To my understanding, we can train our models in kernels or locally and save the weights, and then, we can upload these weights to our final kernel, in order to be able to submit for stages 1 and 2.
\n\nIs the aforementioned sentence correct?
\n\nFurthermore, can we use pretrained models?
", "messages": [ { "id": "507909", "postDate": "04/05/2019 11:26:04", "content": "To my understanding, we can train our models in kernels or locally and save the weights, and then, we can upload these weights to our final kernel, in order to be able to submit for stages 1 and 2.
\n\nIs the aforementioned sentence correct?
\n\nFurthermore, can we use pretrained models?
", "rawMarkdown": "To my understanding, we can train our models in kernels or locally and save the weights, and then, we can upload these weights to our final kernel, in order to be able to submit for stages 1 and 2.\n\nIs the aforementioned sentence correct? \n\nFurthermore, can we use pretrained models?", "votes": null }, { "id": "508037", "postDate": "04/05/2019 14:46:33", "content": "Hi @dkaraflos
\n\nI'm not a competition organizer.
\n\nBut as far as I understood: yes, it is correct.
\n\nAnd yes, you can use models that have been pretrained in Imagenet, etc. The rule says \"Pretrained models may be used to construct the algorithms from publicly released academic datasets (ImageNet, iNaturalist, etc) full citation needed.\"
\n\nBut you must to announce that you're using this pretrained models.\nThe rule say: \"Please specify any and all external data used for training when uploading results in the specified discussion post.\"
\n\nMy interpretation from other competitions, is that you have to announce it if anybody hasn't announced it yet. The idea is that people say \"I'm using Resnet-50 trained in Imagenet\", and then the organizers validate the usage this model \"Yes, it's right\". But they don't expect the thousands of participants repeating the same if this pretrained model have been already validated (I think).
\n\nHope it helps
", "rawMarkdown": "Hi @dkaraflos \n\nI'm not a competition organizer.\n\nBut as far as I understood: yes, it is correct.\n\nAnd yes, you can use models that have been pretrained in Imagenet, etc. The rule says \"Pretrained models may be used to construct the algorithms from publicly released academic datasets (ImageNet, iNaturalist, etc) full citation needed.\"\n\nBut you must to announce that you're using this pretrained models.\nThe rule say: \"Please specify any and all external data used for training when uploading results in the specified discussion post.\"\n\nMy interpretation from other competitions, is that you have to announce it if anybody hasn't announced it yet. The idea is that people say \"I'm using Resnet-50 trained in Imagenet\", and then the organizers validate the usage this model \"Yes, it's right\". But they don't expect the thousands of participants repeating the same if this pretrained model have been already validated (I think).\n\nHope it helps", "votes": null }, { "id": "508133", "postDate": "04/05/2019 16:48:37", "content": "Thank you for yout reply. I hoped that it was this way. I just wanted to clarify this, before i start using any external weights model and then i realize that is not permitted ;)
", "rawMarkdown": "Thank you for yout reply. I hoped that it was this way. I just wanted to clarify this, before i start using any external weights model and then i realize that is not permitted ;)", "votes": null }, { "id": "508137", "postDate": "04/05/2019 17:04:43", "content": "Hello @dkaraflos
\n\nSame, I'm not a competition organizer.\nFrom my understanding, you must ensure that the pretrained model you use is published (see this post: https://www.kaggle.com/c/imet-2019-fgvc6/discussion/87544#latest-506684).
\n\nGood luck, Eric
", "rawMarkdown": "Hello @dkaraflos\n\nSame, I'm not a competition organizer.\nFrom my understanding, you must ensure that the pretrained model you use is published (see this post: https://www.kaggle.com/c/imet-2019-fgvc6/discussion/87544#latest-506684).\n\nGood luck, Eric", "votes": null } ], "comments": [ { "id": 508037, "author_name": "virilo", "author_url": "", "post_date": "04/05/2019 14:46:33", "content": "Hi @dkaraflos
\n\nI'm not a competition organizer.
\n\nBut as far as I understood: yes, it is correct.
\n\nAnd yes, you can use models that have been pretrained in Imagenet, etc. The rule says \"Pretrained models may be used to construct the algorithms from publicly released academic datasets (ImageNet, iNaturalist, etc) full citation needed.\"
\n\nBut you must to announce that you're using this pretrained models.\nThe rule say: \"Please specify any and all external data used for training when uploading results in the specified discussion post.\"
\n\nMy interpretation from other competitions, is that you have to announce it if anybody hasn't announced it yet. The idea is that people say \"I'm using Resnet-50 trained in Imagenet\", and then the organizers validate the usage this model \"Yes, it's right\". But they don't expect the thousands of participants repeating the same if this pretrained model have been already validated (I think).
\n\nHope it helps
", "votes": null, "replies": [ { "id": 508133, "author_name": "dkaraflos", "author_url": "", "post_date": "04/05/2019 16:48:37", "content": "Thank you for yout reply. I hoped that it was this way. I just wanted to clarify this, before i start using any external weights model and then i realize that is not permitted ;)
", "votes": null, "replies": [] } ] }, { "id": 508137, "author_name": "a45632", "author_url": "", "post_date": "04/05/2019 17:04:43", "content": "Hello @dkaraflos
\n\nSame, I'm not a competition organizer.\nFrom my understanding, you must ensure that the pretrained model you use is published (see this post: https://www.kaggle.com/c/imet-2019-fgvc6/discussion/87544#latest-506684).
\n\nGood luck, Eric
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "507909": "To my understanding, we can train our models in kernels or locally and save the weights, and then, we can upload these weights to our final kernel, in order to be able to submit for stages 1 and 2.\n\nIs the aforementioned sentence correct? \n\nFurthermore, can we use pretrained models?", "508037": "Hi @dkaraflos \n\nI'm not a competition organizer.\n\nBut as far as I understood: yes, it is correct.\n\nAnd yes, you can use models that have been pretrained in Imagenet, etc. The rule says \"Pretrained models may be used to construct the algorithms from publicly released academic datasets (ImageNet, iNaturalist, etc) full citation needed.\"\n\nBut you must to announce that you're using this pretrained models.\nThe rule say: \"Please specify any and all external data used for training when uploading results in the specified discussion post.\"\n\nMy interpretation from other competitions, is that you have to announce it if anybody hasn't announced it yet. The idea is that people say \"I'm using Resnet-50 trained in Imagenet\", and then the organizers validate the usage this model \"Yes, it's right\". But they don't expect the thousands of participants repeating the same if this pretrained model have been already validated (I think).\n\nHope it helps", "508133": "Thank you for yout reply. I hoped that it was this way. I just wanted to clarify this, before i start using any external weights model and then i realize that is not permitted ;)", "508137": "Hello @dkaraflos\n\nSame, I'm not a competition organizer.\nFrom my understanding, you must ensure that the pretrained model you use is published (see this post: https://www.kaggle.com/c/imet-2019-fgvc6/discussion/87544#latest-506684).\n\nGood luck, Eric" }, "source": "meta" }