Show HN: Set of trained deep learning models for computer vision | Hacker News Reader

Show HN: Set of trained deep learning models for computer vision

github.com

127 pointsfchollet15 comments

Caffenet also offers a set of pre-trained models in their "model zoo": https://github.com/BVLC/caffe/wiki/Model-Zoo

Can the classify images example be modified to train on other (new) images. For example, images of screenshots to identify elements that are in the images (such as: word processor, browser, command prompt).

Yes, you can use these models for fine-tuning (or feature extraction) on a new dataset. This tutorial would be a good place to start (esp. sections 2 and 3): https://blog.keras.io/building-powerful-image-classification...

Are you allowed to redistribute the models under the MIT License?

The code is under the MIT license, not the weights. The weights are under their respective licenses.

The weights are not included in the git tree and are thus not covered by the LICENSE file. They are automatically downloaded when you run the code.

EDIT: following your comment, I have added a point-by-point breakdown of licensing information in the README. This will avoid any confusion.

Don't take my word for it, but actually as far as I know both in the US and EU data (including model weights) can't be copyrighted.

I'd like to see more information on this.

I ask because I'm surprised more companies haven't gotten into the market of licensing the data they collect (unless they do, in which case, sorry).

https://www.ldc.upenn.edu/

http://kingline.speechocean.com/

http://deeplearning.net/datasets/

http://www.elra.info/en/

Even if this might be true, including a license of some kind is really a best practice at this point. Something like a Creative Commons license is probably more appropriate than the MIT license, but when there's an explicit license there's no doubt about what your rights are.

Question:

Does anyone know of a library for loading models/weights from a registry of some sort?

Yes - pretty much every Deep Learning library. Caffe, Torch, Theano, TensorFlow etc (that's kinda what this link is about?)

Just use Keras on top of TensorFlow as shown at this link.

Well, this is fairly manual. More like:

my_model = registry.get("tensorflow://github.com/asdf/models/imagine/latest") ... my_model.push("...")

That's hundreds and hundreds of MBs you are downloading. It should never change, so it hardly seems a critical piece of functionality.

I guess someone could build it, sort of like the datasets you can download in SciKit or R, or the trained models in NLTK/Spacy.

In-fact I've almost come the full circle on this and think it might be a good idea.

Weird - I didn't think people on the internet could change their mind.

Heh :).

I was thinking that the current "best" model/architecture may change fairly frequently. Obviously you wouldn't want to download 100MB every time the application starts, but maybe amortized every time there is a significant jump would be good.

Anyway, I haven't seen anything like this, so was curious.

You mean something like this:

https://github.com/spacy-io/sputnik

Reply on news.ycombinator.com