Please post here for use of pre-trained models and external data.
\n\nThis competition will allow limited use of external data, such as pre-trained nets. \nYour data/model should be freely available to use by anyone in the community.
", "messages": [ { "id": 119194, "postDate": "2016-05-07T22:27:52.567Z", "content": "[quote=Wendy Kan;119178]
\n\nThis is a bit of a gray area here. The rule of thumb is "if it's free to use by anyone (including State Farm), it's allowed". If you want to use it, please contact the pre-trained model author and confirm that it can be used.
\n\nEDIT: Thanks to @Ctrl+W who found a source that has "unrestricted use" license. Please stick to this version (or others that you can find free license) if you're interested in VGG-16.
\n\n[quote=inversion;119172]
\n\nSo, now this is interesting.
\n\nDoes the VGG-16 licence issue apply if it's been converted to another platform? (e.g., Keras)
\n\n[/quote]
\n\n[/quote]
\n\nThe "no restrictions" license may not be a legally valid here, and in the case of the keras gist, there is no usage license quoted at all. There is no evidence that the IP has been handled correctly in this case, and the publisher in keras-model-zoo has given credit to the gist but changed the licensing terms, which is not allowed unless they have permission from the author.
\n\nMost importantly, from the sponsor's point of view, they would be liable even if they used the source material in good faith based on the stated license found in keras-model-zoo, if the originating VGG team decided to enforce their proper licensing terms.
\n\nSometimes it is not so much a grey area, but rather individuals publishing code or data - often in good faith trying to make something useful - without understanding the fiddly details that affect software licensing.
\n\nIf the authors of the relevant code are here on Kaggle, it would be great to get clarification.
\n\nI feel a little responsible here, since I posted the VGG/Keras gist link on this thread, and know enough about licensing to spot this. I was being a bit naive, as I haven't encountered very many IP problems with machine learning libraries in the past.
", "rawMarkdown": "[quote=Wendy Kan;119178]\r\n\r\nThis is a bit of a gray area here. The rule of thumb is \"if it's free to use by anyone (including State Farm), it's allowed\". If you want to use it, please contact the pre-trained model author and confirm that it can be used. \r\n\r\nEDIT: Thanks to @Ctrl+W who found a [source][1] that has \"unrestricted use\" license. Please stick to this version (or others that you can find free license) if you're interested in VGG-16. \r\n\r\n[quote=inversion;119172]\r\n\r\nSo, now this is interesting. \r\n\r\nDoes the VGG-16 licence issue apply if it's been [converted to another platform][2]? (e.g., Keras)\r\n\r\n\r\n[/quote]\r\n\r\n\r\n [1]: https://github.com/albertomontesg/keras-model-zoo/tree/master/models/VGG-16\r\n [2]: https://gist.github.com/baraldilorenzo/07d7802847aaad0a35d3\r\n\r\n[/quote]\r\n\r\nThe \"no restrictions\" license may not be a legally valid here, and in the case of the keras gist, there is no usage license quoted at all. There is no evidence that the IP has been handled correctly in this case, and the publisher in keras-model-zoo has given credit to the gist but changed the licensing terms, which is not allowed unless they have permission from the author. \r\n\r\nMost importantly, from the sponsor's point of view, they would be liable even if they used the source material in good faith based on the stated license found in keras-model-zoo, if the originating VGG team decided to enforce their proper licensing terms.\r\n\r\nSometimes it is not so much a grey area, but rather individuals publishing code or data - often in good faith trying to make something useful - without understanding the fiddly details that affect software licensing.\r\n\r\nIf the authors of the relevant code are here on Kaggle, it would be great to get clarification.\r\n\r\nI feel a little responsible here, since I posted the VGG/Keras gist link on this thread, and know enough about licensing to spot this. I was being a bit naive, as I haven't encountered very many IP problems with machine learning libraries in the past.\r\n\r\n", "votes": 5 }, { "id": 114941, "postDate": "2016-04-14T22:09:31.907Z", "content": "Please post here for use of pre-trained models and external data.
\n\nThis competition will allow limited use of external data, such as pre-trained nets. \nYour data/model should be freely available to use by anyone in the community.
", "rawMarkdown": "Please post here for use of pre-trained models and external data. \r\n\r\nThis competition will allow limited use of external data, such as pre-trained nets. \r\nYour data/model should be freely available to use by anyone in the community. ", "votes": 5 }, { "id": 117476, "postDate": "2016-04-29T09:04:52.967Z", "content": "[quote=V.AbhijayArora;117469]
\n\n[quote=Luis Andre Dutra e Silva;116906]
\n\nPre-trained models can be compared to software libraries. We could call them "data libraries" as they provide, in the case of imagenet best models, a good start in terms of semantic features already embedded in their binary files.
\n\n[/quote]
\n\nI may be wrong, but isn't it counter-intuitive? Loading weights trained on other images for your dataset, how will that help? Aren't the weights supposed to be specific to a dataset i.e driver images in this case?
\n\n[/quote]
\n\nLow level features such as edges, corners, dots and repeating textures, are common amongst many different types of image. In fact before really deep CNNs took off, there were feature extractors such as Sobel filters, which effectively pre-encoded the same kinds of basic image components.
\n\nThe idea is that a high-performing CNN trained very large number of images (more than we have, or have time for in this competition) will have learned pretty good low-level features, most of which apply to our competition too. Higher-level features less so, but then thats what the fine-tuning is for.
\n\nIt helps if the content is vaguely similar - something trained on people, perhaps a pose classifier, might require less re-training than something classifying breeds of pet.
", "rawMarkdown": "[quote=V.AbhijayArora;117469]\r\n\r\n[quote=Luis Andre Dutra e Silva;116906]\r\n\r\nPre-trained models can be compared to software libraries. We could call them \"data libraries\" as they provide, in the case of imagenet best models, a good start in terms of semantic features already embedded in their binary files. \r\n\r\n[/quote]\r\n\r\nI may be wrong, but isn't it counter-intuitive? Loading weights trained on other images for your dataset, how will that help? Aren't the weights supposed to be specific to a dataset i.e driver images in this case?\r\n\r\n[/quote]\r\n\r\nLow level features such as edges, corners, dots and repeating textures, are common amongst many different types of image. In fact before really deep CNNs took off, there were feature extractors such as Sobel filters, which effectively pre-encoded the same kinds of basic image components.\r\n\r\nThe idea is that a high-performing CNN trained very large number of images (more than we have, or have time for in this competition) will have learned pretty good low-level features, most of which apply to our competition too. Higher-level features less so, but then thats what the fine-tuning is for.\r\n\r\nIt helps if the content is vaguely similar - something trained on people, perhaps a pose classifier, might require less re-training than something classifying breeds of pet.\r\n", "votes": 3 }, { "id": 122668, "postDate": "2016-06-06T10:17:32.877Z", "content": "Situation has changed.
\nThe description of "non-commercial use only" for VGG16 and VGG19 were removed from the caffe zoo.
\nIt is clear that VGG models are provided here by "CC BY 4.0" (commercial use is allowed).
\nPlease rethink about using VGG16 and VGG19.
I am going to use a hand grasping sequence from http://www.hci.iis.u-tokyo.ac.jp/~cai-mj/utgrasp_dataset.html
", "rawMarkdown": "I am going to use a hand grasping sequence from [http://www.hci.iis.u-tokyo.ac.jp/~cai-mj/utgrasp_dataset.html][1]\r\n\r\n\r\n [1]: http://www.hci.iis.u-tokyo.ac.jp/~cai-mj/utgrasp_dataset.html", "votes": 1 }, { "id": 120952, "postDate": "2016-05-22T07:07:58.597Z", "content": "In VGG web site, VGG model is provided by "CC BY 4.0" (commercial use is allowed).
\nIn the caffe zoo, however, VGG_ILSVRC_16_layers is provided by "CC BY-NC 4.0" (non-commercial use only).
I am confused about license of VGG. Is using VGG model allowed ? not allowed ?
\nIs there any way to use VGG model in this competiton ?
Can we use imagenet data in this ? imagenet
", "rawMarkdown": "Can we use imagenet data in this ? [imagenet][1]\r\n\r\n\r\n [1]: http://www.image-net.org/", "votes": 1 }, { "id": 120412, "postDate": "2016-05-18T01:52:08.310Z", "content": "I'm planning on experimenting with the following datasets:
\n\n", "rawMarkdown": "I'm planning on experimenting with the following datasets:\r\n\r\n* [buffy dataset][1]\r\n* [leeds and extended leeds sports][2]\r\n* [FLIC dataset][3]\r\n* [Head pose image dataset][4]\r\n* [MS COCO][5]\r\n\r\n [1]: http://www.robots.ox.ac.uk/~vgg/data/stickmen/\r\n [2]: http://www.comp.leeds.ac.uk/mat4saj/lsp.html\r\n [3]: http://bensapp.github.io/flic-dataset.html\r\n [4]: http://www-prima.inrialpes.fr/perso/Gourier/Faces/HPDatabase.html\r\n [5]: http://mscoco.org/dataset/", "votes": 1 }, { "id": 119272, "postDate": "2016-05-08T18:27:28.083Z", "content": "[quote=threecourse;119243]
\n\nIf so, because caffe VGG model weights are same as weights in VGG site,
\nmy idea is that keras VGG model, whose weights are obtained by converting, is okay.
\nother ideas?
[quote=Neil Slater;119224]
\n\n[quote=threecourse;119196]
\n\nIt says "The models are released under Creative Commons Attribution License." in http://www.robots.ox.ac.uk/~vgg/research/very_deep/
\n\nIs it allowed to convert weights obtained from there?
\n\n[/quote]
\n\nYes, I think that is OK . That license is https://creativecommons.org/licenses/by/4.0/ which allows commercial use, as long as credit is given. It is the caffe model which adds the "no commercial" restriction, which the keras gist then used (and thus inherited the restriction).
\n\nThe keras-model-zoo version appears to be incorrectly licensed, unless separate agreement reached with VGG team. That is not clear.
\n\n[/quote]
\n\n[/quote]
\n\nNo, that is not true, the licensing as written is not good for that. There is work involved in the conversion, and the non-commercial only restriction applies to that work due to how the software has been published, even though the original source doesn't have the restriction. In general it is not safe to use the kind of logical thinking you want to apply, you have to go with what is written for the license.
\n\nEven if what is written doesn't make much sense, that just makes the product unsafe to use commercially. So you might get away with it in your own work - I doubt the VGG team or anyone in the Keras team really wants to sue anyone - but that is not the same as the thing being good to claim a prize in this competition.
\n\nThe simplest thing to do is make the conversion form caffe model to Keras yourself, or hope that someone else does so under a better license (which ideally should be CC BY 4.0, same as the caffe model, just to keep it simple).
\n\nOr even simpler, just assume you aren't going to win and use it anyway . . . that's a safe bet for the vast majority of us after all.
", "rawMarkdown": "[quote=threecourse;119243]\r\n\r\nIf so, because [caffe VGG model][1] weights are same as [weights in VGG site][2], \r\nmy idea is that [keras VGG model][3], whose weights are obtained by converting, is okay. \r\nother ideas?\r\n\r\n[quote=Neil Slater;119224]\r\n\r\n[quote=threecourse;119196]\r\n\r\nIt says \"The models are released under Creative Commons Attribution License.\" in http://www.robots.ox.ac.uk/~vgg/research/very_deep/ \r\n\r\nIs it allowed to convert weights obtained from there? \r\n\r\n[/quote]\r\n\r\nYes, I think that is OK . That license is https://creativecommons.org/licenses/by/4.0/ which *allows* commercial use, as long as credit is given. It is the caffe model which adds the \"no commercial\" restriction, which the keras gist then used (and thus inherited the restriction). \r\n\r\nThe keras-model-zoo version appears to be incorrectly licensed, unless separate agreement reached with VGG team. That is not clear.\r\n\r\n\r\n[/quote]\r\n\r\n\r\n [1]: https://gist.github.com/ksimonyan/211839e770f7b538e2d8#file-readme-md\r\n [2]: http://www.robots.ox.ac.uk/~vgg/software/very_deep/caffe/VGG_ILSVRC_16_layers.caffemodel\r\n [3]: https://gist.github.com/baraldilorenzo/07d7802847aaad0a35d3\r\n\r\n[/quote]\r\n\r\nNo, that is not true, the licensing as written is not good for that. There is work involved in the conversion, and the non-commercial only restriction applies to that work due to how the software has been published, even though the original source doesn't have the restriction. In general it is not safe to use the kind of logical thinking you want to apply, you have to go with what is written for the license. \r\n\r\nEven if what is written doesn't make much sense, that just makes the product unsafe to use commercially. So you might get away with it in your own work - I doubt the VGG team or anyone in the Keras team really wants to sue anyone - but that is not the same as the thing being good to claim a prize in this competition.\r\n\r\nThe simplest thing to do is make the conversion form caffe model to Keras yourself, or hope that someone else does so under a better license (which ideally should be CC BY 4.0, same as the caffe model, just to keep it simple).\r\n\r\nOr even simpler, just assume you aren't going to win and use it anyway . . . that's a safe bet for the vast majority of us after all.\r\n", "votes": 1 }, { "id": 119243, "postDate": "2016-05-08T11:06:11.767Z", "content": "If so, because caffe VGG model weights are same as weights in VGG site,
\nmy idea is that keras VGG model, whose weights are obtained by converting, is okay.
\nother ideas?
[quote=Neil Slater;119224]
\n\n[quote=threecourse;119196]
\n\nIt says "The models are released under Creative Commons Attribution License." in http://www.robots.ox.ac.uk/~vgg/research/very_deep/
\n\nIs it allowed to convert weights obtained from there?
\n\n[/quote]
\n\nYes, I think that is OK . That license is https://creativecommons.org/licenses/by/4.0/ which allows commercial use, as long as credit is given. It is the caffe model which adds the "no commercial" restriction, which the keras gist then used (and thus inherited the restriction).
\n\nThe keras-model-zoo version appears to be incorrectly licensed, unless separate agreement reached with VGG team. That is not clear.
\n\n[/quote]
", "rawMarkdown": "If so, because [caffe VGG model][1] weights are same as [weights in VGG site][2], \r\nmy idea is that [keras VGG model][3], whose weights are obtained by converting, is okay. \r\nother ideas?\r\n\r\n[quote=Neil Slater;119224]\r\n\r\n[quote=threecourse;119196]\r\n\r\nIt says \"The models are released under Creative Commons Attribution License.\" in http://www.robots.ox.ac.uk/~vgg/research/very_deep/ \r\n\r\nIs it allowed to convert weights obtained from there? \r\n\r\n[/quote]\r\n\r\nYes, I think that is OK . That license is https://creativecommons.org/licenses/by/4.0/ which *allows* commercial use, as long as credit is given. It is the caffe model which adds the \"no commercial\" restriction, which the keras gist then used (and thus inherited the restriction). \r\n\r\nThe keras-model-zoo version appears to be incorrectly licensed, unless separate agreement reached with VGG team. That is not clear.\r\n\r\n\r\n[/quote]\r\n\r\n\r\n [1]: https://gist.github.com/ksimonyan/211839e770f7b538e2d8#file-readme-md\r\n [2]: http://www.robots.ox.ac.uk/~vgg/software/very_deep/caffe/VGG_ILSVRC_16_layers.caffemodel\r\n [3]: https://gist.github.com/baraldilorenzo/07d7802847aaad0a35d3", "votes": 1 }, { "id": 119224, "postDate": "2016-05-08T08:07:20.357Z", "content": "[quote=threecourse;119196]
\n\nIt says "The models are released under Creative Commons Attribution License." in http://www.robots.ox.ac.uk/~vgg/research/very_deep/
\n\nIs it allowed to convert weights obtained from there?
\n\n[/quote]
\n\nYes, I think that is OK . That license is https://creativecommons.org/licenses/by/4.0/ which allows commercial use, as long as credit is given. It is the caffe model which adds the "no commercial" restriction, which the keras gist then used (and thus inherited the restriction).
\n\nThe keras-model-zoo version appears to be incorrectly licensed, unless separate agreement reached with VGG team. That is not clear.
", "rawMarkdown": "[quote=threecourse;119196]\r\n\r\nIt says \"The models are released under Creative Commons Attribution License.\" in http://www.robots.ox.ac.uk/~vgg/research/very_deep/ \r\n\r\nIs it allowed to convert weights obtained from there? \r\n\r\n[/quote]\r\n\r\nYes, I think that is OK . That license is https://creativecommons.org/licenses/by/4.0/ which *allows* commercial use, as long as credit is given. It is the caffe model which adds the \"no commercial\" restriction, which the keras gist then used (and thus inherited the restriction). \r\n\r\nThe keras-model-zoo version appears to be incorrectly licensed, unless separate agreement reached with VGG team. That is not clear.\r\n", "votes": 1 }, { "id": 119172, "postDate": "2016-05-07T19:31:16.670Z", "content": "So, now this is interesting.
\n\nDoes the VGG-16 licence issue apply if it's been converted to another platform? (e.g., Keras)
", "rawMarkdown": "So, now this is interesting. \r\n\r\nDoes the VGG-16 licence issue apply if it's been [converted to another platform][1]? (e.g., Keras)\r\n\r\n\r\n [1]: https://gist.github.com/baraldilorenzo/07d7802847aaad0a35d3", "votes": 1 }, { "id": 119156, "postDate": "2016-05-07T17:41:42.597Z", "content": "Hi all,
\n\nSomeone in the community flagged the usage of VGG-16 here. We looked into the license and found this in their disclaimer:
\n\n\n\n\nlicense: http://creativecommons.org/licenses/by-nc/4.0/\n (non-commercial use only)
\n
Since it's non-commercial use only, State Farm won't be able to use it. So the usage of VGG-16 is not allowed.
", "rawMarkdown": "Hi all, \r\n\r\nSomeone in the community flagged the usage of [VGG-16][1] here. We looked into the license and found this in their disclaimer:\r\n\r\n> license: http://creativecommons.org/licenses/by-nc/4.0/\r\n> (non-commercial use only)\r\n\r\nSince it's non-commercial use only, State Farm won't be able to use it. So the usage of VGG-16 is not allowed. \r\n\r\n [1]: https://gist.github.com/ksimonyan/211839e770f7b538e2d8#file-readme-md", "votes": 1 }, { "id": 117804, "postDate": "2016-05-01T01:24:51.910Z", "content": "Another source of external data: Kaggle Facial Key point detection competition.
\n\nhttps://www.kaggle.com/c/facial-keypoints-detection/data
", "rawMarkdown": "Another source of external data: Kaggle Facial Key point detection competition.\r\n\r\nhttps://www.kaggle.com/c/facial-keypoints-detection/data", "votes": 1 }, { "id": 117140, "postDate": "2016-04-27T14:59:50.083Z", "content": "I want to use this code in http://pjreddie.com/darknet/yolo/ for object recognition.
", "rawMarkdown": "I want to use this code in http://pjreddie.com/darknet/yolo/ for object recognition.\r\n", "votes": 1 }, { "id": 116805, "postDate": "2016-04-26T06:00:56.947Z", "content": "[quote=Luis Andre Dutra e Silva;116115]
\n\nI have found the following pre-trained models the most useful ones until now:
\n\nGood luck for everyone!
\n\n[/quote]
\n\nIs the BVLC Google Net model is available on Keras as well?
", "rawMarkdown": "[quote=Luis Andre Dutra e Silva;116115]\r\n\r\nI have found the following pre-trained models the most useful ones until now:\r\n\r\n 1. BVLC - GoogleNet: https://github.com/BVLC/caffe/wiki/Model-Zoo\r\n 2. Facebook - ResNet: https://github.com/facebook/fb.resnet.torch/tree/master/pretrained\r\n 3. Microsoft Research - ResNet 1K: https://github.com/KaimingHe/resnet-1k-layers\r\n 4. I also recommend the use of NVIDIA Digits 3.0 and fine tuning its built-in BVLC GoogleNet. The URL for free registration and download is https://developer.nvidia.com/rdp/form/digits-download-survey\r\n\r\nGood luck for everyone!\r\n\r\n[/quote]\r\n\r\nIs the BVLC Google Net model is available on Keras as well?\r\n", "votes": 1 }, { "id": 116076, "postDate": "2016-04-21T22:11:09.343Z", "content": "", "rawMarkdown": "- [Torch models][1]\r\n- [Matconvnet models][2]\r\n\r\n\r\n [1]: https://github.com/torch/torch7/wiki/ModelZoo \"Torch Models\"\r\n [2]: http://www.vlfeat.org/matconvnet/pretrained/ \"Matconvnet models\"", "votes": 1 }, { "id": 116073, "postDate": "2016-04-21T22:07:22.757Z", "content": "", "rawMarkdown": " - [Lasagne models][1]\r\n - [Tensorflow models][2]\r\n - [Caffe models][3]\r\n\r\n\r\n [1]: https://github.com/Lasagne/Recipes/tree/master/modelzoo \"Lasagne Models\"\r\n [2]: https://github.com/tensorflow/models \"Tensorflow models\"\r\n [3]: https://github.com/BVLC/caffe/wiki/Model-Zoo \"Caffe models\"", "votes": 1 }, { "id": 115947, "postDate": "2016-04-20T23:16:47.763Z", "content": "VGG-19 layers model can be found here : http://www.robots.ox.ac.uk/~vgg/research/very_deep.\nExcept for inceptionV3, the other models are almost based on caffe framework. It is worth noting that pre-trained networks are beneficial but not the guarantee to high accuracy.
", "rawMarkdown": "VGG-19 layers model can be found here : http://www.robots.ox.ac.uk/~vgg/research/very_deep.\r\nExcept for inceptionV3, the other models are almost based on caffe framework. It is worth noting that pre-trained networks are beneficial but not the guarantee to high accuracy.", "votes": 1 }, { "id": 115099, "postDate": "2016-04-16T07:50:07.777Z", "content": "Some pre-built Haar cascade definitions for OpenCV: https://github.com/Itseez/opencv/tree/master/data/haarcascades
\n\n(Note the licenses, some are not suitable for this competition)
", "rawMarkdown": "Some pre-built Haar cascade definitions for OpenCV: https://github.com/Itseez/opencv/tree/master/data/haarcascades\r\n\r\n(Note the licenses, some are not suitable for this competition)", "votes": 1 }, { "id": 115004, "postDate": "2016-04-15T15:05:59.607Z", "content": "ResNet, MIT License\nhttps://github.com/KaimingHe/deep-residual-networks
\n\nResNet-1K,\nhttps://github.com/KaimingHe/resnet-1k-layers
\n\nBVLC-GoogLeNet\nhttps://github.com/BVLC/caffe/tree/master/models/bvlc_googlenet\n"License\nThis model is released for unrestricted use."
", "rawMarkdown": "ResNet, MIT License\r\nhttps://github.com/KaimingHe/deep-residual-networks\r\n\r\nResNet-1K,\r\nhttps://github.com/KaimingHe/resnet-1k-layers\r\n\r\nBVLC-GoogLeNet\r\nhttps://github.com/BVLC/caffe/tree/master/models/bvlc_googlenet\r\n\"License\r\nThis model is released for unrestricted use.\"", "votes": 1 }, { "id": 126124, "postDate": "2016-07-06T15:18:27.990Z", "content": "Some other stuff, some of which might have been mentioned already:
\n\nhttps://github.com/ishay2b/VanillaCNN\nhttp://pjreddie.com/darknet/yolo/\nhttps://github.com/rbgirshick/fast-rcnn\nhttps://github.com/rbgirshick/py-faster-rcnn\nhttps://github.com/mitmul/deeppose\nhttps://github.com/shihenw/convolutional-pose-machines-release\nhttps://github.com/anewell/pose-hg-train\nhttp://mscoco.org/ (includes phone/drink data)
\n\nA good collection:\nhttps://github.com/kjw0612/awesome-deep-vision
", "rawMarkdown": "Some other stuff, some of which might have been mentioned already:\r\n\r\nhttps://github.com/ishay2b/VanillaCNN\r\nhttp://pjreddie.com/darknet/yolo/\r\nhttps://github.com/rbgirshick/fast-rcnn\r\nhttps://github.com/rbgirshick/py-faster-rcnn\r\nhttps://github.com/mitmul/deeppose\r\nhttps://github.com/shihenw/convolutional-pose-machines-release\r\nhttps://github.com/anewell/pose-hg-train\r\nhttp://mscoco.org/ (includes phone/drink data)\r\n\r\nA good collection:\r\nhttps://github.com/kjw0612/awesome-deep-vision", "votes": 2 }, { "id": 119815, "postDate": "2016-05-12T21:12:56.530Z", "content": "@Matteo Presutto,
\n\nYes, semi-supervised learning is fine.
", "rawMarkdown": "@Matteo Presutto, \r\n\r\nYes, semi-supervised learning is fine.", "votes": 2 }, { "id": 114978, "postDate": "2016-04-15T08:37:49.723Z", "content": "There is a Keras import of VGG16 here: https://gist.github.com/baraldilorenzo/07d7802847aaad0a35d3
", "rawMarkdown": "There is a Keras import of VGG16 here: https://gist.github.com/baraldilorenzo/07d7802847aaad0a35d3\r\n", "votes": 2 }, { "id": 119178, "postDate": "2016-05-07T20:29:33.213Z", "content": "This is a bit of a gray area here. The rule of thumb is "if it's free to use by anyone (including State Farm), it's allowed". If you want to use it, please contact the pre-trained model author and confirm that it can be used.
\n\nEDIT: Thanks to @Ctrl+W who found a source that has "unrestricted use" license. Please stick to this version (or others that you can find free license) if you're interested in VGG-16.
\n\n[quote=inversion;119172]
\n\nSo, now this is interesting.
\n\nDoes the VGG-16 licence issue apply if it's been converted to another platform? (e.g., Keras)
\n\n[/quote]
", "rawMarkdown": "This is a bit of a gray area here. The rule of thumb is \"if it's free to use by anyone (including State Farm), it's allowed\". If you want to use it, please contact the pre-trained model author and confirm that it can be used. \r\n\r\nEDIT: Thanks to @Ctrl+W who found a [source][1] that has \"unrestricted use\" license. Please stick to this version (or others that you can find free license) if you're interested in VGG-16. \r\n\r\n[quote=inversion;119172]\r\n\r\nSo, now this is interesting. \r\n\r\nDoes the VGG-16 licence issue apply if it's been [converted to another platform][2]? (e.g., Keras)\r\n\r\n\r\n[/quote]\r\n\r\n\r\n [1]: https://github.com/albertomontesg/keras-model-zoo/tree/master/models/VGG-16\r\n [2]: https://gist.github.com/baraldilorenzo/07d7802847aaad0a35d3" }, { "id": 117469, "postDate": "2016-04-29T08:18:30.263Z", "content": "[quote=Luis Andre Dutra e Silva;116906]
\n\nPre-trained models can be compared to software libraries. We could call them "data libraries" as they provide, in the case of imagenet best models, a good start in terms of semantic features already embedded in their binary files.
\n\n[/quote]
\n\nI may be wrong, but isn't it counter-intuitive? Loading weights trained on other images for your dataset, how will that help? Aren't the weights supposed to be specific to a dataset i.e driver images in this case?
", "rawMarkdown": "[quote=Luis Andre Dutra e Silva;116906]\r\n\r\nPre-trained models can be compared to software libraries. We could call them \"data libraries\" as they provide, in the case of imagenet best models, a good start in terms of semantic features already embedded in their binary files. \r\n\r\n[/quote]\r\n\r\nI may be wrong, but isn't it counter-intuitive? Loading weights trained on other images for your dataset, how will that help? Aren't the weights supposed to be specific to a dataset i.e driver images in this case?" }, { "id": 129592, "postDate": "2016-07-31T17:55:36.373Z", "content": "Didn't see this on this thread:\nhttps://github.com/metalbubble/places365
", "rawMarkdown": "Didn't see this on this thread:\r\nhttps://github.com/metalbubble/places365" }, { "id": 128932, "postDate": "2016-07-25T09:29:02.213Z", "content": "Maybe a duplicate, but:
\n\nhttps://github.com/Lasagne/Recipes/blob/master/modelzoo/inception_v3.py
\n\nhttps://github.com/KaimingHe/deep-residual-networks
\n\nhttps://github.com/facebook/fb.resnet.torch
", "rawMarkdown": "Maybe a duplicate, but:\r\n\r\n[https://github.com/Lasagne/Recipes/blob/master/modelzoo/inception_v3.py][1]\r\n\r\n[https://github.com/KaimingHe/deep-residual-networks][2]\r\n\r\n[https://github.com/facebook/fb.resnet.torch][3]\r\n\r\n\r\n [1]: https://github.com/Lasagne/Recipes/blob/master/modelzoo/inception_v3.py \"Inception V3\"\r\n [2]: https://github.com/KaimingHe/deep-residual-networks \"ResNet in Caffe\"\r\n [3]: https://github.com/facebook/fb.resnet.torch \"ResNet in Torch\"" }, { "id": 128857, "postDate": "2016-07-24T21:24:02.563Z", "content": "https://github.com/lim0606/caffe-googlenet-bn\nBN-inception trained with Caffe.\nMight be a duplicate
", "rawMarkdown": "https://github.com/lim0606/caffe-googlenet-bn\r\nBN-inception trained with Caffe.\r\nMight be a duplicate" }, { "id": 125989, "postDate": "2016-07-05T10:17:47.563Z", "content": "could you please help with sourses to download caffe models, pretrained on specific data after being trained on imagenet. for example here http://img.cs.uec.ac.jp/pub/conf15/150703yanai_0.pdf written about experiments, where finetuning cnn on data from certain field (food images, mined from internet) helps to aquire hi accuracy where you have small own dataset
", "rawMarkdown": "could you please help with sourses to download caffe models, pretrained on specific data after being trained on imagenet. for example here http://img.cs.uec.ac.jp/pub/conf15/150703yanai_0.pdf written about experiments, where finetuning cnn on data from certain field (food images, mined from internet) helps to aquire hi accuracy where you have small own dataset" }, { "id": 125662, "postDate": "2016-07-01T11:49:36.220Z", "content": "you guys can try this (both caffe and tensorflow versions available)!\nIt looks awesome, see the attachment demo image from the website.
\n\nLicense:\nThe pre-trained models and the CAM technique are released for unrestricted use.\nContact Bolei Zhou if you have questions.
", "rawMarkdown": "you guys can try this (both caffe and tensorflow versions available)!\r\nIt looks awesome, see the attachment demo image from the website.\r\n\r\n - http://cnnlocalization.csail.mit.edu/\r\n - https://github.com/metalbubble/CAM\r\n - https://github.com/jazzsaxmafia/Weakly_detector\r\n\r\nLicense:\r\nThe pre-trained models and the CAM technique are released for unrestricted use.\r\nContact Bolei Zhou if you have questions.\r\n" }, { "id": 125606, "postDate": "2016-06-30T20:05:54.530Z", "content": "@ Florian: I am working on it. They seem needs a proper crop to reach those performances of VGG.
", "rawMarkdown": "@ Florian: I am working on it. They seem needs a proper crop to reach those performances of VGG." }, { "id": 125586, "postDate": "2016-06-30T16:17:33.600Z", "content": "[quote=all_random;125543]
\n\nWorking on:
\n\n[/quote]
\n\nAre you getting good results with networks pretrained on CIFAR-10?
\n\nI was able to replicate the preactivation and wide residual network results in Lasagne. If anyone wants to use those networks they can be found here: https://github.com/FlorianMuellerklein/Identity-Mapping-ResNet-Lasagne
", "rawMarkdown": "[quote=all_random;125543]\r\n\r\nWorking on:\r\n\r\n1. Pretrained caffe model, googlenet-bn: https://github.com/lim0606/caffe-googlenet-bn (license: unrestricted use)\r\n2. Pretrained lasagne model, resnet: https://github.com/Lasagne/Recipes/tree/master/papers/deep_residual_learning (license: unspecified)\r\n3. Pretrained caffe model, WRN: https://github.com/revilokeb/wide_residual_nets_caffe (license: MIT)\r\n\r\n[/quote]\r\n\r\n\r\nAre you getting good results with networks pretrained on CIFAR-10? \r\n\r\nI was able to replicate the preactivation and wide residual network results in Lasagne. If anyone wants to use those networks they can be found here: https://github.com/FlorianMuellerklein/Identity-Mapping-ResNet-Lasagne " }, { "id": 125543, "postDate": "2016-06-30T08:18:05.447Z", "content": "Working on:
\n\n[quote=Matteo Presutto;119524]
\n\nTake a look at this https://github.com/Lasagne/Recipes/blob/master/examples/Using%20a%20Caffe%20Pretrained%20Network%20-%20CIFAR10.ipynb , I had to implement a custom scale layer to import resnet50 though, plus I think caffe doesn't flip kernels (it does cross-correlation) while lasagne does by default
\n\n[/quote]
\n\nIs there copy written by R?
", "rawMarkdown": "[quote=Matteo Presutto;119524]\r\n\r\nTake a look at this https://github.com/Lasagne/Recipes/blob/master/examples/Using%20a%20Caffe%20Pretrained%20Network%20-%20CIFAR10.ipynb , I had to implement a custom scale layer to import resnet50 though, plus I think caffe doesn't flip kernels (it does cross-correlation) while lasagne does by default\r\n\r\n[/quote]\r\n\r\nIs there copy written by R?\r\n" }, { "id": 124791, "postDate": "2016-06-22T12:16:27.247Z", "content": "[quote=tmain;117804]
\n\nAnother source of external data: Kaggle Facial Key point detection competition.
\n\nhttps://www.kaggle.com/c/facial-keypoints-detection/data
\n\n[/quote]
\n\nI think but not sure this tutorial cannot help for State farm competition it has another approach , also I need feed-back from others.
", "rawMarkdown": "[quote=tmain;117804]\r\n\r\nAnother source of external data: Kaggle Facial Key point detection competition.\r\n\r\nhttps://www.kaggle.com/c/facial-keypoints-detection/data\r\n\r\n[/quote]\r\n\r\nI think but not sure this tutorial cannot help for State farm competition it has another approach , also I need feed-back from others." }, { "id": 124263, "postDate": "2016-06-16T14:57:48.323Z", "content": "Hi Wendy,
\n\nAs others have also pointed out, the license of VGGnet has been changed. \nPlease see here: http://www.robots.ox.ac.uk/~vgg/research/very_deep/.
\n\nCan you comfirm whether the use of VGGnet is allowed or not.\nThanks
\n\n[quote=Wendy Kan;119156]
\n\nHi all,
\n\nSomeone in the community flagged the usage of VGG-16 here. We looked into the license and found this in their disclaimer:
\n\n\n\n\nlicense: http://creativecommons.org/licenses/by-nc/4.0/\n (non-commercial use only)
\n
Since it's non-commercial use only, State Farm won't be able to use it. So the usage of VGG-16 is not allowed.
\n\n[/quote]
", "rawMarkdown": "Hi Wendy,\r\n\r\nAs others have also pointed out, the license of VGGnet has been changed. \r\nPlease see here: http://www.robots.ox.ac.uk/~vgg/research/very_deep/.\r\n\r\nCan you comfirm whether the use of VGGnet is allowed or not.\r\nThanks\r\n\r\n[quote=Wendy Kan;119156]\r\n\r\nHi all, \r\n\r\nSomeone in the community flagged the usage of [VGG-16][1] here. We looked into the license and found this in their disclaimer:\r\n\r\n> license: http://creativecommons.org/licenses/by-nc/4.0/\r\n> (non-commercial use only)\r\n\r\nSince it's non-commercial use only, State Farm won't be able to use it. So the usage of VGG-16 is not allowed. \r\n\r\n [1]: https://gist.github.com/ksimonyan/211839e770f7b538e2d8#file-readme-md\r\n\r\n[/quote]\r\n" }, { "id": 122365, "postDate": "2016-06-03T12:14:40.343Z", "content": "Trying VGG16:\nhttps://github.com/machrisaa/tensorflow-vgg
", "rawMarkdown": "Trying VGG16:\r\nhttps://github.com/machrisaa/tensorflow-vgg" }, { "id": 121360, "postDate": "2016-05-25T20:34:47.083Z", "content": "Admins - I would like to know if we can use ImageNet as well.
\n\nMore specifically: (a) is it ok to use ImageNet directly, and (b) is it ok to train a model on ImageNet, release that under an open-source license, and then use it. In the case of (b), which is the "external data", ImageNet or the model? When does the model have to be released?
\n\nPlease give us a ruling on this. Thanks!
\n\n[quote=Argmen;120490]
\n\nCan we use imagenet data in this ? imagenet
\n\n[/quote]
", "rawMarkdown": "Admins - I would like to know if we can use ImageNet as well. \r\n\r\nMore specifically: (a) is it ok to use ImageNet directly, and (b) is it ok to train a model on ImageNet, release that under an open-source license, and then use it. In the case of (b), which is the \"external data\", ImageNet or the model? When does the model have to be released? \r\n\r\nPlease give us a ruling on this. Thanks!\r\n\r\n* By ImageNet I mean the dataset for the ILSVRC2012 task 1.\r\n\r\n[quote=Argmen;120490]\r\n\r\nCan we use imagenet data in this ? [imagenet][1]\r\n\r\n\r\n [1]: http://www.image-net.org/\r\n\r\n[/quote]\r\n" }, { "id": 120945, "postDate": "2016-05-21T23:34:54.623Z", "content": "Using the following pre-trained model:
\n\nname: BVLC GoogleNet Model\ncaffemodel: bvlc_googlenet.caffemodel\ncaffemodel_url: http://dl.caffe.berkeleyvision.org/bvlc_googlenet.caffemodel\nlicense: unrestricted
", "rawMarkdown": "Using the following pre-trained model:\r\n\r\nname: BVLC GoogleNet Model\r\ncaffemodel: bvlc_googlenet.caffemodel\r\ncaffemodel_url: http://dl.caffe.berkeleyvision.org/bvlc_googlenet.caffemodel\r\nlicense: unrestricted" }, { "id": 120667, "postDate": "2016-05-19T21:33:22.167Z", "content": "BVLC models available at https://github.com/BVLC/caffe/tree/master/models
", "rawMarkdown": "BVLC models available at https://github.com/BVLC/caffe/tree/master/models" }, { "id": 120134, "postDate": "2016-05-15T19:33:50.453Z", "content": "", "rawMarkdown": "[Lasagne Modelzoo][1]\r\n\r\n\r\n [1]: https://github.com/Lasagne/Recipes/tree/master/modelzoo" }, { "id": 120119, "postDate": "2016-05-15T16:44:52.547Z", "content": "If you go ahead, you should probably reference the original unposed models here for people to get them if they want to copy the idea - just posting the idea here is similar to saying "I'm going to use some Creative Commons photos of car interiors - is that ok?" - you would need to link your source for the photos, so you should similarly link your source for the 3D models. IMO, that doesn't mean you need to upload anything like a completed Blender scene, with all the assets in place. Just the meshes, textures, accessories and morphs that you haven't coded yourself should be sufficient.
\n\nI think it's a cool idea, but I also think it will be hard work to create synthetic data for this problem that will help the model generalise (and not just extend the problem domain to the synthetic images).
", "rawMarkdown": "If you go ahead, you should probably reference the original unposed models here for people to get them if they want to copy the idea - just posting the idea here is similar to saying \"I'm going to use some Creative Commons photos of car interiors - is that ok?\" - you would need to link your source for the photos, so you should similarly link your source for the 3D models. IMO, that doesn't mean you need to upload anything like a completed Blender scene, with all the assets in place. Just the meshes, textures, accessories and morphs that you haven't coded yourself should be sufficient.\r\n\r\nI think it's a cool idea, but I also think it will be hard work to create synthetic data for this problem that will help the model generalise (and not just extend the problem domain to the synthetic images).\r\n" }, { "id": 120046, "postDate": "2016-05-14T21:57:50.460Z", "content": "I'm going to use make human models, with morphs. All the 3d models (such as car/cloth/hair) I will use are going to be available for free-to use for any purpose. Rendering will use free software (Blender).
\n\nThinking about this again, I think I can upload the blender files.
", "rawMarkdown": "I'm going to use make human models, with morphs. All the 3d models (such as car/cloth/hair) I will use are going to be available for free-to use for any purpose. Rendering will use free software (Blender).\r\n\r\nThinking about this again, I think I can upload the blender files.\r\n\r\n" }, { "id": 120039, "postDate": "2016-05-14T20:05:28.040Z", "content": "[quote=Obben;120021]
\n\nIf I write a script that would generate some synthetic training data, do I still need to share the synthetic data ? The way I see is my program has 3 phases: 1. Generating synthetic training data, 2. Using that data to train an intermediate model, 3. Use the competition data, and output of intermediate model, for final model building, all of these are going to be submitted as the final program.
\n\n[/quote]
\n\nTruly synthetic data, as opposed to data augmentation, likely has a data source. For instance, if you are building Poser or Daz models and rendering them, you should probably reference the 3D model data you are using.
\n\nThe only way this isn't external data would be if you coded synthetic data renderings without any reference material, or generated your own graphic reference material from scratch. For human forms that is quite hard, but I suppose possible if you have some artistic skill.
", "rawMarkdown": "[quote=Obben;120021]\r\n\r\nIf I write a script that would generate some synthetic training data, do I still need to share the synthetic data ? The way I see is my program has 3 phases: 1. Generating synthetic training data, 2. Using that data to train an intermediate model, 3. Use the competition data, and output of intermediate model, for final model building, all of these are going to be submitted as the final program.\r\n\r\n[/quote]\r\n\r\nTruly synthetic data, as opposed to data augmentation, likely has a data source. For instance, if you are building Poser or Daz models and rendering them, you should probably reference the 3D model data you are using.\r\n\r\nThe only way this isn't external data would be if you coded synthetic data renderings without any reference material, or generated your own graphic reference material from scratch. For human forms that is quite hard, but I suppose possible if you have some artistic skill.\r\n" }, { "id": 120021, "postDate": "2016-05-14T17:38:58.703Z", "content": "If I write a script that would generate some synthetic training data, do I still need to share the synthetic data ? The way I see is my program has 3 phases: 1. Generating synthetic training data, 2. Using that data to train an intermediate model, 3. Use the competition data, and output of intermediate model, for final model building, all of these are going to be submitted as the final program.
", "rawMarkdown": "If I write a script that would generate some synthetic training data, do I still need to share the synthetic data ? The way I see is my program has 3 phases: 1. Generating synthetic training data, 2. Using that data to train an intermediate model, 3. Use the competition data, and output of intermediate model, for final model building, all of these are going to be submitted as the final program.\r\n\r\n\r\n\r\n" }, { "id": 119803, "postDate": "2016-05-12T20:39:44.677Z", "content": "Wendy Kan, is it possible to use semi-supervised learning? From what I red in the rules there should be no problem
", "rawMarkdown": "Wendy Kan, is it possible to use semi-supervised learning? From what I red in the rules there should be no problem" }, { "id": 119524, "postDate": "2016-05-11T06:44:12.080Z", "content": "Take a look at this https://github.com/Lasagne/Recipes/blob/master/examples/Using%20a%20Caffe%20Pretrained%20Network%20-%20CIFAR10.ipynb , I had to implement a custom scale layer to import resnet50 though, plus I think caffe doesn't flip kernels (it does cross-correlation) while lasagne does by default
", "rawMarkdown": "Take a look at this https://github.com/Lasagne/Recipes/blob/master/examples/Using%20a%20Caffe%20Pretrained%20Network%20-%20CIFAR10.ipynb , I had to implement a custom scale layer to import resnet50 though, plus I think caffe doesn't flip kernels (it does cross-correlation) while lasagne does by default" }, { "id": 119411, "postDate": "2016-05-10T02:30:00.757Z", "content": "does anyone have an idea about how to convert the pre-trained models in caffe to theano format (for lasagne)? It seems there are lots of pre-trained model data in caffe, but not in lasage. I'm talking about the ResNet50/101/152.. Thank you~
", "rawMarkdown": "does anyone have an idea about how to convert the pre-trained models in caffe to theano format (for lasagne)? It seems there are lots of pre-trained model data in caffe, but not in lasage. I'm talking about the ResNet50/101/152.. Thank you~" }, { "id": 119196, "postDate": "2016-05-08T00:45:07.593Z", "content": "It says "The models are released under Creative Commons Attribution License." in http://www.robots.ox.ac.uk/~vgg/research/very_deep/
\n\nIs it allowed to convert weights obtained from there?
", "rawMarkdown": "It says \"The models are released under Creative Commons Attribution License.\" in http://www.robots.ox.ac.uk/~vgg/research/very_deep/ \r\n\r\nIs it allowed to convert weights obtained from there? " }, { "id": 119174, "postDate": "2016-05-07T19:58:30.947Z", "content": "This is what I am wondering too.
\n\n[quote=inversion;119172]
\n\nSo, now this is interesting.
\n\nDoes the VGG-16 licence issue apply if it's been converted to another platform? (e.g., Keras)
\n\n[/quote]
", "rawMarkdown": "This is what I am wondering too.\r\n\r\n[quote=inversion;119172]\r\n\r\nSo, now this is interesting. \r\n\r\nDoes the VGG-16 licence issue apply if it's been [converted to another platform][1]? (e.g., Keras)\r\n\r\n\r\n [1]: https://gist.github.com/baraldilorenzo/07d7802847aaad0a35d3\r\n\r\n[/quote]\r\n" }, { "id": 118630, "postDate": "2016-05-04T14:49:05.757Z", "content": "neon models: \nhttp://neon.nervanasys.com/docs/latest/model_zoo.html
", "rawMarkdown": "neon models: \r\nhttp://neon.nervanasys.com/docs/latest/model_zoo.html" }, { "id": 116937, "postDate": "2016-04-26T15:47:02.430Z", "content": "[quote=V.AbhijayArora;116807]
\n\n[quote=Luis Andre Dutra e Silva;116408]
\n\nSorry, but neither I am much familiar with Torch/Lua. There are many tutorials on web about it. For example: https://www.lua.org/pil/contents.html
\n\n[/quote]
\n\nI'm a newbie here, so excuse me if my question sounds naiive. I wanted to know what's the benefit of using a pre-trained model?
\n\n[/quote]
\n\nPlease check this thread: https://www.kaggle.com/c/state-farm-distracted-driver-detection/forums/t/20336/what-is-the-best-score-using-pre-trained-model
", "rawMarkdown": "[quote=V.AbhijayArora;116807]\r\n\r\n[quote=Luis Andre Dutra e Silva;116408]\r\n\r\nSorry, but neither I am much familiar with Torch/Lua. There are many tutorials on web about it. For example: https://www.lua.org/pil/contents.html\r\n\r\n[/quote]\r\n\r\nI'm a newbie here, so excuse me if my question sounds naiive. I wanted to know what's the benefit of using a pre-trained model? \r\n\r\n[/quote]\r\n\r\nPlease check this thread: https://www.kaggle.com/c/state-farm-distracted-driver-detection/forums/t/20336/what-is-the-best-score-using-pre-trained-model\r\n" }, { "id": 116807, "postDate": "2016-04-26T06:06:51.227Z", "content": "[quote=Luis Andre Dutra e Silva;116408]
\n\nSorry, but neither I am much familiar with Torch/Lua. There are many tutorials on web about it. For example: https://www.lua.org/pil/contents.html
\n\n[/quote]
\n\nI'm a newbie here, so excuse me if my question sounds naiive. I wanted to know what's the benefit of using a pre-trained model?
", "rawMarkdown": "[quote=Luis Andre Dutra e Silva;116408]\r\n\r\nSorry, but neither I am much familiar with Torch/Lua. There are many tutorials on web about it. For example: https://www.lua.org/pil/contents.html\r\n\r\n[/quote]\r\n\r\nI'm a newbie here, so excuse me if my question sounds naiive. I wanted to know what's the benefit of using a pre-trained model? " }, { "id": 116693, "postDate": "2016-04-25T17:04:56.163Z", "content": "Hi ebverybody,\nVGG-16 for Keras can be found here : https://github.com/albertomontesg/keras-model-zoo/tree/master/models/VGG-16\nLicense: unrestricted use
", "rawMarkdown": "Hi ebverybody,\r\nVGG-16 for Keras can be found here : https://github.com/albertomontesg/keras-model-zoo/tree/master/models/VGG-16\r\nLicense: unrestricted use" }, { "id": 116409, "postDate": "2016-04-24T00:31:02.030Z", "content": "[quote=mind.cool ;116408]
\n\nSorry, but neither I am much familiar with Torch/Lua. There are many tutorials on the web about it. For example: https://www.lua.org/pil/contents.html
\n\n[/quote]
\n\nThanks for the link. I was just avoiding going through new language syntax. I will just use Python or R to call command line to generate the probabilities for test images.
", "rawMarkdown": "[quote=mind.cool ;116408]\r\n\r\nSorry, but neither I am much familiar with Torch/Lua. There are many tutorials on the web about it. For example: https://www.lua.org/pil/contents.html\r\n\r\n[/quote]\r\n\r\nThanks for the link. I was just avoiding going through new language syntax. I will just use Python or R to call command line to generate the probabilities for test images.\r\n\r\n" }, { "id": 116405, "postDate": "2016-04-23T23:53:32.273Z", "content": "[quote=mind.cool ;116115]
\n\nI have found the following pre-trained models the most useful ones until now:
\n\nGood luck for everyone!
\n\n[/quote]
\n\nThanks for sharing. I am not much familiar with torch/Lua. Can you please share some code how to predict on new data after fine tuning the model and saving it into CSV.
\n\nThanks in advance.
", "rawMarkdown": "[quote=mind.cool ;116115]\r\n\r\nI have found the following pre-trained models the most useful ones until now:\r\n\r\n 1. BVLC - GoogleNet: https://github.com/BVLC/caffe/wiki/Model-Zoo\r\n 2. Facebook - ResNet: https://github.com/facebook/fb.resnet.torch/tree/master/pretrained\r\n 3. Microsoft Research - ResNet 1K: https://github.com/KaimingHe/resnet-1k-layers\r\n 4. I also recommend the use of NVIDIA Digits 3.0 and fine tuning its built-in BVLC GoogleNet. The URL for free registration and download is https://developer.nvidia.com/rdp/form/digits-download-survey\r\n\r\nGood luck for everyone!\r\n\r\n[/quote]\r\n\r\nThanks for sharing. I am not much familiar with torch/Lua. Can you please share some code how to predict on new data after fine tuning the model and saving it into CSV.\r\n\r\nThanks in advance." }, { "id": 115720, "postDate": "2016-04-19T17:55:58.437Z", "content": "Extraction pretrained model: http://pjreddie.com/darknet/imagenet/#extraction
\n\nResNet pretrained models: https://github.com/facebook/fb.resnet.torch/tree/master/pretrained
", "rawMarkdown": "Extraction pretrained model: http://pjreddie.com/darknet/imagenet/#extraction\r\n\r\nResNet pretrained models: https://github.com/facebook/fb.resnet.torch/tree/master/pretrained" }, { "id": 115586, "postDate": "2016-04-19T08:15:47.030Z", "content": "ResNet Pretrained models: https://github.com/facebook/fb.resnet.torch/tree/master/pretrained
", "rawMarkdown": "ResNet Pretrained models: https://github.com/facebook/fb.resnet.torch/tree/master/pretrained", "replies": [ { "id": 411280, "postDate": "2018-10-27T19:04:59.913Z", "content": "hey loc.. did you chose same for segmentation problem like Airbus also...
", "rawMarkdown": "hey loc.. did you chose same for segmentation problem like Airbus also..." } ] }, { "id": 115088, "postDate": "2016-04-16T06:11:47.083Z", "content": "Inception-v3 https://github.com/tensorflow/models/tree/master/inception
", "rawMarkdown": "Inception-v3 https://github.com/tensorflow/models/tree/master/inception\r\n" }, { "id": 122425, "postDate": "2016-06-04T00:07:02.750Z", "rawMarkdown": "", "isDeleted": true }, { "id": 121845, "postDate": "2016-05-30T09:31:09.163Z", "rawMarkdown": "", "isDeleted": true }, { "id": 116906, "postDate": "2016-04-26T14:23:51.607Z", "rawMarkdown": "", "isDeleted": true }, { "id": 116408, "postDate": "2016-04-24T00:19:15.230Z", "rawMarkdown": "", "isDeleted": true }, { "id": 116115, "postDate": "2016-04-22T04:46:21.287Z", "rawMarkdown": "", "votes": 4, "isDeleted": true } ], "comments": [ { "id": 119194, "author_name": "Neil Slater", "author_url": "", "post_date": "2016-05-07T22:27:52.567000", "content": "[quote=Wendy Kan;119178]
\n\nThis is a bit of a gray area here. The rule of thumb is "if it's free to use by anyone (including State Farm), it's allowed". If you want to use it, please contact the pre-trained model author and confirm that it can be used.
\n\nEDIT: Thanks to @Ctrl+W who found a source that has "unrestricted use" license. Please stick to this version (or others that you can find free license) if you're interested in VGG-16.
\n\n[quote=inversion;119172]
\n\nSo, now this is interesting.
\n\nDoes the VGG-16 licence issue apply if it's been converted to another platform? (e.g., Keras)
\n\n[/quote]
\n\n[/quote]
\n\nThe "no restrictions" license may not be a legally valid here, and in the case of the keras gist, there is no usage license quoted at all. There is no evidence that the IP has been handled correctly in this case, and the publisher in keras-model-zoo has given credit to the gist but changed the licensing terms, which is not allowed unless they have permission from the author.
\n\nMost importantly, from the sponsor's point of view, they would be liable even if they used the source material in good faith based on the stated license found in keras-model-zoo, if the originating VGG team decided to enforce their proper licensing terms.
\n\nSometimes it is not so much a grey area, but rather individuals publishing code or data - often in good faith trying to make something useful - without understanding the fiddly details that affect software licensing.
\n\nIf the authors of the relevant code are here on Kaggle, it would be great to get clarification.
\n\nI feel a little responsible here, since I posted the VGG/Keras gist link on this thread, and know enough about licensing to spot this. I was being a bit naive, as I haven't encountered very many IP problems with machine learning libraries in the past.
", "votes": 5, "replies": [] }, { "id": 117476, "author_name": "Neil Slater", "author_url": "", "post_date": "2016-04-29T09:04:52.967000", "content": "[quote=V.AbhijayArora;117469]
\n\n[quote=Luis Andre Dutra e Silva;116906]
\n\nPre-trained models can be compared to software libraries. We could call them "data libraries" as they provide, in the case of imagenet best models, a good start in terms of semantic features already embedded in their binary files.
\n\n[/quote]
\n\nI may be wrong, but isn't it counter-intuitive? Loading weights trained on other images for your dataset, how will that help? Aren't the weights supposed to be specific to a dataset i.e driver images in this case?
\n\n[/quote]
\n\nLow level features such as edges, corners, dots and repeating textures, are common amongst many different types of image. In fact before really deep CNNs took off, there were feature extractors such as Sobel filters, which effectively pre-encoded the same kinds of basic image components.
\n\nThe idea is that a high-performing CNN trained very large number of images (more than we have, or have time for in this competition) will have learned pretty good low-level features, most of which apply to our competition too. Higher-level features less so, but then thats what the fine-tuning is for.
\n\nIt helps if the content is vaguely similar - something trained on people, perhaps a pose classifier, might require less re-training than something classifying breeds of pet.
", "votes": 3, "replies": [] }, { "id": 122668, "author_name": "toshi_k", "author_url": "", "post_date": "2016-06-06T10:17:32.877000", "content": "Situation has changed.
\nThe description of "non-commercial use only" for VGG16 and VGG19 were removed from the caffe zoo.
\nIt is clear that VGG models are provided here by "CC BY 4.0" (commercial use is allowed).
\nPlease rethink about using VGG16 and VGG19.
I am going to use a hand grasping sequence from http://www.hci.iis.u-tokyo.ac.jp/~cai-mj/utgrasp_dataset.html
", "votes": 1, "replies": [] }, { "id": 120952, "author_name": "toshi_k", "author_url": "", "post_date": "2016-05-22T07:07:58.597000", "content": "In VGG web site, VGG model is provided by "CC BY 4.0" (commercial use is allowed).
\nIn the caffe zoo, however, VGG_ILSVRC_16_layers is provided by "CC BY-NC 4.0" (non-commercial use only).
I am confused about license of VGG. Is using VGG model allowed ? not allowed ?
\nIs there any way to use VGG model in this competiton ?
Can we use imagenet data in this ? imagenet
", "votes": 1, "replies": [] }, { "id": 120412, "author_name": "keithshep", "author_url": "", "post_date": "2016-05-18T01:52:08.310000", "content": "I'm planning on experimenting with the following datasets:
\n\n", "votes": 1, "replies": [] }, { "id": 119272, "author_name": "Neil Slater", "author_url": "", "post_date": "2016-05-08T18:27:28.083000", "content": "[quote=threecourse;119243]
\n\nIf so, because caffe VGG model weights are same as weights in VGG site,
\nmy idea is that keras VGG model, whose weights are obtained by converting, is okay.
\nother ideas?
[quote=Neil Slater;119224]
\n\n[quote=threecourse;119196]
\n\nIt says "The models are released under Creative Commons Attribution License." in http://www.robots.ox.ac.uk/~vgg/research/very_deep/
\n\nIs it allowed to convert weights obtained from there?
\n\n[/quote]
\n\nYes, I think that is OK . That license is https://creativecommons.org/licenses/by/4.0/ which allows commercial use, as long as credit is given. It is the caffe model which adds the "no commercial" restriction, which the keras gist then used (and thus inherited the restriction).
\n\nThe keras-model-zoo version appears to be incorrectly licensed, unless separate agreement reached with VGG team. That is not clear.
\n\n[/quote]
\n\n[/quote]
\n\nNo, that is not true, the licensing as written is not good for that. There is work involved in the conversion, and the non-commercial only restriction applies to that work due to how the software has been published, even though the original source doesn't have the restriction. In general it is not safe to use the kind of logical thinking you want to apply, you have to go with what is written for the license.
\n\nEven if what is written doesn't make much sense, that just makes the product unsafe to use commercially. So you might get away with it in your own work - I doubt the VGG team or anyone in the Keras team really wants to sue anyone - but that is not the same as the thing being good to claim a prize in this competition.
\n\nThe simplest thing to do is make the conversion form caffe model to Keras yourself, or hope that someone else does so under a better license (which ideally should be CC BY 4.0, same as the caffe model, just to keep it simple).
\n\nOr even simpler, just assume you aren't going to win and use it anyway . . . that's a safe bet for the vast majority of us after all.
", "votes": 1, "replies": [] }, { "id": 119243, "author_name": "threecourse", "author_url": "", "post_date": "2016-05-08T11:06:11.767000", "content": "If so, because caffe VGG model weights are same as weights in VGG site,
\nmy idea is that keras VGG model, whose weights are obtained by converting, is okay.
\nother ideas?
[quote=Neil Slater;119224]
\n\n[quote=threecourse;119196]
\n\nIt says "The models are released under Creative Commons Attribution License." in http://www.robots.ox.ac.uk/~vgg/research/very_deep/
\n\nIs it allowed to convert weights obtained from there?
\n\n[/quote]
\n\nYes, I think that is OK . That license is https://creativecommons.org/licenses/by/4.0/ which allows commercial use, as long as credit is given. It is the caffe model which adds the "no commercial" restriction, which the keras gist then used (and thus inherited the restriction).
\n\nThe keras-model-zoo version appears to be incorrectly licensed, unless separate agreement reached with VGG team. That is not clear.
\n\n[/quote]
", "votes": 1, "replies": [] }, { "id": 119224, "author_name": "Neil Slater", "author_url": "", "post_date": "2016-05-08T08:07:20.357000", "content": "[quote=threecourse;119196]
\n\nIt says "The models are released under Creative Commons Attribution License." in http://www.robots.ox.ac.uk/~vgg/research/very_deep/
\n\nIs it allowed to convert weights obtained from there?
\n\n[/quote]
\n\nYes, I think that is OK . That license is https://creativecommons.org/licenses/by/4.0/ which allows commercial use, as long as credit is given. It is the caffe model which adds the "no commercial" restriction, which the keras gist then used (and thus inherited the restriction).
\n\nThe keras-model-zoo version appears to be incorrectly licensed, unless separate agreement reached with VGG team. That is not clear.
", "votes": 1, "replies": [] }, { "id": 119172, "author_name": "inversion", "author_url": "", "post_date": "2016-05-07T19:31:16.670000", "content": "So, now this is interesting.
\n\nDoes the VGG-16 licence issue apply if it's been converted to another platform? (e.g., Keras)
", "votes": 1, "replies": [] }, { "id": 119156, "author_name": "Wendy Kan", "author_url": "", "post_date": "2016-05-07T17:41:42.597000", "content": "Hi all,
\n\nSomeone in the community flagged the usage of VGG-16 here. We looked into the license and found this in their disclaimer:
\n\n\n\n\nlicense: http://creativecommons.org/licenses/by-nc/4.0/\n (non-commercial use only)
\n
Since it's non-commercial use only, State Farm won't be able to use it. So the usage of VGG-16 is not allowed.
", "votes": 1, "replies": [] }, { "id": 117804, "author_name": "tmain", "author_url": "", "post_date": "2016-05-01T01:24:51.910000", "content": "Another source of external data: Kaggle Facial Key point detection competition.
\n\nhttps://www.kaggle.com/c/facial-keypoints-detection/data
", "votes": 1, "replies": [] }, { "id": 117140, "author_name": "tereka", "author_url": "", "post_date": "2016-04-27T14:59:50.083000", "content": "I want to use this code in http://pjreddie.com/darknet/yolo/ for object recognition.
", "votes": 1, "replies": [] }, { "id": 116805, "author_name": "Abhijay Arora", "author_url": "", "post_date": "2016-04-26T06:00:56.947000", "content": "[quote=Luis Andre Dutra e Silva;116115]
\n\nI have found the following pre-trained models the most useful ones until now:
\n\nGood luck for everyone!
\n\n[/quote]
\n\nIs the BVLC Google Net model is available on Keras as well?
", "votes": 1, "replies": [] }, { "id": 116076, "author_name": "Jeremy Howard", "author_url": "", "post_date": "2016-04-21T22:11:09.343000", "content": "", "votes": 1, "replies": [] }, { "id": 116073, "author_name": "Jeremy Howard", "author_url": "", "post_date": "2016-04-21T22:07:22.757000", "content": "", "votes": 1, "replies": [] }, { "id": 115947, "author_name": "Xingzhong Du", "author_url": "", "post_date": "2016-04-20T23:16:47.763000", "content": "VGG-19 layers model can be found here : http://www.robots.ox.ac.uk/~vgg/research/very_deep.\nExcept for inceptionV3, the other models are almost based on caffe framework. It is worth noting that pre-trained networks are beneficial but not the guarantee to high accuracy.
", "votes": 1, "replies": [] }, { "id": 115099, "author_name": "Neil Slater", "author_url": "", "post_date": "2016-04-16T07:50:07.777000", "content": "Some pre-built Haar cascade definitions for OpenCV: https://github.com/Itseez/opencv/tree/master/data/haarcascades
\n\n(Note the licenses, some are not suitable for this competition)
", "votes": 1, "replies": [] }, { "id": 115004, "author_name": "old-ufo", "author_url": "", "post_date": "2016-04-15T15:05:59.607000", "content": "ResNet, MIT License\nhttps://github.com/KaimingHe/deep-residual-networks
\n\nResNet-1K,\nhttps://github.com/KaimingHe/resnet-1k-layers
\n\nBVLC-GoogLeNet\nhttps://github.com/BVLC/caffe/tree/master/models/bvlc_googlenet\n"License\nThis model is released for unrestricted use."
", "votes": 1, "replies": [] }, { "id": 126124, "author_name": "bobutis", "author_url": "", "post_date": "2016-07-06T15:18:27.990000", "content": "Some other stuff, some of which might have been mentioned already:
\n\nhttps://github.com/ishay2b/VanillaCNN\nhttp://pjreddie.com/darknet/yolo/\nhttps://github.com/rbgirshick/fast-rcnn\nhttps://github.com/rbgirshick/py-faster-rcnn\nhttps://github.com/mitmul/deeppose\nhttps://github.com/shihenw/convolutional-pose-machines-release\nhttps://github.com/anewell/pose-hg-train\nhttp://mscoco.org/ (includes phone/drink data)
\n\nA good collection:\nhttps://github.com/kjw0612/awesome-deep-vision
", "votes": 2, "replies": [] }, { "id": 119815, "author_name": "Wendy Kan", "author_url": "", "post_date": "2016-05-12T21:12:56.530000", "content": "@Matteo Presutto,
\n\nYes, semi-supervised learning is fine.
", "votes": 2, "replies": [] }, { "id": 114978, "author_name": "Neil Slater", "author_url": "", "post_date": "2016-04-15T08:37:49.723000", "content": "There is a Keras import of VGG16 here: https://gist.github.com/baraldilorenzo/07d7802847aaad0a35d3
", "votes": 2, "replies": [] }, { "id": 119178, "author_name": "Wendy Kan", "author_url": "", "post_date": "2016-05-07T20:29:33.213000", "content": "This is a bit of a gray area here. The rule of thumb is "if it's free to use by anyone (including State Farm), it's allowed". If you want to use it, please contact the pre-trained model author and confirm that it can be used.
\n\nEDIT: Thanks to @Ctrl+W who found a source that has "unrestricted use" license. Please stick to this version (or others that you can find free license) if you're interested in VGG-16.
\n\n[quote=inversion;119172]
\n\nSo, now this is interesting.
\n\nDoes the VGG-16 licence issue apply if it's been converted to another platform? (e.g., Keras)
\n\n[/quote]
", "votes": 0, "replies": [] }, { "id": 117469, "author_name": "Abhijay Arora", "author_url": "", "post_date": "2016-04-29T08:18:30.263000", "content": "[quote=Luis Andre Dutra e Silva;116906]
\n\nPre-trained models can be compared to software libraries. We could call them "data libraries" as they provide, in the case of imagenet best models, a good start in terms of semantic features already embedded in their binary files.
\n\n[/quote]
\n\nI may be wrong, but isn't it counter-intuitive? Loading weights trained on other images for your dataset, how will that help? Aren't the weights supposed to be specific to a dataset i.e driver images in this case?
", "votes": 0, "replies": [] }, { "id": 129592, "author_name": "datapool", "author_url": "", "post_date": "2016-07-31T17:55:36.373000", "content": "Didn't see this on this thread:\nhttps://github.com/metalbubble/places365
", "votes": 0, "replies": [] }, { "id": 128932, "author_name": "Andrey Rykov", "author_url": "", "post_date": "2016-07-25T09:29:02.213000", "content": "Maybe a duplicate, but:
\n\nhttps://github.com/Lasagne/Recipes/blob/master/modelzoo/inception_v3.py
\n\nhttps://github.com/KaimingHe/deep-residual-networks
\n\nhttps://github.com/facebook/fb.resnet.torch
", "votes": 0, "replies": [] }, { "id": 128857, "author_name": "Guanshuo Xu", "author_url": "", "post_date": "2016-07-24T21:24:02.563000", "content": "https://github.com/lim0606/caffe-googlenet-bn\nBN-inception trained with Caffe.\nMight be a duplicate
", "votes": 0, "replies": [] }, { "id": 125989, "author_name": "Sergei Baramzin", "author_url": "", "post_date": "2016-07-05T10:17:47.563000", "content": "could you please help with sourses to download caffe models, pretrained on specific data after being trained on imagenet. for example here http://img.cs.uec.ac.jp/pub/conf15/150703yanai_0.pdf written about experiments, where finetuning cnn on data from certain field (food images, mined from internet) helps to aquire hi accuracy where you have small own dataset
", "votes": 0, "replies": [] }, { "id": 125662, "author_name": "", "author_url": "", "post_date": "2016-07-01T11:49:36.220000", "content": "", "votes": 0, "replies": [] }, { "id": 125606, "author_name": "", "author_url": "", "post_date": "2016-06-30T20:05:54.530000", "content": "", "votes": 0, "replies": [] }, { "id": 125586, "author_name": "", "author_url": "", "post_date": "2016-06-30T16:17:33.600000", "content": "", "votes": 0, "replies": [] }, { "id": 125543, "author_name": "", "author_url": "", "post_date": "2016-06-30T08:18:05.447000", "content": "", "votes": 0, "replies": [] }, { "id": 124792, "author_name": "", "author_url": "", "post_date": "2016-06-22T12:20:00.293000", "content": "", "votes": 0, "replies": [] }, { "id": 124791, "author_name": "", "author_url": "", "post_date": "2016-06-22T12:16:27.247000", "content": "", "votes": 0, "replies": [] }, { "id": 124263, "author_name": "", "author_url": "", "post_date": "2016-06-16T14:57:48.323000", "content": "", "votes": 0, "replies": [] }, { "id": 122365, "author_name": "", "author_url": "", "post_date": "2016-06-03T12:14:40.343000", "content": "", "votes": 0, "replies": [] }, { "id": 121360, "author_name": "", "author_url": "", "post_date": "2016-05-25T20:34:47.083000", "content": "", "votes": 0, "replies": [] }, { "id": 120945, "author_name": "", "author_url": "", "post_date": "2016-05-21T23:34:54.623000", "content": "", "votes": 0, "replies": [] }, { "id": 120667, "author_name": "", "author_url": "", "post_date": "2016-05-19T21:33:22.167000", "content": "", "votes": 0, "replies": [] }, { "id": 120134, "author_name": "", "author_url": "", "post_date": "2016-05-15T19:33:50.453000", "content": "", "votes": 0, "replies": [] }, { "id": 120119, "author_name": "", "author_url": "", "post_date": "2016-05-15T16:44:52.547000", "content": "", "votes": 0, "replies": [] }, { "id": 120046, "author_name": "", "author_url": "", "post_date": "2016-05-14T21:57:50.460000", "content": "", "votes": 0, "replies": [] }, { "id": 120039, "author_name": "", "author_url": "", "post_date": "2016-05-14T20:05:28.040000", "content": "", "votes": 0, "replies": [] }, { "id": 120021, "author_name": "", "author_url": "", "post_date": "2016-05-14T17:38:58.703000", "content": "", "votes": 0, "replies": [] }, { "id": 119803, "author_name": "", "author_url": "", "post_date": "2016-05-12T20:39:44.677000", "content": "", "votes": 0, "replies": [] }, { "id": 119524, "author_name": "", "author_url": "", "post_date": "2016-05-11T06:44:12.080000", "content": "", "votes": 0, "replies": [] }, { "id": 119411, "author_name": "", "author_url": "", "post_date": "2016-05-10T02:30:00.757000", "content": "", "votes": 0, "replies": [] }, { "id": 119196, "author_name": "", "author_url": "", "post_date": "2016-05-08T00:45:07.593000", "content": "", "votes": 0, "replies": [] }, { "id": 119174, "author_name": "", "author_url": "", "post_date": "2016-05-07T19:58:30.947000", "content": "", "votes": 0, "replies": [] }, { "id": 118630, "author_name": "", "author_url": "", "post_date": "2016-05-04T14:49:05.757000", "content": "", "votes": 0, "replies": [] }, { "id": 116937, "author_name": "", "author_url": "", "post_date": "2016-04-26T15:47:02.430000", "content": "", "votes": 0, "replies": [] }, { "id": 116807, "author_name": "", "author_url": "", "post_date": "2016-04-26T06:06:51.227000", "content": "", "votes": 0, "replies": [] }, { "id": 116693, "author_name": "", "author_url": "", "post_date": "2016-04-25T17:04:56.163000", "content": "", "votes": 0, "replies": [] }, { "id": 116409, "author_name": "", "author_url": "", "post_date": "2016-04-24T00:31:02.030000", "content": "", "votes": 0, "replies": [] }, { "id": 116405, "author_name": "", "author_url": "", "post_date": "2016-04-23T23:53:32.273000", "content": "", "votes": 0, "replies": [] }, { "id": 115720, "author_name": "", "author_url": "", "post_date": "2016-04-19T17:55:58.437000", "content": "", "votes": 0, "replies": [] }, { "id": 115586, "author_name": "", "author_url": "", "post_date": "2016-04-19T08:15:47.030000", "content": "", "votes": 0, "replies": [ { "id": 411280, "author_name": "", "author_url": "", "post_date": "2018-10-27T19:04:59.913000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 115088, "author_name": "", "author_url": "", "post_date": "2016-04-16T06:11:47.083000", "content": "", "votes": 0, "replies": [] }, { "id": 122425, "author_name": "", "author_url": "", "post_date": "2016-06-04T00:07:02.750000", "content": "", "votes": 0, "replies": [] }, { "id": 121845, "author_name": "", "author_url": "", "post_date": "2016-05-30T09:31:09.163000", "content": "", "votes": 0, "replies": [] }, { "id": 116906, "author_name": "", "author_url": "", "post_date": "2016-04-26T14:23:51.607000", "content": "", "votes": 0, "replies": [] }, { "id": 116408, "author_name": "", "author_url": "", "post_date": "2016-04-24T00:19:15.230000", "content": "", "votes": 0, "replies": [] }, { "id": 116115, "author_name": "", "author_url": "", "post_date": "2016-04-22T04:46:21.287000", "content": "", "votes": 4, "replies": [] } ], "raw_markdown_by_id": { "119194": "[quote=Wendy Kan;119178]\r\n\r\nThis is a bit of a gray area here. The rule of thumb is \"if it's free to use by anyone (including State Farm), it's allowed\". If you want to use it, please contact the pre-trained model author and confirm that it can be used. \r\n\r\nEDIT: Thanks to @Ctrl+W who found a [source][1] that has \"unrestricted use\" license. Please stick to this version (or others that you can find free license) if you're interested in VGG-16. \r\n\r\n[quote=inversion;119172]\r\n\r\nSo, now this is interesting. \r\n\r\nDoes the VGG-16 licence issue apply if it's been [converted to another platform][2]? (e.g., Keras)\r\n\r\n\r\n[/quote]\r\n\r\n\r\n [1]: https://github.com/albertomontesg/keras-model-zoo/tree/master/models/VGG-16\r\n [2]: https://gist.github.com/baraldilorenzo/07d7802847aaad0a35d3\r\n\r\n[/quote]\r\n\r\nThe \"no restrictions\" license may not be a legally valid here, and in the case of the keras gist, there is no usage license quoted at all. There is no evidence that the IP has been handled correctly in this case, and the publisher in keras-model-zoo has given credit to the gist but changed the licensing terms, which is not allowed unless they have permission from the author. \r\n\r\nMost importantly, from the sponsor's point of view, they would be liable even if they used the source material in good faith based on the stated license found in keras-model-zoo, if the originating VGG team decided to enforce their proper licensing terms.\r\n\r\nSometimes it is not so much a grey area, but rather individuals publishing code or data - often in good faith trying to make something useful - without understanding the fiddly details that affect software licensing.\r\n\r\nIf the authors of the relevant code are here on Kaggle, it would be great to get clarification.\r\n\r\nI feel a little responsible here, since I posted the VGG/Keras gist link on this thread, and know enough about licensing to spot this. I was being a bit naive, as I haven't encountered very many IP problems with machine learning libraries in the past.\r\n\r\n", "114941": "Please post here for use of pre-trained models and external data. \r\n\r\nThis competition will allow limited use of external data, such as pre-trained nets. \r\nYour data/model should be freely available to use by anyone in the community. ", "117476": "[quote=V.AbhijayArora;117469]\r\n\r\n[quote=Luis Andre Dutra e Silva;116906]\r\n\r\nPre-trained models can be compared to software libraries. We could call them \"data libraries\" as they provide, in the case of imagenet best models, a good start in terms of semantic features already embedded in their binary files. \r\n\r\n[/quote]\r\n\r\nI may be wrong, but isn't it counter-intuitive? Loading weights trained on other images for your dataset, how will that help? Aren't the weights supposed to be specific to a dataset i.e driver images in this case?\r\n\r\n[/quote]\r\n\r\nLow level features such as edges, corners, dots and repeating textures, are common amongst many different types of image. In fact before really deep CNNs took off, there were feature extractors such as Sobel filters, which effectively pre-encoded the same kinds of basic image components.\r\n\r\nThe idea is that a high-performing CNN trained very large number of images (more than we have, or have time for in this competition) will have learned pretty good low-level features, most of which apply to our competition too. Higher-level features less so, but then thats what the fine-tuning is for.\r\n\r\nIt helps if the content is vaguely similar - something trained on people, perhaps a pose classifier, might require less re-training than something classifying breeds of pet.\r\n", "122668": "Situation has changed.