search

ENCP / CNNdroid

Open Source Library for GPU-Accelerated Execution of Trained Deep Convolutional Neural Networks on Android
MIT License
539 stars 181 forks source link

Support TensorFlow conversion #22

Open john1deer opened 7 years ago

john1deer commented 7 years ago

The conversion tools support most popular frameworks - but not the most popular of them all - TensorFlow. It would be nice to have an easy way to port a TF network to CNNdroid.

akbarxie commented 7 years ago

I have never use TF ,but in my project we used the model we defined ourselves,just try to get the weight and bias from TF moedel and send it to CNNdroid layers,it may be faster than get the next version of CNNdroid.(By the way the author is a very nice person) We will help you whenever you encounter problems.

someonelikeypp commented 5 years ago

Here are some questions to ask

  1. Can this project support the TensorFlow model now
  2. Is there any conditional limit: mobile phone system or opengl es version, because only tensorflow-lite is available on TensorFlow and the device should support at least opengl es 3.1

Hope to be able to get help

latifisalar commented 5 years ago
  1. Unfortunately, we have not developed the TF conversion script. But, if you extract the weights and biases from network and save them in MessagePack binary format, you should be able to run them on CNNdroid. You can have a look at Caffe to CNNdroid conversion script to get an idea about how to save them in MessagePack.
  2. The only system compatibility constraint is to have Android 4.0 and later running on your mobile phone. Other than that, you need to check the supported layers and network configurations. For example, residual connections are not supported.

Thanks

someonelikeypp commented 5 years ago

Thanks so much. It helps a lot.