{ "id": 13793, "title": "Using Pre-Trained networks", "url": "/competitions/diabetic-retinopathy-detection/discussion/13793", "author_name": "", "post_date": "2015-04-26T21:51:23.360Z", "votes": 2, "comment_count": 8, "views": 2892, "content": "

Hi,

\n

Can we use a pre-trained network (e.g. http://www.vlfeat.org/matconvnet/pretrained/) as a part of our classifier?

", "messages": [ { "id": "74923", "postDate": "04/26/2015 21:51:23", "content": "

Hi,

\n

Can we use a pre-trained network (e.g. http://www.vlfeat.org/matconvnet/pretrained/) as a part of our classifier?

", "rawMarkdown": "", "votes": null }, { "id": "75198", "postDate": "04/28/2015 07:10:43", "content": "

No reason why you not.

\n\n

(AKA - Yes)

", "rawMarkdown": "", "votes": null }, { "id": "75204", "postDate": "04/28/2015 08:09:56", "content": "

As I understand, according to rule:

\n

"Use of external data is not permitted"

\n

We can not use pre-trained networks. Why? Because they use external data. Imagine that you have similar database with 1M images. Then pre-training on it will gave you really good performance. This is why I believe that using pre-trained model is not permitted. 

", "rawMarkdown": "", "votes": null }, { "id": "80432", "postDate": "05/30/2015 04:08:56", "content": "

I would really like to have an official position on this, since it looks to me like a rather strict and artificial limitation. Having 1M retinal images to train your model would give you an unfair advantage, but general-purpose pre-trained models are available to everyone. 

\n

In addition, transfer learning from general datasets (like ImageNet) is a key technique for making small datasets like the ones available for Computer-Aided Diagnosis feasible for Deep Learning.

\n

Let me be clear : for the moment our team is not using pre-trained models, exactly because the rule is ambiguous, but I'd feel we'd shot ourselves on the foot if later the organizers say "oh ! that is not really external external data".

", "rawMarkdown": "", "votes": null }, { "id": "80445", "postDate": "05/30/2015 10:08:26", "content": "

Agreed. Organizers can you please respond?

", "rawMarkdown": "", "votes": null }, { "id": "80527", "postDate": "05/31/2015 15:05:30", "content": "

Pre-trained networks are considered external data, thus not allowed.

", "rawMarkdown": "", "votes": null }, { "id": "80538", "postDate": "05/31/2015 19:02:36", "content": "

Got it

", "rawMarkdown": "Got it", "votes": null }, { "id": "80559", "postDate": "06/01/2015 09:18:13", "content": "

Can i use a couple of labeled images from stare dataset only to train my vessel detector? or i have do it this time consuming task personally on images in here?

", "rawMarkdown": "", "votes": null }, { "id": "80575", "postDate": "06/01/2015 13:03:22", "content": "

[quote=Mehdi Pedram;80559]

\n

Can i use a couple of labeled images from stare dataset only to train my vessel detector? or i have do it this time consuming task personally on images in here?

\n

[/quote]

\n

Sorry, but use of external data is not permitted.

", "rawMarkdown": "", "votes": null } ], "comments": [ { "id": 75198, "author_name": "danofer", "author_url": "", "post_date": "04/28/2015 07:10:43", "content": "

No reason why you not.

\n\n

(AKA - Yes)

", "votes": null, "replies": [] }, { "id": 75204, "author_name": "melgor", "author_url": "", "post_date": "04/28/2015 08:09:56", "content": "

As I understand, according to rule:

\n

"Use of external data is not permitted"

\n

We can not use pre-trained networks. Why? Because they use external data. Imagine that you have similar database with 1M images. Then pre-training on it will gave you really good performance. This is why I believe that using pre-trained model is not permitted. 

", "votes": null, "replies": [] }, { "id": 80432, "author_name": "evalle", "author_url": "", "post_date": "05/30/2015 04:08:56", "content": "

I would really like to have an official position on this, since it looks to me like a rather strict and artificial limitation. Having 1M retinal images to train your model would give you an unfair advantage, but general-purpose pre-trained models are available to everyone. 

\n

In addition, transfer learning from general datasets (like ImageNet) is a key technique for making small datasets like the ones available for Computer-Aided Diagnosis feasible for Deep Learning.

\n

Let me be clear : for the moment our team is not using pre-trained models, exactly because the rule is ambiguous, but I'd feel we'd shot ourselves on the foot if later the organizers say "oh ! that is not really external external data".

", "votes": null, "replies": [] }, { "id": 80445, "author_name": "jwjohnson314", "author_url": "", "post_date": "05/30/2015 10:08:26", "content": "

Agreed. Organizers can you please respond?

", "votes": null, "replies": [] }, { "id": 80527, "author_name": "wcukierski", "author_url": "", "post_date": "05/31/2015 15:05:30", "content": "

Pre-trained networks are considered external data, thus not allowed.

", "votes": null, "replies": [] }, { "id": 80538, "author_name": "hexiang", "author_url": "", "post_date": "05/31/2015 19:02:36", "content": "

Got it

", "votes": null, "replies": [] }, { "id": 80559, "author_name": "pedram", "author_url": "", "post_date": "06/01/2015 09:18:13", "content": "

Can i use a couple of labeled images from stare dataset only to train my vessel detector? or i have do it this time consuming task personally on images in here?

", "votes": null, "replies": [] }, { "id": 80575, "author_name": "wcukierski", "author_url": "", "post_date": "06/01/2015 13:03:22", "content": "

[quote=Mehdi Pedram;80559]

\n

Can i use a couple of labeled images from stare dataset only to train my vessel detector? or i have do it this time consuming task personally on images in here?

\n

[/quote]

\n

Sorry, but use of external data is not permitted.

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "74923": "", "75198": "", "75204": "", "80432": "", "80445": "", "80527": "", "80538": "Got it", "80559": "", "80575": "" }, "source": "meta" }