{
  "id": 53663,
  "title": "Typical MaskRCNN Loss Function Values for this problem?",
  "url": "/competitions/data-science-bowl-2018/discussion/53663",
  "author_name": "",
  "post_date": "2018-04-03T07:58:11.336476600Z",
  "votes": 2,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Could anyone please tell what are the typical loss function values when trained on MaskRCNN? In particular values of \n1. train loss\n2. rpn_class_loss\n3. rpn_bbox_loss\n4. mrcnn_class_loss\n5. mrcnn_bbox_loss\n6. mrcnn_mask_loss\n7. val loss</p>",
  "messages": [
    {
      "id": "308248",
      "postDate": "04/03/2018 07:58:11",
      "content": "<p>Could anyone please tell what are the typical loss function values when trained on MaskRCNN? In particular values of \n1. train loss\n2. rpn_class_loss\n3. rpn_bbox_loss\n4. mrcnn_class_loss\n5. mrcnn_bbox_loss\n6. mrcnn_mask_loss\n7. val loss</p>",
      "rawMarkdown": "Could anyone please tell what are the typical loss function values when trained on MaskRCNN? In particular values of \n1. train loss\n2. rpn_class_loss\n3. rpn_bbox_loss\n4. mrcnn_class_loss\n5. mrcnn_bbox_loss\n6. mrcnn_mask_loss\n7. val loss",
      "votes": null
    },
    {
      "id": "309359",
      "postDate": "04/05/2018 05:56:59",
      "content": "<p>Hi Puneet, for me it varies a lot for each experiment. here are my values for one experiment after 14 epochs.\n- loss: 0.8560 - rpn_class_loss: 0.0338 - rpn_bbox_loss: 0.2324 - mrcnn_class_loss: 0.1592 - mrcnn_bbox_loss: 0.1643 - mrcnn_mask_loss: 0.2663 - val_loss: 0.8723 - val_rpn_class_loss: 0.0544 - val_rpn_bbox_loss: 0.2183 - val_mrcnn_class_loss: 0.2033 - val_mrcnn_bbox_loss: 0.1392 - val_mrcnn_mask_loss: 0.2571</p>\n\n<p>I only rely on the val_* numbers ; in my opinion the other training numbers are less useful. we have a large model and small dataset; the model could easily be over-fitting the data. </p>\n\n<p>One question: after how many epochs did you get your result and what was your training schedule and Learning rate.\nThanks\nMehul</p>",
      "rawMarkdown": "Hi Puneet, for me it varies a lot for each experiment. here are my values for one experiment after 14 epochs.\n- loss: 0.8560 - rpn_class_loss: 0.0338 - rpn_bbox_loss: 0.2324 - mrcnn_class_loss: 0.1592 - mrcnn_bbox_loss: 0.1643 - mrcnn_mask_loss: 0.2663 - val_loss: 0.8723 - val_rpn_class_loss: 0.0544 - val_rpn_bbox_loss: 0.2183 - val_mrcnn_class_loss: 0.2033 - val_mrcnn_bbox_loss: 0.1392 - val_mrcnn_mask_loss: 0.2571\n\nI only rely on the val_* numbers ; in my opinion the other training numbers are less useful. we have a large model and small dataset; the model could easily be over-fitting the data. \n\nOne question: after how many epochs did you get your result and what was your training schedule and Learning rate.\nThanks\nMehul",
      "votes": null
    },
    {
      "id": "309376",
      "postDate": "04/05/2018 07:21:27",
      "content": "<p>Hi Mehul,</p>\n\n<p>For me below is the output after 9 epochs:</p>\n\n<p>Epoch 9/15\n 54/600 [=&gt;............................] - ETA: 31:20 - loss: 0.5847 - rpn_class_loss: 0.0225 - rpn_bbox_loss: 0.1530 - mrcnn_class_loss: 0.1012 - mrcnn_bbox_loss: 0.1072 - mrcnn_mask_loss: 0.2008</p>\n\n<p>I am not using any post processing for now? Is this loss good to go or I need to train for more epochs?</p>",
      "rawMarkdown": "Hi Mehul,\n\nFor me below is the output after 9 epochs:\n\nEpoch 9/15\n 54/600 [=&gt;............................] - ETA: 31:20 - loss: 0.5847 - rpn_class_loss: 0.0225 - rpn_bbox_loss: 0.1530 - mrcnn_class_loss: 0.1012 - mrcnn_bbox_loss: 0.1072 - mrcnn_mask_loss: 0.2008\n\nI am not using any post processing for now? Is this loss good to go or I need to train for more epochs?",
      "votes": null
    },
    {
      "id": "309386",
      "postDate": "04/05/2018 07:42:22",
      "content": "<p>For only 9 epochs that is good! You can try 20 epochs more and see if falls further. The main issue is to avoid over training. Did you split into train and validaton sets too ?  I als used image augmentation like rotation but that did not help me. It helped some others. </p>\n\n<p>Did you do any preprocessing on the data ? Like scaling ? </p>",
      "rawMarkdown": "For only 9 epochs that is good! You can try 20 epochs more and see if falls further. The main issue is to avoid over training. Did you split into train and validaton sets too ?  I als used image augmentation like rotation but that did not help me. It helped some others. \n\nDid you do any preprocessing on the data ? Like scaling ?",
      "votes": null
    },
    {
      "id": "309616",
      "postDate": "04/05/2018 17:25:01",
      "content": "<p>yes I have kept validation set based on vgg clustering. Also as pre-processing I have done only scaling.</p>",
      "rawMarkdown": "yes I have kept validation set based on vgg clustering. Also as pre-processing I have done only scaling.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 309359,
      "author_name": "mpsampat",
      "author_url": "",
      "post_date": "04/05/2018 05:56:59",
      "content": "<p>Hi Puneet, for me it varies a lot for each experiment. here are my values for one experiment after 14 epochs.\n- loss: 0.8560 - rpn_class_loss: 0.0338 - rpn_bbox_loss: 0.2324 - mrcnn_class_loss: 0.1592 - mrcnn_bbox_loss: 0.1643 - mrcnn_mask_loss: 0.2663 - val_loss: 0.8723 - val_rpn_class_loss: 0.0544 - val_rpn_bbox_loss: 0.2183 - val_mrcnn_class_loss: 0.2033 - val_mrcnn_bbox_loss: 0.1392 - val_mrcnn_mask_loss: 0.2571</p>\n\n<p>I only rely on the val_* numbers ; in my opinion the other training numbers are less useful. we have a large model and small dataset; the model could easily be over-fitting the data. </p>\n\n<p>One question: after how many epochs did you get your result and what was your training schedule and Learning rate.\nThanks\nMehul</p>",
      "votes": null,
      "replies": [
        {
          "id": 309376,
          "author_name": "puneetvashisht786",
          "author_url": "",
          "post_date": "04/05/2018 07:21:27",
          "content": "<p>Hi Mehul,</p>\n\n<p>For me below is the output after 9 epochs:</p>\n\n<p>Epoch 9/15\n 54/600 [=&gt;............................] - ETA: 31:20 - loss: 0.5847 - rpn_class_loss: 0.0225 - rpn_bbox_loss: 0.1530 - mrcnn_class_loss: 0.1012 - mrcnn_bbox_loss: 0.1072 - mrcnn_mask_loss: 0.2008</p>\n\n<p>I am not using any post processing for now? Is this loss good to go or I need to train for more epochs?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 309386,
          "author_name": "mpsampat",
          "author_url": "",
          "post_date": "04/05/2018 07:42:22",
          "content": "<p>For only 9 epochs that is good! You can try 20 epochs more and see if falls further. The main issue is to avoid over training. Did you split into train and validaton sets too ?  I als used image augmentation like rotation but that did not help me. It helped some others. </p>\n\n<p>Did you do any preprocessing on the data ? Like scaling ? </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 309616,
          "author_name": "puneetvashisht786",
          "author_url": "",
          "post_date": "04/05/2018 17:25:01",
          "content": "<p>yes I have kept validation set based on vgg clustering. Also as pre-processing I have done only scaling.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "308248": "Could anyone please tell what are the typical loss function values when trained on MaskRCNN? In particular values of \n1. train loss\n2. rpn_class_loss\n3. rpn_bbox_loss\n4. mrcnn_class_loss\n5. mrcnn_bbox_loss\n6. mrcnn_mask_loss\n7. val loss",
    "309359": "Hi Puneet, for me it varies a lot for each experiment. here are my values for one experiment after 14 epochs.\n- loss: 0.8560 - rpn_class_loss: 0.0338 - rpn_bbox_loss: 0.2324 - mrcnn_class_loss: 0.1592 - mrcnn_bbox_loss: 0.1643 - mrcnn_mask_loss: 0.2663 - val_loss: 0.8723 - val_rpn_class_loss: 0.0544 - val_rpn_bbox_loss: 0.2183 - val_mrcnn_class_loss: 0.2033 - val_mrcnn_bbox_loss: 0.1392 - val_mrcnn_mask_loss: 0.2571\n\nI only rely on the val_* numbers ; in my opinion the other training numbers are less useful. we have a large model and small dataset; the model could easily be over-fitting the data. \n\nOne question: after how many epochs did you get your result and what was your training schedule and Learning rate.\nThanks\nMehul",
    "309376": "Hi Mehul,\n\nFor me below is the output after 9 epochs:\n\nEpoch 9/15\n 54/600 [=&gt;............................] - ETA: 31:20 - loss: 0.5847 - rpn_class_loss: 0.0225 - rpn_bbox_loss: 0.1530 - mrcnn_class_loss: 0.1012 - mrcnn_bbox_loss: 0.1072 - mrcnn_mask_loss: 0.2008\n\nI am not using any post processing for now? Is this loss good to go or I need to train for more epochs?",
    "309386": "For only 9 epochs that is good! You can try 20 epochs more and see if falls further. The main issue is to avoid over training. Did you split into train and validaton sets too ?  I als used image augmentation like rotation but that did not help me. It helped some others. \n\nDid you do any preprocessing on the data ? Like scaling ?",
    "309616": "yes I have kept validation set based on vgg clustering. Also as pre-processing I have done only scaling."
  },
  "source": "meta"
}