Hello everyone,
\n\nSo as many of you know the deep learning tutorial provided by Kaggle uses Sunnybrook data and the label provided in that dataset is the contour of LV, not diastole or systole volume of LV. \nNow the training data provided by Kaggle has label of diastole and systole volume but not the contour. Are there anyways to continue training the network with Kaggle's training data, in a way similar to the script provided by Bing Xu in "End-to-End Deep Learning Tutorial (0.0392)"?
\n\nI currently have the model in a .caffemodel file, so I would probably stay with caffe instead of trying to switch to MXNet... unless there's an easy way to transfer learning to MXNet?
", "messages": [ { "id": "109436", "postDate": "02/26/2016 01:33:55", "content": "Hello everyone,
\n\nSo as many of you know the deep learning tutorial provided by Kaggle uses Sunnybrook data and the label provided in that dataset is the contour of LV, not diastole or systole volume of LV. \nNow the training data provided by Kaggle has label of diastole and systole volume but not the contour. Are there anyways to continue training the network with Kaggle's training data, in a way similar to the script provided by Bing Xu in "End-to-End Deep Learning Tutorial (0.0392)"?
\n\nI currently have the model in a .caffemodel file, so I would probably stay with caffe instead of trying to switch to MXNet... unless there's an easy way to transfer learning to MXNet?
", "rawMarkdown": "Hello everyone,\r\n\r\nSo as many of you know the deep learning tutorial provided by Kaggle uses Sunnybrook data and the label provided in that dataset is the contour of LV, not diastole or systole volume of LV. \r\nNow the training data provided by Kaggle has label of diastole and systole volume but not the contour. Are there anyways to continue training the network with Kaggle's training data, in a way similar to the script provided by Bing Xu in [\"End-to-End Deep Learning Tutorial (0.0392)\"][1]?\r\n\r\nI currently have the model in a .caffemodel file, so I would probably stay with caffe instead of trying to switch to MXNet... unless there's an easy way to transfer learning to MXNet?\r\n\r\n\r\n [1]: https://www.kaggle.com/c/second-annual-data-science-bowl/forums/t/18079/end-to-end-deep-learning-tutorial-0-0392", "votes": null }, { "id": "109463", "postDate": "02/26/2016 11:59:35", "content": "I'm using a different approach.
", "rawMarkdown": "I'm using a different approach.", "votes": null }, { "id": "109480", "postDate": "02/26/2016 16:29:01", "content": "So what is your approach??? Also does anyone else have any suggestions?
", "rawMarkdown": "So what is your approach??? Also does anyone else have any suggestions?", "votes": null }, { "id": "109485", "postDate": "02/26/2016 17:30:02", "content": "@solder, you are a little late to the party, but I would suggest you look at the Fourier tutorial instead of the deep learning tutorial if you want to get contour information out of thin air. Be advised there are many rabbit holes to jump into…choose wisely :)
", "rawMarkdown": "solder, you are a little late to the party, but I would suggest you look at the Fourier tutorial instead of the deep learning tutorial if you want to get contour information out of thin air. Be advised there are many rabbit holes to jump into…choose wisely :)", "votes": null }, { "id": "109486", "postDate": "02/26/2016 17:34:55", "content": "@DavidGbodiOdaibo\nWhat if I do not try to get contour information from the Kaggle data? Is it possible to continue training the neural net on the Kaggle dataset using diastole/systole volume data after it has been trained on the Sunnybrook data with contour data?
", "rawMarkdown": "DavidGbodiOdaibo\r\nWhat if I do not try to get contour information from the Kaggle data? Is it possible to continue training the neural net on the Kaggle dataset using diastole/systole volume data after it has been trained on the Sunnybrook data with contour data?", "votes": null }, { "id": "109489", "postDate": "02/26/2016 18:31:35", "content": "@solder The modality of the Kaggle training data labels (sys/dia) do not match that of the Sunnybrook dataset, so probably not. Trust me, earlier in the completion people tried the Sunnybrook approach and got horrible results. Save you the trouble, most of the people at the top of the leaderboard are using a variant of the Fourier tutorial combined with a convolutional neural net.
", "rawMarkdown": "solder The modality of the Kaggle training data labels (sys/dia) do not match that of the Sunnybrook dataset, so probably not. Trust me, earlier in the completion people tried the Sunnybrook approach and got horrible results. Save you the trouble, most of the people at the top of the leaderboard are using a variant of the Fourier tutorial combined with a convolutional neural net.", "votes": null }, { "id": "109490", "postDate": "02/26/2016 18:42:02", "content": "@DavidGbodiOdaibo\nThanks for your response! Just curious what kind of input/output format in the CNNs are people at the top most likely using? The deep learning tutorial takes in input slice by slice and outputs the predicted contour slice by slice as well. And then it combines slices to calculate volume.
\n\nIn Bing Xu's post with MXNet, he seems to be feeding in all slices at a single time point at once and outputting the CDF of volume directly. Which way is likely going to work better?
", "rawMarkdown": "DavidGbodiOdaibo\r\nThanks for your response! Just curious what kind of input/output format in the CNNs are people at the top most likely using? The deep learning tutorial takes in input slice by slice and outputs the predicted contour slice by slice as well. And then it combines slices to calculate volume. \r\n\r\nIn Bing Xu's post with MXNet, he seems to be feeding in all slices at a single time point at once and outputting the CDF of volume directly. Which way is likely going to work better?", "votes": null }, { "id": "109495", "postDate": "02/26/2016 19:17:13", "content": "@solder The people at the top are probably using the various things in the Fourier tutorial to better localize the heart reduce/condense the images ( e.t.c) there are many hidden gems in there (I think). I have come to that conclusion based on my own empirical evidence using (goolenet, ResNet, VGG, blah blah) and all the big bad powerful neural nets, that there is a ceiling, and my own observation that whenever a neural net guy joins forces with a Fourier segmentation guy they shoot to the top of the leaderboard in a few days sometimes hours. The key is in the preprocessing of the training images. So, their inputs are preprocessed images and their output are, Who Knows, could be the volumes if they are doing linear regression or Softmax probabilities if they are doing classification or the CDF directly.
\n\nBing Xu’s approach will work better than the deep learning approach.
", "rawMarkdown": "solder The people at the top are probably using the various things in the Fourier tutorial to better localize the heart reduce/condense the images ( e.t.c) there are many hidden gems in there (I think). I have come to that conclusion based on my own empirical evidence using (goolenet, ResNet, VGG, blah blah) and all the big bad powerful neural nets, that there is a ceiling, and my own observation that whenever a neural net guy joins forces with a Fourier segmentation guy they shoot to the top of the leaderboard in a few days sometimes hours. The key is in the preprocessing of the training images. So, their inputs are preprocessed images and their output are, Who Knows, could be the volumes if they are doing linear regression or Softmax probabilities if they are doing classification or the CDF directly.\r\n\r\nBing Xu’s approach will work better than the deep learning approach.", "votes": null }, { "id": "109568", "postDate": "02/27/2016 21:26:29", "content": "@I Love Solder Soon :3
", "rawMarkdown": "I Love Solder Soon :3", "votes": null }, { "id": "109575", "postDate": "02/27/2016 23:34:43", "content": "[quote=DavidGbodiOdaibo;109495]
\n\n@solder The people at the top are probably using the various things in the Fourier tutorial to better localize the heart reduce/condense the images ( e.t.c) there are many hidden gems in there (I think). I have come to that conclusion based on my own empirical evidence using (goolenet, ResNet, VGG, blah blah) and all the big bad powerful neural nets, that there is a ceiling, and my own observation that whenever a neural net guy joins forces with a Fourier segmentation guy they shoot to the top of the leaderboard in a few days sometimes hours. The key is in the preprocessing of the training images. So, their inputs are preprocessed images and their output are, Who Knows, could be the volumes if they are doing linear regression or Softmax probabilities if they are doing classification or the CDF directly.
\n\nBing Xu’s approach will work better than the deep learning approach.
\n\n[/quote]
\n\nThanks for your response! I will read into the MXnet and Fourier tutorial more see if I can get some inspirations from those scripts!
", "rawMarkdown": "[quote=DavidGbodiOdaibo;109495]\r\n\r\n@solder The people at the top are probably using the various things in the Fourier tutorial to better localize the heart reduce/condense the images ( e.t.c) there are many hidden gems in there (I think). I have come to that conclusion based on my own empirical evidence using (goolenet, ResNet, VGG, blah blah) and all the big bad powerful neural nets, that there is a ceiling, and my own observation that whenever a neural net guy joins forces with a Fourier segmentation guy they shoot to the top of the leaderboard in a few days sometimes hours. The key is in the preprocessing of the training images. So, their inputs are preprocessed images and their output are, Who Knows, could be the volumes if they are doing linear regression or Softmax probabilities if they are doing classification or the CDF directly.\r\n\r\nBing Xu’s approach will work better than the deep learning approach.\r\n\r\n\r\n[/quote]\r\n\r\nThanks for your response! I will read into the MXnet and Fourier tutorial more see if I can get some inspirations from those scripts!", "votes": null } ], "comments": [ { "id": 109463, "author_name": "alvaroosvaldo", "author_url": "", "post_date": "02/26/2016 11:59:35", "content": "I'm using a different approach.
", "votes": null, "replies": [] }, { "id": 109480, "author_name": "ilovesolder", "author_url": "", "post_date": "02/26/2016 16:29:01", "content": "So what is your approach??? Also does anyone else have any suggestions?
", "votes": null, "replies": [] }, { "id": 109485, "author_name": "godaibo", "author_url": "", "post_date": "02/26/2016 17:30:02", "content": "@solder, you are a little late to the party, but I would suggest you look at the Fourier tutorial instead of the deep learning tutorial if you want to get contour information out of thin air. Be advised there are many rabbit holes to jump into…choose wisely :)
", "votes": null, "replies": [] }, { "id": 109486, "author_name": "ilovesolder", "author_url": "", "post_date": "02/26/2016 17:34:55", "content": "@DavidGbodiOdaibo\nWhat if I do not try to get contour information from the Kaggle data? Is it possible to continue training the neural net on the Kaggle dataset using diastole/systole volume data after it has been trained on the Sunnybrook data with contour data?
", "votes": null, "replies": [] }, { "id": 109489, "author_name": "godaibo", "author_url": "", "post_date": "02/26/2016 18:31:35", "content": "@solder The modality of the Kaggle training data labels (sys/dia) do not match that of the Sunnybrook dataset, so probably not. Trust me, earlier in the completion people tried the Sunnybrook approach and got horrible results. Save you the trouble, most of the people at the top of the leaderboard are using a variant of the Fourier tutorial combined with a convolutional neural net.
", "votes": null, "replies": [] }, { "id": 109490, "author_name": "ilovesolder", "author_url": "", "post_date": "02/26/2016 18:42:02", "content": "@DavidGbodiOdaibo\nThanks for your response! Just curious what kind of input/output format in the CNNs are people at the top most likely using? The deep learning tutorial takes in input slice by slice and outputs the predicted contour slice by slice as well. And then it combines slices to calculate volume.
\n\nIn Bing Xu's post with MXNet, he seems to be feeding in all slices at a single time point at once and outputting the CDF of volume directly. Which way is likely going to work better?
", "votes": null, "replies": [] }, { "id": 109495, "author_name": "godaibo", "author_url": "", "post_date": "02/26/2016 19:17:13", "content": "@solder The people at the top are probably using the various things in the Fourier tutorial to better localize the heart reduce/condense the images ( e.t.c) there are many hidden gems in there (I think). I have come to that conclusion based on my own empirical evidence using (goolenet, ResNet, VGG, blah blah) and all the big bad powerful neural nets, that there is a ceiling, and my own observation that whenever a neural net guy joins forces with a Fourier segmentation guy they shoot to the top of the leaderboard in a few days sometimes hours. The key is in the preprocessing of the training images. So, their inputs are preprocessed images and their output are, Who Knows, could be the volumes if they are doing linear regression or Softmax probabilities if they are doing classification or the CDF directly.
\n\nBing Xu’s approach will work better than the deep learning approach.
", "votes": null, "replies": [] }, { "id": 109568, "author_name": "alvaroosvaldo", "author_url": "", "post_date": "02/27/2016 21:26:29", "content": "@I Love Solder Soon :3
", "votes": null, "replies": [] }, { "id": 109575, "author_name": "ilovesolder", "author_url": "", "post_date": "02/27/2016 23:34:43", "content": "[quote=DavidGbodiOdaibo;109495]
\n\n@solder The people at the top are probably using the various things in the Fourier tutorial to better localize the heart reduce/condense the images ( e.t.c) there are many hidden gems in there (I think). I have come to that conclusion based on my own empirical evidence using (goolenet, ResNet, VGG, blah blah) and all the big bad powerful neural nets, that there is a ceiling, and my own observation that whenever a neural net guy joins forces with a Fourier segmentation guy they shoot to the top of the leaderboard in a few days sometimes hours. The key is in the preprocessing of the training images. So, their inputs are preprocessed images and their output are, Who Knows, could be the volumes if they are doing linear regression or Softmax probabilities if they are doing classification or the CDF directly.
\n\nBing Xu’s approach will work better than the deep learning approach.
\n\n[/quote]
\n\nThanks for your response! I will read into the MXnet and Fourier tutorial more see if I can get some inspirations from those scripts!
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "109436": "Hello everyone,\r\n\r\nSo as many of you know the deep learning tutorial provided by Kaggle uses Sunnybrook data and the label provided in that dataset is the contour of LV, not diastole or systole volume of LV. \r\nNow the training data provided by Kaggle has label of diastole and systole volume but not the contour. Are there anyways to continue training the network with Kaggle's training data, in a way similar to the script provided by Bing Xu in [\"End-to-End Deep Learning Tutorial (0.0392)\"][1]?\r\n\r\nI currently have the model in a .caffemodel file, so I would probably stay with caffe instead of trying to switch to MXNet... unless there's an easy way to transfer learning to MXNet?\r\n\r\n\r\n [1]: https://www.kaggle.com/c/second-annual-data-science-bowl/forums/t/18079/end-to-end-deep-learning-tutorial-0-0392", "109463": "I'm using a different approach.", "109480": "So what is your approach??? Also does anyone else have any suggestions?", "109485": "solder, you are a little late to the party, but I would suggest you look at the Fourier tutorial instead of the deep learning tutorial if you want to get contour information out of thin air. Be advised there are many rabbit holes to jump into…choose wisely :)", "109486": "DavidGbodiOdaibo\r\nWhat if I do not try to get contour information from the Kaggle data? Is it possible to continue training the neural net on the Kaggle dataset using diastole/systole volume data after it has been trained on the Sunnybrook data with contour data?", "109489": "solder The modality of the Kaggle training data labels (sys/dia) do not match that of the Sunnybrook dataset, so probably not. Trust me, earlier in the completion people tried the Sunnybrook approach and got horrible results. Save you the trouble, most of the people at the top of the leaderboard are using a variant of the Fourier tutorial combined with a convolutional neural net.", "109490": "DavidGbodiOdaibo\r\nThanks for your response! Just curious what kind of input/output format in the CNNs are people at the top most likely using? The deep learning tutorial takes in input slice by slice and outputs the predicted contour slice by slice as well. And then it combines slices to calculate volume. \r\n\r\nIn Bing Xu's post with MXNet, he seems to be feeding in all slices at a single time point at once and outputting the CDF of volume directly. Which way is likely going to work better?", "109495": "solder The people at the top are probably using the various things in the Fourier tutorial to better localize the heart reduce/condense the images ( e.t.c) there are many hidden gems in there (I think). I have come to that conclusion based on my own empirical evidence using (goolenet, ResNet, VGG, blah blah) and all the big bad powerful neural nets, that there is a ceiling, and my own observation that whenever a neural net guy joins forces with a Fourier segmentation guy they shoot to the top of the leaderboard in a few days sometimes hours. The key is in the preprocessing of the training images. So, their inputs are preprocessed images and their output are, Who Knows, could be the volumes if they are doing linear regression or Softmax probabilities if they are doing classification or the CDF directly.\r\n\r\nBing Xu’s approach will work better than the deep learning approach.", "109568": "I Love Solder Soon :3", "109575": "[quote=DavidGbodiOdaibo;109495]\r\n\r\n@solder The people at the top are probably using the various things in the Fourier tutorial to better localize the heart reduce/condense the images ( e.t.c) there are many hidden gems in there (I think). I have come to that conclusion based on my own empirical evidence using (goolenet, ResNet, VGG, blah blah) and all the big bad powerful neural nets, that there is a ceiling, and my own observation that whenever a neural net guy joins forces with a Fourier segmentation guy they shoot to the top of the leaderboard in a few days sometimes hours. The key is in the preprocessing of the training images. So, their inputs are preprocessed images and their output are, Who Knows, could be the volumes if they are doing linear regression or Softmax probabilities if they are doing classification or the CDF directly.\r\n\r\nBing Xu’s approach will work better than the deep learning approach.\r\n\r\n\r\n[/quote]\r\n\r\nThanks for your response! I will read into the MXnet and Fourier tutorial more see if I can get some inspirations from those scripts!" }, "source": "meta" }