Now that this competition is almost over, I've been thinking ahead to future Kaggle competitions and wondering why there was a prohibition in this competition on using pretrained CNNs? After all, this is now a commonly used technique and since it still requires skill in tweaking hyperparameters and even CNN structure for a specific problem, it doesn't make a lot of sense to keep it out of the competition. Banning pretrained networks seems, at least to me, to make as much sense as banning the use of NN libraries like Caffe or Tensorflow.
", "messages": [ { "id": "111458", "postDate": "03/14/2016 19:35:22", "content": "Now that this competition is almost over, I've been thinking ahead to future Kaggle competitions and wondering why there was a prohibition in this competition on using pretrained CNNs? After all, this is now a commonly used technique and since it still requires skill in tweaking hyperparameters and even CNN structure for a specific problem, it doesn't make a lot of sense to keep it out of the competition. Banning pretrained networks seems, at least to me, to make as much sense as banning the use of NN libraries like Caffe or Tensorflow.
", "rawMarkdown": "Now that this competition is almost over, I've been thinking ahead to future Kaggle competitions and wondering why there was a prohibition in this competition on using pretrained CNNs? After all, this is now a commonly used technique and since it still requires skill in tweaking hyperparameters and even CNN structure for a specific problem, it doesn't make a lot of sense to keep it out of the competition. Banning pretrained networks seems, at least to me, to make as much sense as banning the use of NN libraries like Caffe or Tensorflow.", "votes": null }, { "id": "111461", "postDate": "03/14/2016 19:42:07", "content": "[quote=SpammySmith;111458]
\n\nwhy there was a prohibition in this competition on using pretrained CNNs?
\n\n[/quote]
\n\nIt is hard to be certain. I guess sponsor wanted something new. They could have used pre-trained networks themselves without much effort and wanted something that requires significant time and knowledge to develop
", "rawMarkdown": "[quote=SpammySmith;111458]\r\n\r\nwhy there was a prohibition in this competition on using pretrained CNNs? \r\n\r\n[/quote]\r\n\r\nIt is hard to be certain. I guess sponsor wanted something new. They could have used pre-trained networks themselves without much effort and wanted something that requires significant time and knowledge to develop", "votes": null }, { "id": "111470", "postDate": "03/14/2016 21:06:02", "content": "I think that's because the organizer wants to avoid using external data. Not every has access to the same kind of external data or pretrained model, so I think it's a fairness issue.
", "rawMarkdown": "I think that's because the organizer wants to avoid using external data. Not every has access to the same kind of external data or pretrained model, so I think it's a fairness issue.", "votes": null }, { "id": "111483", "postDate": "03/14/2016 23:13:08", "content": "@AL: A pre-trained network is just a starting point. There still would be plenty of time consuming tuning and modifications to do to improve the accuracy of the model, and some of that tuning might involve, as mentioned, changes to the architecture of the later stages of the CNN. Plenty of room for innovation.
\n\n@Wei Dong: There are plenty of pre-trained models that are now available. See the Caffe model zoo or Google's pre-trained Inception net. And there are many other parts of Kaggle contests that are unfair; for instance, if you have access to a GeForce TITAN GPU cluster, then you certainly have an edge if you are competing with me and my 5 year old non-GPU equipped laptop.
", "rawMarkdown": "AL: A pre-trained network is just a starting point. There still would be plenty of time consuming tuning and modifications to do to improve the accuracy of the model, and some of that tuning might involve, as mentioned, changes to the architecture of the later stages of the CNN. Plenty of room for innovation.\r\n\r\n@Wei Dong: There are plenty of pre-trained models that are now available. See [the Caffe model zoo][1] or [Google's pre-trained Inception net][2]. And there are many other parts of Kaggle contests that are unfair; for instance, if you have access to a GeForce TITAN GPU cluster, then you certainly have an edge if you are competing with me and my 5 year old non-GPU equipped laptop.\r\n\r\n\r\n [1]: http://caffe.berkeleyvision.org/model_zoo.html\r\n [2]: http://googleresearch.blogspot.com/2016/03/train-your-own-image-classifier-with.html", "votes": null } ], "comments": [ { "id": 111461, "author_name": "alexlzzz", "author_url": "", "post_date": "03/14/2016 19:42:07", "content": "[quote=SpammySmith;111458]
\n\nwhy there was a prohibition in this competition on using pretrained CNNs?
\n\n[/quote]
\n\nIt is hard to be certain. I guess sponsor wanted something new. They could have used pre-trained networks themselves without much effort and wanted something that requires significant time and knowledge to develop
", "votes": null, "replies": [] }, { "id": 111470, "author_name": "aaalgo", "author_url": "", "post_date": "03/14/2016 21:06:02", "content": "I think that's because the organizer wants to avoid using external data. Not every has access to the same kind of external data or pretrained model, so I think it's a fairness issue.
", "votes": null, "replies": [] }, { "id": 111483, "author_name": "spammy", "author_url": "", "post_date": "03/14/2016 23:13:08", "content": "@AL: A pre-trained network is just a starting point. There still would be plenty of time consuming tuning and modifications to do to improve the accuracy of the model, and some of that tuning might involve, as mentioned, changes to the architecture of the later stages of the CNN. Plenty of room for innovation.
\n\n@Wei Dong: There are plenty of pre-trained models that are now available. See the Caffe model zoo or Google's pre-trained Inception net. And there are many other parts of Kaggle contests that are unfair; for instance, if you have access to a GeForce TITAN GPU cluster, then you certainly have an edge if you are competing with me and my 5 year old non-GPU equipped laptop.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "111458": "Now that this competition is almost over, I've been thinking ahead to future Kaggle competitions and wondering why there was a prohibition in this competition on using pretrained CNNs? After all, this is now a commonly used technique and since it still requires skill in tweaking hyperparameters and even CNN structure for a specific problem, it doesn't make a lot of sense to keep it out of the competition. Banning pretrained networks seems, at least to me, to make as much sense as banning the use of NN libraries like Caffe or Tensorflow.", "111461": "[quote=SpammySmith;111458]\r\n\r\nwhy there was a prohibition in this competition on using pretrained CNNs? \r\n\r\n[/quote]\r\n\r\nIt is hard to be certain. I guess sponsor wanted something new. They could have used pre-trained networks themselves without much effort and wanted something that requires significant time and knowledge to develop", "111470": "I think that's because the organizer wants to avoid using external data. Not every has access to the same kind of external data or pretrained model, so I think it's a fairness issue.", "111483": "AL: A pre-trained network is just a starting point. There still would be plenty of time consuming tuning and modifications to do to improve the accuracy of the model, and some of that tuning might involve, as mentioned, changes to the architecture of the later stages of the CNN. Plenty of room for innovation.\r\n\r\n@Wei Dong: There are plenty of pre-trained models that are now available. See [the Caffe model zoo][1] or [Google's pre-trained Inception net][2]. And there are many other parts of Kaggle contests that are unfair; for instance, if you have access to a GeForce TITAN GPU cluster, then you certainly have an edge if you are competing with me and my 5 year old non-GPU equipped laptop.\r\n\r\n\r\n [1]: http://caffe.berkeleyvision.org/model_zoo.html\r\n [2]: http://googleresearch.blogspot.com/2016/03/train-your-own-image-classifier-with.html" }, "source": "meta" }