search

Syn-McJ / TFClassify-Unity

An example of using Tensorflow with Unity for image classification and object detection.
MIT License
167 stars 47 forks source link

any update on how to fix the performance issue #17

Closed ICL-Developer closed 4 years ago

ICL-Developer commented 6 years ago

HI,

Any update on how fix the performance issue?

kindly waiting for you reply.

Thank You.

AvishekhDas commented 6 years ago

May be quantize the TF graph to further reduce the model size? Did you check that possibility? I deployed a pre-trained VGG model and the app crashed initially because of huge model size (>500MB). I then quantized the graph and size came down to < 130 MB. The app runs now, but the inference time is 4-5 seconds, which is still bad. Considering, imagenet models, the size can be brought down to may be 14-15 MB, but haven't tried it yet.

Syn-McJ commented 6 years ago

Hey folks, The best performance I achieved was using mobilenet models, which makes sense given that it was created specifically for mobile usage and features quick inference times though with a bit lower accuracy. Other than that no, I'm not sure how to improve performance.

AvishekhDas commented 6 years ago

Andrey,

Yes, Mobilenets are extremely fast, but isn't very accurate, at least in my scenario. Quantization works very well. For an InceptionV3 model, which is ~95 MB, Quantization reduces the size to ~24 MB without any loss of accuracy. You can check out quantization here: https://petewarden.com/2016/05/03/how-to-quantize-neural-networks-with-tensorflow/

You can also explore pruning, but I haven't tried it yet: https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/model_pruning

Syn-McJ commented 6 years ago

@AvishekhDas, Thanks for the links, I'll check them out!

Syn-McJ commented 4 years ago

My new example using new Unity Barracuda inference engine, it gives better performance: https://github.com/Syn-McJ/TFClassify-Unity-Barracuda