02/2016 04:44:40",
      "content": "<p>@Jeff din't see that :D :D</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 147809,
      "author_name": "anishshah",
      "author_url": "",
      "post_date": "12/02/2016 04:55:50",
      "content": "<p>Sorry for the late reply! Thanks @Jeff. Using pre-trained network worked!\nBTW, I'm on 3rd place. :P</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 147810,
      "author_name": "agent007",
      "author_url": "",
      "post_date": "12/02/2016 04:57:32",
      "content": "<p>Congrats, Anish. Did you try other network architectures? Inception? ResNet?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 147823,
      "author_name": "anishshah",
      "author_url": "",
      "post_date": "12/02/2016 05:30:18",
      "content": "<p>Yes! I used ResNet50 where I got 23rd rank. And then Inception which improved the rank to top 10</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 148018,
      "author_name": "agent007",
      "author_url": "",
      "post_date": "12/02/2016 19:45:09",
      "content": "<p>That's awesome, Anish. How did were you able to tame Keras' ResNet50? I've tried lower and lower learning rates and waiting 32 epochs, but it doesn't seem to perform as quickly and reliably as VGG16.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 148022,
      "author_name": "anishshah",
      "author_url": "",
      "post_date": "12/02/2016 20:02:48",
      "content": "<p>I used  a learning rate of 1e-4 and momentum 0.9... somewhat like in the paper..</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "146024": "Hello everyone,\r\n\r\nI have created a VGG16 Net for this binary classification, but the accuracy stagnates at 50% and loss at ~0.69.\r\n\r\nhttps://gist.github.com/anonymous/865df085c6bef6ba2d23e8038f76cba6\r\n\r\nCan anybody tell me what's wrong with the network?",
    "146077": "I think this may be a fresh VGG16 net. You want a VGG16 net that includes its pre-trained weights.",
    "147796": "Not sure about tensorflow, but the loss function I used in keras was binary cross entropy. As suggested by  @Jeff finetuning pretrained VGG16 model is far better as compared to training from scratch.",
    "147804": "Now that Anish is 23rd place, I'm pretty sure he somehow figured it out, Anjith ;-)",
    "147806": "Jeff din't see that :D :D",
    "147809": "Sorry for the late reply! Thanks @Jeff. Using pre-trained network worked!\r\nBTW, I'm on 3rd place. :P",
    "147810": "Congrats, Anish. Did you try other network architectures? Inception? ResNet?",
    "147823": "Yes! I used ResNet50 where I got 23rd rank. And then Inception which improved the rank to top 10",
    "148018": "That's awesome, Anish. How did were you able to tame Keras' ResNet50? I've tried lower and lower learning rates and waiting 32 epochs, but it doesn't seem to perform as quickly and reliably as VGG16.",
    "148022": "I used  a learning rate of 1e-4 and momentum 0.9... somewhat like in the paper.."
  },
  "source": "meta"
}