search

keras-team / keras

Deep Learning for humans
http://keras.io/
Apache License 2.0
61.93k stars 19.46k forks source link

Adapting Yahoo Open_NSFW model to keras #4140

Closed ArmenAg closed 7 years ago

ArmenAg commented 8 years ago

How would you guys recommend adapting the yahoo opennsfw model (https://github.com/yahoo/open_nsfw) to keras. I tried using MarcBS's code but could not get it to work (https://github.com/MarcBS/keras)

Any help is appreciated.

rohrl commented 7 years ago

@ArmenAg have you succeeded ? I'm trying to do that too.

stale[bot] commented 7 years ago

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs, but feel free to re-open it if needed.

dribnet commented 6 years ago

note that this port has been done by @kashif though there might be issues with the current implementation: https://github.com/fchollet/deep-learning-models/pull/30

kashif commented 6 years ago

I have a new version @dribnet but only with tensorflow weights... I could push that?

dribnet commented 6 years ago

sure, I'd appreciate the update (and maybe others would too)!

dribnet commented 6 years ago

Hi @kashif - just pinging you again as I'm using your keras open_nsfw for a project currently and would love to get updated tensorflow weights if those are available and work better than the most current weights path currently found in your open_nsfw branch.

kashif commented 6 years ago

@dribnet here you go:

https://www.dropbox.com/s/os39us48nfepgll/open_nsfw_keras.zip?dl=0

Its not the cleanest and I got stuck trying to refactor the code... but hopefully it works!

dribnet commented 6 years ago

sorry - could you check the link or repost this? I was going to take a stab and refactoring this.

kashif commented 6 years ago

@dribnet ok sure... I moment

kashif commented 6 years ago

@dribnet should work again

dribnet commented 6 years ago

got it, thanks!

jessetrana commented 6 years ago

@dribnet Thanks for the hard work! I was giving this a try and have still been seeing some differences in relative to the caffe version. Can you elaborate on what parts were causing problems during the refactor? I've noticed that mmdnn is giving similar results to your translation, so I'm wondering if there is some fundamental mismatch on caffe vs. keras.