search

GantMan / nsfw_model

Keras model of NSFW detector
Other
1.8k stars 279 forks source link

Why not resnet? #23

Open yangbisheng2009 opened 5 years ago

yangbisheng2009 commented 5 years ago

Since resnet is very strong now, why didn't your try it? Is it something wrong or some theory not suitable for nsfw？

Looking forward to your guidance

GantMan commented 5 years ago

Good question! To be fair, when I started this model, I was shooting for the best.

I think resnet is a great suggestion. I'll see about training a resnet version next!

TechnikEmpire commented 4 years ago

In 1.2.1, an upcoming release, there may be Resnet 50 V2 models that I've trained. However, Efficientnet models are also trained that are superior, they just can't be deployed to JS yet.

BerryRB commented 4 years ago

@TechnikEmpire Hi, can you tell me when will Version 1.2.1 be released?

TechnikEmpire commented 4 years ago

@BerryRB just look at the open PR I have for it. There are linked models there. You wont want resnet though its huge and offers little advantage over other models. In my fork I've made a readme page where you can see the stats of all trained models I've linked in the PR. @GantMan is the repo owner and its up to him to merge and officially post models but last time we spoke, he said he had a crazy rammed schedule so it might be a while before an official release.

yangbisheng2009 commented 4 years ago

https://github.com/yangbisheng2009/nsfw-resnet

I implement NSFW by pytorch-resnet101 Highly thanks GantMan，give me excellent suggestion

BerryRB commented 4 years ago

@TechnikEmpire Thank for your reply and great work! Can you tell me the amount of your training data and test data? Can you share your test data with me if you can? I currently have 2W cleaned data, which I can share with you if you need

TechnikEmpire commented 4 years ago

@BerryRB I just used about 6K images per category. I made the tool use a 90%/10% train/val split instead of the 80/20 used before.

With regards to my data, I won't risk distributing it due to a multitude of legal concerns.

I'm done with training these models personally as I believe my latest models have reached near the peak of capability for these networks on this task. I appreciate your offer to share your data but I'm simply moving on from this kind of project. All the best.