Hi Yahoo people.

Your blog post about the release of this model asks that if people develop their own model, to share it here in a bug ticket.

In the same blog post, you mention that you built a squeezenet 1.0 model that performed better, but released the ResNet-thin model (based on fine-tuning imagenet models).

I have trained a squeezenet 1.1 model from-scratch and have achieved, in my tests, slightly better overall accuracy than your released model, so I'd like to share it here and hopefully, as the blog post describes, get it included in your readme.

The model can be found here:

https://github.com/TechnikEmpire/NsfwSqueezenet

I achieved, again in my tests, ~99.1% overall accuracy on ~24,000 random test images, ~12K from each class. The best part is, using OpenCV's DNN module along with Intel's now open-source Inference Engine as the back-end, evaluation on the CPU on modern hardware*, including JPEG decoding and preprocessing, can be executed in ~3msec, or ~333fps. This is more than sufficient for real-time video classification on the CPU, and even ~13x realtime. With the same module, your model executes on the same hardware at ~8msec.

A cutoff (described in the project wiki) was determined through a process of averaging positive classification probability.

This work was sponsored in part by CloudVeil Technologies so I'd like to give them a nod there.

Thank you for open sourcing your model. It inspired me to to develop my model, because it was a clear demonstration of the possibilities in tackling this problem with deep learning.

• "Modern Hardware" here is Intel i7-6700 @ 3.4 ghz. Also of note, this model executes at ~30msec on a piddly little Intel Atom processor.