{
  "id": 71759,
  "title": "Help！！！Some problem in tf",
  "url": "/competitions/human-protein-atlas-image-classification/discussion/71759",
  "author_name": "",
  "post_date": "2018-11-16T08:38:58.217868800Z",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I use resnet50 to train the datasets,but i find  that when let the last fc layer through the sigmoid function ,the output of some images are the same.Why？I set batch size 64 ,and make the train data every 64 a group then feed it in my model every batch size.loss decreases slowly .when i use the model to predict the test data,the outputs of some pictures are same too.I simplily use sigmoid with cross_entropy as the loss function.I have checked that data process is right.I have  no idea about how to solve the prblem,if anyone know how to solve it ,please help me .Thanks</p>",
  "messages": [
    {
      "id": "422453",
      "postDate": "11/16/2018 08:38:58",
      "content": "<p>I use resnet50 to train the datasets,but i find  that when let the last fc layer through the sigmoid function ,the output of some images are the same.Why？I set batch size 64 ,and make the train data every 64 a group then feed it in my model every batch size.loss decreases slowly .when i use the model to predict the test data,the outputs of some pictures are same too.I simplily use sigmoid with cross_entropy as the loss function.I have checked that data process is right.I have  no idea about how to solve the prblem,if anyone know how to solve it ,please help me .Thanks</p>",
      "rawMarkdown": "I use resnet50 to train the datasets,but i find  that when let the last fc layer through the sigmoid function ,the output of some images are the same.Why？I set batch size 64 ,and make the train data every 64 a group then feed it in my model every batch size.loss decreases slowly .when i use the model to predict the test data,the outputs of some pictures are same too.I simplily use sigmoid with cross_entropy as the loss function.I have checked that data process is right.I have  no idea about how to solve the prblem,if anyone know how to solve it ,please help me .Thanks",
      "votes": null
    },
    {
      "id": "422551",
      "postDate": "11/16/2018 11:51:57",
      "content": "<p>It has been reported that there are some duplicate images, \nsee here:\n<a href=\"https://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/69979\">https://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/69979</a></p>\n\n<p>you can generate the figures of the same output to see if this is the case.</p>",
      "rawMarkdown": "It has been reported that there are some duplicate images, \nsee here:\nhttps://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/69979\n\nyou can generate the figures of the same output to see if this is the case.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 422551,
      "author_name": "zhijianli",
      "author_url": "",
      "post_date": "11/16/2018 11:51:57",
      "content": "<p>It has been reported that there are some duplicate images, \nsee here:\n<a href=\"https://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/69979\">https://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/69979</a></p>\n\n<p>you can generate the figures of the same output to see if this is the case.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "422453": "I use resnet50 to train the datasets,but i find  that when let the last fc layer through the sigmoid function ,the output of some images are the same.Why？I set batch size 64 ,and make the train data every 64 a group then feed it in my model every batch size.loss decreases slowly .when i use the model to predict the test data,the outputs of some pictures are same too.I simplily use sigmoid with cross_entropy as the loss function.I have checked that data process is right.I have  no idea about how to solve the prblem,if anyone know how to solve it ,please help me .Thanks",
    "422551": "It has been reported that there are some duplicate images, \nsee here:\nhttps://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/69979\n\nyou can generate the figures of the same output to see if this is the case."
  },
  "source": "meta"
}